Chapter 12: Behavioral and Collective Decisions

Algorithmic Game Theory
Prospect theoryLoss aversionQuantal responseSocial preferencesThe commons and OstromPublic goods and punishmentThreshold projectsInformation cascadesCondorcet jury theorem
Sercan KÜLCÜ · Giresun University · Computer Engineering

Behavioral Decisions

Warm-up: The Public Goods Game

The rules

Each round, four players get 10 tokens each and put any number into a common pool; the pool is multiplied by 1.6 and split equally among all four. Tokens not contributed are kept.

Think about it
  • In a group of four, how many tokens do you think are contributed on average in the first round? After ten rounds?
  • If anyone contributing fewer than 3 tokens is fined 2 tokens (a punishment variant), how do contributions change?
common pool1 token × 1.6contribute 10.40.40.40.4YouP2P3P4Net to you −1 + 0.4 = −0.6; +1.2 in total to the other three

Behavioral Decisions

Learning Outcomes

Behavioral decisionsValue a gamble with prospect theory; explain loss aversion, the endowment effect and framing; show how quantal response changes the Nash prediction.
Social preferencesExplain deviations in dictator, ultimatum and trust games by inequity aversion and reciprocity; compute the Fehr–Schmidt acceptance threshold.
Commons and public goodsCompare the commons equilibrium with the optimum; interpret Ostrom's principles, punishment and second-order free riding, and equilibria of threshold projects.
Information aggregationAnalyze the wisdom of crowds with the diversity theorem, information cascades with Bayes' rule, and the Condorcet jury theorem with binomial probabilities.

Behavioral Decisions

Are People Always Rational? Bounded Rationality

  • Bounded rationality (Simon): information, time and computing power are limited; people look for "good enough", not the best.
  • Deviations are systematic, not random: the same mistakes recur, so they can be modeled.
  • Causes: missing information, time pressure, emotions, shortcuts (heuristics).
OverconfidenceTrusting one's own estimates too much: investors trade too often.
Framing effectThe decision changes when the same option is presented differently.
Loss aversionA loss hurts more than an equal gain pleases: panic selling.
Base-rate neglectIgnoring how common a disease is (the prior) when reading a positive test result.

Behavioral Decisions

Prospect Theory

  • Kahneman and Tversky (1979): outcomes are judged as gains or losses relative to a reference point, not as final wealth.
  • The value function is concave for gains, convex for losses, and steeper for losses:
    v(x) = x0.88 (x ≥ 0); v(x) = −2.25·(−x)0.88 (x < 0)
  • x0.88 grows more slowly than x: v(200) ≈ 105.9, 1.84 times v(100), not 2. Concave: the curve flattens to the right.
  • λ = 2.25 is the loss aversion coefficient (Tversky and Kahneman's 1992 estimate).
  • Consequence (ignoring weights): risk aversion for gains, risk seeking for losses.
v(100) = 57.5v(−100) = −129.5gainlossvalue vreference
Reference point example: someone expecting a 10,000 TL bonus who gets 9,000 TL feels a 1,000 TL loss; to someone expecting nothing, the same amount is a big gain.

Behavioral Decisions

Loss Aversion: A Coin-Flip Bet

The betHeads +100 TL, tails −100 TL. Expected value 0.
Prospect value½·v(100) + ½·v(−100) = ½·57.5 − ½·129.5 = −36.0 < 0 ⇒ rejected.
Acceptance threshold½·G0.88 = ½·2.25·1000.88 ⇒ (G/100)0.88 = 2.25 ⇒ G = 100·2.251/0.88 ≈ 251 TL.
To risk losing 100 TL, a person needs a chance to win about 251 TL. In experiments, too, people typically ask for more than twice the loss.
Probability weighting w(p)p = 0.01p = 0.50p = 0.99
Perceived weight0.0550.4210.912

The probability weighting function w(p) shows how a probability feels (Tversky and Kahneman's 1992 estimate): 1% feels like about 5.5%, 50% like about 42%, 99% like about 91%. Small probabilities are overweighted and large ones underweighted: the same person may buy both lottery tickets and insurance.

