Behavioral Economics
–
Decisions under risk
Lecturer:
Prof. dr. Joël van der Weele
email: [Link]@[Link]
1 / 54
Outline
▶ EU utility and standard theory
▶ Anomalies
▶ Allais Paradox
▶ Certainty effect
▶ Reflection effect
▶ Prospect Theory
▶ Probability weighting
▶ Loss aversion (again)
▶ Cognitive foundations of prospect theory
2 / 54
Decision under risk
Suppose that we make a decision that yields different amounts of
money xi , each with probability pi .
▶ We assume probabilities are known (otherwise: ambiguity)
▶ “Known” can mean: subjectively constructed.
Definition (Simple Lottery)
A simple lottery Ls is a probability distribution over the set of
outcomes X = {x1 , . . . , xn },
Ls = (x1 , p1 ; x2 , p2 . . . ; xn , pn ).
Pn
▶ We assume pi ≥ 0 and i=1 pi = 1.
3 / 54
Decision under risk
How would you make a decision?
Simple rules of thumb
▶ Pick lottery with highest maximum outcome (maxmax)
▶ Pick lottery with highest minimum outcome (maxmin)
This ignores probabilities.
We could maximize expected value instead:
n
X
E[u] = pi xi
i=1
4 / 54
Bernoulli’s St.-Petersburg Paradox
Consider the following gamble:
▶ A fair coin is tossed until it lands on tail.
▶ In case it lands on tail for the first time on the k-th toss, you will
receive 2k dollars.
# 1 2 3 4 5 6 7 ...
$ 2 4 8 16 32 64 128 ...
▶ What is the maximum amount of money you would be willing to
pay for this lottery?
▶ Most people say dollars.
▶ The expected value of this lottery is
∞ k
X 1
E[u] = 2k =
2
k=1
5 / 54
Bernoulli’s St.-Petersburg Paradox #
▶ The Swiss mathematician D. Bernoulli
thought about this paradox in the 18th
century.
▶ Bernoulli hypothesized that the decision maker is not maximizing
the expected value of money but the expectation of the utility or
pleasure they receive from the money E[u] = ni=1 pi u(xi ).
P
▶ Bernoulli assumed that u(x) = ln(x). So the expected utility from
the gamble is
∞ ∞ k
X X 1
E[u] = pk ln(xk ) = ln(2k ) = ln(4)
2
k=1 k=1
6 / 54
Bernoullis St.-Petersburg Paradox
▶ Four dollars for certain would yield as much utility as the
participation in the gamble.
▶ So a willingness to pay of 4 dollars appears to have some empirical
validity.
▶ But, of course, the assumption that u(x) = ln(x) is somewhat
arbitrary.
Let’s consider a more rigorous axiomatic approach.
7 / 54
Axioms of Expected Utility Theory
Von Neumann and Morgenstern considered some normatively attractive
properties that preferences over lotteries should satisfy.
1. Rationality
2. Continuity
3. Consequentialism (Reduction of Compound Lotteries)
4. Independence
8 / 54
Rationality #
Axiom (Rationality)
The preference relation ⪰ over lotteries is complete, and transitive, i.e.,
(i) Either L ⪰ L̂ or L ⪯ L̂ (or both).
(iii) L ⪰ L̂ and L̂ ⪰ L̃ =⇒ L ⪰ L̃
9 / 54
Continuity #
Axiom (Continuity)
The preference relation ⪰ is continuous, i.e., for all L, L̂, L̃ with
L ≻ L̂ ≻ L̃ there exist α, β ∈ (0, 1) such that
(α, L ; (1 − α), L̃) ≻ L̂ ≻ (β, L ; (1 − β), L̃).
If you prefer the prize Mountain Bike over the prize Scooter and both
are preferred over winning nothing, then there is a probability α (e.g.,
99.9%) for which you will prefer the lottery of winning a Mountain Bike
with prob. α and nothing with 1 − α over getting a Scooter.
10 / 54
Reduction of Compound Lotteries #
Axiom (Consequentialism)
For all L and L̂ that induce the same probability distribution over X, we
have L ∼ L̂.
This means that the lotteries
L̂ = (L̃, 0.5 ; 100, 0.5) with L̃ = (0, 0.5 ; 100, 0.5)
and
L = (0, 0.25 ; 100, 0.75)
are as equivalent.