Behavioral Decisions

Code Trace: When Is a Bet Accepted?

0+100 / −100+200 / −100+251 / −100+300 / −100expected valueprospect valueaccept: prospect value > 0

Four coin-flip bets: win +G with probability ½, lose 100 TL with probability ½. Expected value (G − 100)/2; prospect value uses v.

0+100 / −1000.0−36.0+200 / −100+251 / −100+300 / −100expected valueprospect valueaccept: prospect value > 0

G = 100: v(100) = 57.5 and v(−100) = −129.5 ⇒ ½·57.5 − ½·129.5 = −36.0. A bet with expected value 0 is rejected.

0+100 / −1000.0−36.0+200 / −10050.0−11.8+251 / −100+300 / −100expected valueprospect valueaccept: prospect value > 0

G = 200: v(200) = 105.9 ⇒ ½·105.9 − ½·129.5 = −11.8. Still rejected, although the expected value is +50.

0+100 / −1000.0−36.0+200 / −10050.0−11.8+251 / −10075.5−0.1+300 / −100expected valueprospect valueaccept: prospect value > 0

G = 251: v(251) = 129.3 ⇒ ½·129.3 − ½·129.5 ≈ −0.1: right at the acceptance threshold (≈ 251 TL).

0+100 / −1000.0−36.0+200 / −10050.0−11.8+251 / −10075.5−0.1+300 / −100100.010.9expected valueprospect valueaccept: prospect value > 0

G = 300: v(300) = 151.3 ⇒ ½·151.3 − ½·129.5 = +10.9 > 0: now the bet is accepted.

0+100 / −1000.0−36.0+200 / −10050.0−11.8+251 / −10075.5−0.1+300 / −100100.010.9expected valueprospect valueaccept: prospect value > 0

A risk-neutral person takes the +200 / −100 bet (expected value +50) at once; with λ = 2.25 the loss weighs so much that the bet is rejected until the gain exceeds 251 TL.

def v(x):
    if x >= 0: return x ** 0.88
    return -2.25 * (-x) ** 0.88
def prospect(G, K):      # ½: +G, ½: −K
    return 0.5 * v(G) + 0.5 * v(-K)
for G in [100, 200, 251, 300]:
    print(G, round(prospect(G, 100), 1))

Behavioral Decisions

A Trace of Loss Aversion: The Endowment Effect

  • Experiment (Kahneman, Knetsch and Thaler, 1990): half the students were randomly given a mug. Owners' minimum selling price was about twice the maximum price buyers were willing to pay, and far fewer trades took place than expected.
  • Explanation: giving up the mug is a loss from the reference point, getting it is a gain; the loss weighs λ times as much.
  • Status quo bias: leaving the default also feels like a loss, which is why automatic enrollment raises pension plan participation.
Example

The mug's use value is 20 TL, the value function is linear, λ = 2.25.

  • Owner: giving up the mug feels like −2.25·20 = −45 ⇒ asks at least 45 TL.
  • Buyer: getting the mug is +20 ⇒ pays at most 20 TL.
  • Ratio 45/20 = 2.25 = λ: no trade at prices in between.
Common mistake"The owner likes the mug more." No: the object is the same; only the reference point changes.

Behavioral Decisions

Framing: Same Option, Different Choice

Tversky and Kahneman (1981): two programs for an outbreak expected to affect 600 people.

Gain frame

  • A: 200 people are saved.
  • B: with probability 1/3, 600 are saved; with probability 2/3, no one is.
  • 72% of participants chose A (the sure thing).

Loss frame

  • C: 400 people die.
  • D: with probability 1/3, no one dies; with probability 2/3, 600 die.
  • 78% of participants chose D (the gamble).
A ≡ C and B ≡ D (200 expected survivors in each). Expected utility theory says the choice should not change; prospect theory explains it by risk aversion for gains and risk seeking for losses.