11 / 54
Independence #
Axiom (Independence)
For all L, L̂ and L̃ and α ∈ (0, 1], we have
L ⪰ L̂ ⇐⇒ (α · L ; (1 − α) · L̃) ⪰ (α · L̂ ; (1 − α) · L̃).
Suppose that you prefer going to the theatre over going to a basketball
game. Then, will you also prefer a lottery where you can win a theatre
ticket with prob. 10% over a lottery where you can win a basketball
game ticket with prob. 10%? If your answer is yes, then your choice
satisfies the independence (of irrelevant alternatives) axiom.
12 / 54
Expected Utility Theory #
Theorem (von Neumann–Morgenstern)
For any agent satisfying axioms rationality, continuity, consequentialism,
and independence, there exists a function u : X → R such that
n
X n
X
L ⪰ L̂ ⇐⇒ pi u(xi ) ≥ p̂i u(xi ).
i=1 i=1
▶ If the axioms are satisfied, then we can assign utilities to each
outcome of a lottery such that choosing the best lottery according
to the preference ⪰ amounts to choosing the lottery with the
highest expected utility.
13 / 54
Risk preferences
For the expected utility maximizer, the shape of the utility function
determines their risk preferences.
▶ A decision maker is risk averse if they prefer the expected value of
a lottery over the lottery itself. This is the case whenever their
utility function is concave, u ′′ (x) < 0
▶ A decision maker is risk neutral if they are indifferent between the
expected value of a lottery and the lottery itself. This is the case
whenever their utility function is linear, u ′′ (x) = 0
▶ A decision maker is risk loving if they prefer a lottery over its
expected value. This is the case whenever their utility function is
convex, u ′′ (x) > 0
14 / 54
Risk preferences
▶ Expected utility from a 50:50 lottery over 1 or 3 dollars for risk
averse and risk loving preferences.
Risk averse pref. Risk loving pref.
So we can explain gambling, or insurance, but not both.
15 / 54
Comments on EUT
▶ The expected utility theorem implies that when we model an agent
as maximizing expected utility, all we are really saying is that their
choice behavior adheres to some simple axioms.
▶ But we can use expected utility to make predictions about how
people actually behave.
▶ Expected utility theory is extremely general, so in principle can
apply to all kinds of decisions.
▶ In the laboratory, we can construct lotteries to test the theory in
detail.
16 / 54
Anomalies
17 / 54
Calibration theorem
Rabin, M. (2013). Risk aversion and expected-utility theory: A
calibration theorem. Econometrica, 68(5), 1281-1292.
▶ Suppose that an EU maximizer turns down
ℓ = (e11, 0.5; −e10, 0.5)
at any wealth level W .
▶ Consider an alternative lottery
ℓ′ = (e11x, 0.5; −e100, 0.5)
▶ What is the biggest value of x for which she will turn down ℓ′ ?
1. 10 2. 50 3. 1000 4. 1 million 5. ∞
18 / 54
Calibration theorem
Answer: ..... (See problem set).
Why?
▶ Rejecting small stakes lottery ℓ implies enormous concavity of the
utility function.
▶ The implied concavity means that no amount of gains suffice to
offset the potential loss of e100.
19 / 54
Calibration theorem
This conclusion seems absurd, even in the absence of (experimental)
evidence.
Lessons:
▶ Distinguish small and large stakes
▶ EUT can be a theory of large stakes, or small stakes, but not both.
▶ Caution with interpretation of small stakes, e.g. in laboratory
experiments
▶ Maybe reference points can offer a solution?
20 / 54
Towards an alternative for EUT
Kahneman, D. & Tversky, A. (1979). Prospect theory: An analysis of
decision under risk. Econometrica, 47(2), 363-391.
▶ Famous article (more than 88,000 citations on G. Scholar)
▶ Experiments on next slides taken from article (unless otherwise
noted)
▶ Shorthand: KT79.
21 / 54
Allais Paradox
Which lottery would you rather have?
Lottery A:
▶ 2500 dollars with 33 percent chance
▶ 2400 dollars with 66 percent chance
▶ 0 dollars with 1 percent chance
Lottery B:
▶ 2400 dollars for sure
22 / 54
Allais Paradox
Which lottery would you rather have?