Behavioral Decisions

Common Mistakes: Prospect Theory

"Loss aversion is risk aversion."No. Risk aversion comes from the curvature of utility, loss aversion from the kink at the reference point (λ). The ±100 TL bet is rejected because of λ even if v is nearly linear for gains.
"w(p) is a probability."No, it is a decision weight: w(0.01) + w(0.99) = 0.055 + 0.912 = 0.967 ≠ 1, and w(0.5) = 0.421.
"The reference point is always zero."It can be an expectation, the status quo or a goal. To someone expecting a 10,000 TL bonus, 9,000 TL is a loss.
"In the loss frame everyone gambles."78% is a majority tendency. The theory explains average behavior; it does not predict individuals.

Behavioral Decisions

Noisy but Smart: Quantal Response

  • People choose the better option more often, but not always; the bigger the difference, the fewer the errors.
  • Logit choice: P(a) = eλ·u(a) / Σb eλ·u(b) (e ≈ 2.718; ex grows fast in x). At λ = 0 choice is random; as λ → ∞ it becomes a best response.
  • This λ has nothing to do with the loss-aversion λ = 2.25; it measures how "sharp" the choice is.
  • With two options the formula simplifies: P(better) = 1 / (1 + e−λ·diff). With diff 2 and λ = 1, P ≈ 0.88.
  • Quantal response equilibrium (QRE; McKelvey and Palfrey, 1995): every player gives a logit response to the others' noisy choices; as λ → ∞ it approaches a Nash equilibrium.
λP(payoff 10)
00.500
0.50.731
10.881
20.982
∞1

Two options, payoffs 10 and 8 (diff 2): P = 1/(1 + e−2λ)

A different model, level-k thinking, limits depth of reasoning rather than adding noise: a level-k player treats the others as level k−1 (Nagel, 1995).

Behavioral Decisions

Quantal Response: Asymmetric Matching Pennies

LeftRight
Top320, 4040, 80
Bottom40, 8080, 40
  • p that makes Column indifferent: 40p + 80(1 − p) = 80p + 40(1 − p) ⇒ p = 1/2.
  • q that makes Row indifferent: 320q + 40(1 − q) = 40q + 80(1 − q) ⇒ q = 1/8.
  • Nash: whatever Row's own 320 is, the probability of Top stays 1/2.
λp = P(Top)q = P(Left)
00.500.50
0.020.830.37
0.10.710.15
10.520.125
∞ (Nash)0.500.125

The values come from solving the two logit equations jointly (by computer). With small λ, Row is drawn to the 320 and plays Top often (0.83). As λ grows, Row accounts for Column's shift to Right and returns to 1/2.

In experiments, Row players clearly shifted toward the big payoff (320) (Goeree and Holt, 2001). QRE captures this own-payoff effect at moderate λ. Common mistake: in a mixed Nash equilibrium a player's probabilities are pinned down by the opponent's payoffs; real people also react to their own.

Social Preferences

Social Preferences: Is It Only About Money?

  • Dictator game: one person splits 10 tokens between herself and a recipient, who can do nothing. Someone who cares only about money gives 0.
  • In a meta-analysis of over a hundred studies the average share given is about 28%; a sizable fraction gives nothing (Engel, 2011).
  • Ultimatum game: once the recipient can reject, offers rise to 40–50% and low offers are rejected.
0%25%50% 0 ≈ 28% Dictator 0–10% 40–50% Ultimatum theory observed (typical share)
In the dictator game the share given comes from fairness or altruism; the extra in the ultimatum game comes from fear of rejection (strategic). Social preference models build these motives into the utility function.

Social Preferences

Inequity Aversion: The Fehr–Schmidt Model

For two people: ui = xi − α·max(xj − xi, 0) − β·max(xi − xj, 0),  0 ≤ β ≤ α, β < 1 (Fehr and Schmidt, 1999)
  • α: the pain of being behind (envy); β: the discomfort of being ahead (guilt). People mind being behind more: α ≥ β.
  • α = β = 0 gives the classic player who cares only about money.
UltimatumA responder offered s out of 10 accepts if s − α(10 − 2s) ≥ 0 ⇔ s ≥ 10α/(1 + 2α). α = 1 ⇒ s ≥ 3.33.
DictatorGiving s ≤ 5 keeps her ahead, so only the β term matters: u = (10 − s) − β·(10 − 2s) = 10 − 10β + s(2β − 1). If β > 1/2 she splits equally; if β < 1/2 she gives nothing; at β = 1/2 every share from 0 to 5 gives the same utility.
Public goodsA sufficiently inequity-averse player contributes if the others do and not otherwise: a conditional cooperator.