Lottery C:
▶ 2500 dollars with 33 percent chance
▶ 0 dollars with 67 percent chance
Lottery D:
▶ 2400 dollars with 34 percent chance
▶ 0 dollars with 66 percent chance
23 / 54
0.33 e 2.500
Choice 1: A 0.66 e 2.400 or B 1.00 e 2.400
0.01 e0
0.33 e 2.500 0.34 e 2.400
Choice 2: C or D
0.67 e0 0.66 e0
Result:
▶ In Choice 1: 82% chose B.
▶ In Choice 2: 83% chose C.
▶ More specifically, 61% chose B in P1 and C in P2.
24 / 54
Your behavior
Result:
▶ In Choice 1: 58% chose B.
▶ In Choice 2: 79% chose C.
▶ More specifically, 42% chose B in P1 and C in P2.
42% of you showed a preference reversal
25 / 54
Allais Paradox
▶ Typical argument: A ≺ B in Choice 1 since 2400 is already a lot
and the chance of winning 2500 is not worth the risk of getting
nothing. But C ≻ D in Choice 2 since chances of winning 2500
instead of 2400 are only marginally smaller.
▶ But according to EUT, if A ≻ B ⇔ C ≻ D. Subtracting a 0.66
chance of winning 2400 from Choice 1 options yields Choice 2.
▶ This reduction has a stronger effect on utility if a certain outcome
becomes uncertain than if both the initial and the final lottery are
uncertain.
Possibility Effect
Relative to impossible outcomes, possible outcomes with very small
probabilities are overweighted.
26 / 54
Another experiment
0.8 e 4.000
Choice 3: A or B 1.00 e 3000
0.2 e0
0.2 e 4.000 0.25 e 3000
Choice 4: C or D
0.8 e0 0.75 e0
27 / 54
Certainty Effect
Result:
▶ In Choice 1: 80% chose B. (N=95)
▶ In Choice 2: 65% chose C. (N=95)
▶ More than 50% of subjects choose B in Choice 1 and C in Choice 2.
Interpretation:
▶ Again, according to EUT, if B ≻ A ⇔ C ≻ D. Why?
▶ KT(1979, p.266): “Apparently, reducing the probability of winning
from 1 to 0.25 has a bigger effect than the reduction from .8 to
.2.”
Certainty Effect
Relative to certain outcomes, uncertain outcomes with large
probabilities are underweighted.
28 / 54
Your behavior
Result:
▶ In Choice 1: 83% chose B.
▶ In Choice 2: 67% chose C.
▶ More specifically, 54% chose B in C1 and C in C2.
54% of you showed a preference reversal
29 / 54
More experiments!
0.8 e -4.000
Choice 5: A or B 1.00 e -3000
0.2 e0
0.2 e -4.000 0.25 e -3000
Choice 6: C or D
0.8 e0 0.75 e0
30 / 54
Reflection Effect
Result:
▶ In Choice 5: 92% chose A. (N=95)
▶ In Choice 6: 58% chose D. (N=95)
Observation:
▶ Observed choices again contradict EUT, which A ≻ B ⇔ C ≻ D.
▶ Compared to the previous experiment, the reflection of gains and
losses causes a preference reversal:
▶ In Choice 3, the majority prefers B (safe), but in Choice 5 most
prefer A (risky) .
▶ In Choice 4, the majority prefers C, but in Choice 6 most prefer D.
31 / 54
Your behavior
Result:
▶ In Choice 1: 71% chose A.
▶ In Choice 2: 63% chose D.
▶ More specifically, 54% chose B in C1 and C in C2.
54% of you showed a preference reversal
32 / 54
Reflection Effect
Reflection Effect
Mirroring lottery outcomes around zero (reflection) causes a preference
reversal.
▶ DM is risk averse in the gain domain and risk seeking in the loss
domain.
▶ This is incompatible with EUT because the EU maximizer only
cares about final wealth and not about gains and losses.
▶ Note that we are not just varying gains and losses, but also
(potential) final wealth levels.
33 / 54
Asian disease
Tversky, A., & Kahneman, D. (1981). The framing of decisions and the
psychology of choice. Science, 211(4481), 453-458.
“Imagine that the U.S. is preparing for the outbreak of an unusual Asian
disease, which is expected to kill 600 people. Two alternative programs
to combat the disease have been proposed. Assume that the exact
scientific estimate of the consequences of the programs are as follows:
▶ If program A is adopted, 200 people will be saved.
▶ If program B is adopted, there is 1/3 probability that 600 people
will be saved and 2/3 probability that no people will be saved.
Which of the two programs would you favor?”
72% chooses A, 28% chooses B (N = 152)
34 / 54
Asian disease
“Imagine that the U.S. is preparing for the outbreak of an unusual Asian
disease, which is expected to kill 600 people. Two alternative programs
to combat the disease have been proposed. Assume that the exact
scientific estimate of the consequences of the programs are as follows:
▶ If program A is adopted, 400 people will die.
▶ If program B is adopted, there is 1/3 probability that nobody will
die and 2/3 probability that 600 people will die.
Which of the two programs would you favor?
78% chooses B, 22% chooses A (N = 155)
Another demonstration of the reflection effect.
35 / 54
Your behavior
Result (between subject!):
▶ In Save Frame: 80% chose A.
▶ In Die frame: 75% chose B.
36 / 54
A Final Experiment
Choice 7: You have been given 1000. Choose between
0.5 e 1.000
A or B 1.00 e 500
0.5 e0
Choice 8: You have been given 2000. Choose between
0.5 e -1.000
A or B 1.00 e -500
0.5 e0
37 / 54
Isolation Effect
Result
▶ In Choice 7: 84% chose B. N=70
▶ In Choice 8: 69% chose C. N=69
Interpretation:
▶ Violates EUT: both lotteries yield same final wealth, so choice
should be the same.
▶ Subjects appear to ignore endowments, that are common to both
options, and focus on...
▶ ...additional gains in Choice 1;
▶ ...additional losses in Choice 2.
▶ Then, choices are motivated by reflection effect.
38 / 54
Your behavior
Result :
▶ In Choice 1: 67% chose B.
▶ In Choice 2: 54% chose C.
▶ More specifically, 33% chose B in C1 and C in C2.
33% of you showed a preference reversal
39 / 54
Reference dependence
Isolation Effect
People focus on components in which alternatives differ and disregard
components that the alternatives share.
KT 1979: “The apparent neglect of a bonus that was common to both
options [...] implies that the carriers of value or utility are changes of
wealth, rather than final asset positions that include current wealth.
This conclusion is the cornerstone of an alternative theory of risky
choice [...] ”
40 / 54
Summary of anomalies
EUT anomalies suggest that
1) Probabilities weighted in a non-linear way, overestimating small
probabilities. (Certainty effect and Possibility effect).
⇒ Probability weighting.
2) Utility/value of outcomes is not only determined by final wealth
levels, but by changes in wealth (isolation effect). It matters
whether changes are perceived as gains or losses. People are risk
averse for losses, and risk loving for gains (Reflection effect).
⇒ Reference dependence.
41 / 54
Four-fold pattern
The combination of probability weighting and reference dependence may
lead to a four-fold pattern.1
Gains Losses
Small probability Risk seeking Risk averse
Large probability Risk averse Risk seeking
Can explain simultaneous gambling and insurance
▶ enter a lottery offering a small probability of large prize
▶ take out insurance against a small probability of large loss
1 Tversky, Amos, and Daniel Kahneman. ”Advances in Prospect Theory:
Cumulative Representation of Uncertainty.” Journal of Risk and Uncertainty, 1992,
5, 297-323.
42 / 54
Prospect Theory
43 / 54
Prospect Theory: Introduction
▶ Inspired by the discussed EUT anomalies, Kahneman and Tversky
proposed an alternative theory: prospect theory (PT).
▶ The decision process consists of two phases:
▶ Editing: Simplification of choice problem (also called ”framing”),
▶ Evaluation: Evaluation of previously edited problem.
44 / 54
Prospect Theory: Editing
(i) Coding: DM defines a reference point (e.g., current wealth level)
and classifies the outcomes of a lottery either as gain or loss in
relation to this reference point.
(ii) Combination: Lotteries are simplified such that the same outcomes
are summed.
(iii) Segregation: Certain outcomes are separated from uncertain
outcomes. E.g.: L = (200, 0.2 ; 300, 0.8) is considered as a
certain gain of 200 plus the lottery L̂ = (0, 0.2 ; 100, 0.8).