Social Preferences

Step by Step: The Ultimatum Acceptance Threshold (α = 1)

0reject = 0s = 1 · (9, 1)s = 2 · (8, 2)s = 3 · (7, 3)s = 4 · (6, 4)s = 5 · (5, 5)

With α = 1 the responder gets u = s − 1·(10 − 2s) by accepting and 0 by rejecting. The proposer offers (10 − s, s).

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)s = 3 · (7, 3)s = 4 · (6, 4)s = 5 · (5, 5)

s = 1: u = 1 − 8 = −7 < 0 ⇒ reject. One token does not make up for the pain of an 8-token gap.

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)−4rejects = 3 · (7, 3)s = 4 · (6, 4)s = 5 · (5, 5)

s = 2: u = 2 − 6 = −4 ⇒ reject.

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)−4rejects = 3 · (7, 3)−1rejects = 4 · (6, 4)s = 5 · (5, 5)

s = 3: u = 3 − 4 = −1 ⇒ still reject (barely).

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)−4rejects = 3 · (7, 3)−1rejects = 4 · (6, 4)2accepts = 5 · (5, 5)

s = 4: u = 4 − 2 = +2 > 0 ⇒ accept.

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)−4rejects = 3 · (7, 3)−1rejects = 4 · (6, 4)2accepts = 5 · (5, 5)5accept

s = 5: u = 5, no gap ⇒ accept.

0reject = 0s = 1 · (9, 1)−7rejects = 2 · (8, 2)−4rejects = 3 · (7, 3)−1rejects = 4 · (6, 4)2accepts = 5 · (5, 5)5acceptthreshold s = 10/3

Threshold 10α/(1 + 2α) = 10/3 ≈ 3.33: the smallest acceptable whole-token offer is 4. A proposer who knows this offers 4 and keeps 6: rejecting low offers reflects a sense of fairness.

Social Preferences

Reciprocity: The Trust Game

  • Trust game (Berg, Dickhaut and McCabe, 1995): the sender sends any part of $10, the amount is tripled, and the receiver returns any part of it. Played one-shot and fully anonymously.
  • Theory (money only): the receiver returns nothing; knowing this, the sender sends 0.
  • In fact senders sent $5.16 on average, and $4.66 came back.
  • Reciprocity: responding to intentions: kindness for kindness, punishment for harm (Rabin, 1993). Inequity aversion looks at the distribution of outcomes, reciprocity at intentions.
Sender10 $ Receiver × 3x3xy back average: x = 5.16 · 3x = 15.48 · y = 4.66
Common mistake: "Trusting paid off." On average the amount returned (4.66) was less than the amount sent (5.16); still, most people sent something.

Collective Decisions

Collective Decision Making: Individual vs Group Interest

  • In group decisions, social influence, conformity pressure and shared interests all play a role.
  • Communities decide by majority, by consensus or by voting rules.
  • Individual interest maximizes one's own utility, collective interest the group's; the two often conflict.
TrafficEveryone takes their own shortest route → total delay is larger than necessary (selfish routing; Pigou's example, the Price of Anarchy).
Work teamsEveryone focuses on their own tasks → the joint project suffers (a public good).
Common resourcesFisheries, pastures, groundwater, the atmosphere → overuse.
The classic name for the last case: the tragedy of the commons (Hardin 1968).

Collective Decisions

The Tragedy of the Commons: A Numerical Model

0508010001,6002,500total animals Gtotal payoff G·(100 − G)

4 herders graze animals on a common pasture. Each animal is worth 100 − G (G = total animals): the more crowded the pasture, the less each animal eats. Herder i's payoff is ui = gi(100 − G).