(iv) Cancellation: If all lottery alternatives have the same components,
these are ignored.
(v) Detection of Dominance: If lottery L̂ is first-order stochastically
dominated by lottery L̃, then L̂ is deleted in the choice set.
45 / 54
Editing: Comments
▶ The editing phase manipulates the choice set and may already
violate EUT.
▶ E.g.: “Cancellation” can explain the “isolation effect”.
Problems:
▶ Predicted choices can depend on the order in which the different
editing steps are applied.
▶ Formalization was too vague to become a time-proven concept. In
their revision of prospect theory (cumulative prospect theory), KT
disregard editing as “no formal theory of framing is available”
(Tversky und Kahneman, 1992, p.299).
▶ In the following, we suppose that the lotteries we consider have
already passed the editing phase.
46 / 54
Shape and Features of the PT Function
Evaluation of simple lottery L
n
X
UP T (L|r ) = πi (pi ) v (xi − r )
i=1
▶ r ∈ R is the reference point, to which every outcome is compared;
▶ v (·) is the value function, which assigns a value to every deviation
from the reference point, with v (0) = 0;
▶ πi (pi ) is the decision weight assigned to outcome xi , as a function
of the actual probability.
We discuss these elements in turn.
47 / 54
Value Function: Shape
What do we know about the shape of the value function?
▶ Isolation effect: Changes in wealth with respect to a reference point
determine well-being. Gains are perceived differently than losses.
▶ Need for a reference point.
▶ Reflection effect: DM is risk averse in the domain of gains and risk
seeking in the domain of losses.
▶ Value function should be concave in the gain domain.
▶ Value function should be convex in the loss domain.
48 / 54
Value Function: Shape
What do we know about the shape of the value function?
▶ Most decision makers dislike symmetric lotteries like
L = (x, 0.5 ; −x, 0.5). They are risk averse around the reference
point (here equal to zero).
▶ The unattractiveness of symmetric lotteries increases with the size
of the outcomes, i.e., for x > y > 0 we typically have:
(y , 0.5 ; −y , 0.5) ≻ (x, 0.5 ; −x, 0.5).
▶ For r = 0, the previous condition implies that:
v (y ) + v (−y ) > v (x) + v (−x).
If y = 0, then v (x) < −v (−x): Losses loom larger than same-sized
gains – DM is loss averse.
49 / 54
The prospect theory value function
Value v (x)
Gains
Outcome (x)
0
Losses
50 / 54
Probability weighting
▶ Overweighting of small probabilities, underweighting of large
probabilities
▶ Insensitivity to changes in intermediate probabilities.
▶ Note that weights are not probabilities: may sum to less than one.
51 / 54
Development of Prospect Theory
Prospect theory
▶ Can explain all anomalies described previously
▶ Decades-long research program to calibrate and refine the theory.
▶ e.g. Functional forms for value function and probability weighting
▶ Advances in measurement
New anomalies and modifications
▶ PT can lead to strange predictions: i.e. preference for
stochastically dominated options (see problem set)
▶ E.g. favors lotteries with many improbable (and hence
overweighted) outcomes
▶ Revised version called ”cumulative prospect theory”, which
recalculates the probability weight.
52 / 54
Applications
Despite this research program, prospect theory has been slow to be
applied (Barberis, J. of Economic Perspectives, 2013)
▶ More complicated than EUT
▶ No theory of the reference point
▶ Work by Köszegi and Rabin: reference points as (rational)
expectations
Nevertheless,
▶ loss aversion is often used to explain behavior in labor supply,
bargaining, law (see last week)
▶ probability weighting is to explain (over)insurance, gambling
53 / 54
Interpretation
Do anomalies represent preferences over risk, or confusion/complexity?
Some evidence for confusion
▶ S-shape pattern is typical of “attenuation” in cognitive science.
▶ Decisions are often less sensitive further away from natural
boundaries (like 0 or 1)
▶ Some evidence for anomalies in non-risky decisions (Oprea 2024 -
but controversial)
▶ Are abstract lotteries really “risk”?
Does it matter?
▶ PT remains a descriptive theory of experiments
▶ External validity is in question
▶ Welfare prescriptions depend on true preferences.
54 / 54