0508010001,6002,500total animals Gtotal payoff G·(100 − G)optimum: G = 50 → 2,500

Social optimum: maximize G(100 − G) ⇒ G = 50, total 2,500. Shared equally, each herder has 12.5 animals and earns 625.

0508010001,6002,500total animals Gtotal payoff G·(100 − G)optimum: G = 50 → 2,500one deviation: G = 68.75

But while the other three graze 37.5 animals in total (12.5 each), herder 1 earns g·(100 − 37.5 − g). Setting the derivative 62.5 − 2g = 0 gives g = 31.25: a payoff of 976.6 > 625. The others' payoff falls to 390.6.

0508010001,6002,500total animals Gtotal payoff G·(100 − G)optimum: G = 50 → 2,500one deviation: G = 68.75Nash: G = 80 → 1,600

When everyone best-responds: g = (100 − 3g)/2 ⇒ g* = 20, G = 80. Check: with the others at 60, the best response is (100 − 60)/2 = 20. 400 per herder, 1,600 in total. Herders taking turns best-responding also converge to this equilibrium.

0508010001,6002,500total animals Gtotal payoff G·(100 − G)optimum: G = 50 → 2,500one deviation: G = 68.75Nash: G = 80 → 1,600

The tragedy: equilibrium total is 1,600, 64% of the optimum. Each herder feels only 1/4 of the damage his own animal does to the pasture (an externality).

Collective Decisions

More Herders, and Remedies

ng* (per herder)G*totalshare of optimum
150502,500100%
233.3366.72,22288.9%
420801,60064%
109.0990.982633.1%
1000.9999983.9%

In equilibrium g* = 100/(n + 1); as n → ∞, G* → 100 and the total → 0

Pigouvian taxWith a tax, a herder earns g·(100 − G − t); the same computation gives the symmetric equilibrium g = (100 − t)/5. t = 37.5 ⇒ g = 12.5, G = 50.
Quotas / permitsRights to 50 animals in total are allocated and can be traded.
Ostrom (Nobel 2009)Local communities can sustain the resource through monitoring, graduated sanctions and shared rules.

Collective Decisions

State or Market? Ostrom's Third Way

Hardin's remedy: central control or privatization. Ostrom (1990) found shared principles in local arrangements lasting centuries (Nobel 2009).

1. Clear boundariesWho may use which resource is well defined.
2. Rules fit local conditionsRules are tailored to the resource.
3. Collective choiceThose affected by the rules can change them.
4. MonitoringMonitors accountable to the users.
5. Graduated sanctionsLight first penalties, heavier for repeats.
6. Cheap conflict resolutionDisputes are settled quickly and locally.
7. Right to organizeExternal authorities recognize local rules.
8. Nested governanceLayered institutions for large resources.
Ostrom's example, Alanya's inshore fishers: spots are drawn by lottery each season, and boats rotate as the fish migrate.

Collective Decisions

Evaluating Collective Choices: The Veil of Ignorance

Three welfare criteriaThe utilitarian (SW = Σ ui), egalitarian (min ui) and Nash social welfare (Π ui) criteria can rank the same options differently. The commons optimum (G = 50, total 2,500) is the utilitarian best.
Veil of ignorance: a policy is chosen without knowing who will be in which position. Give each position equal probability and maximize expected utility (the Laplace rule) and you get the utilitarian criterion (Harsanyi, 1955); apply maximin and you get the egalitarian one (Rawls, 1971).
ExampleThree positions: P = (10, 6, 2), R = (5, 5, 5). Laplace: P's average 6 > 5 ⇒ P. Maximin: worst case 2 < 5 ⇒ R.
Decision under uncertaintyThe veil turns social choice into a decision theory problem: the same table recommends different policies depending on the uncertainty criterion.

Public Goods and Punishment

Public Goods in the Lab: What Do We See?

  • Each token contributed returns 0.4 to the contributor and 1.6 to the group; not contributing is dominant, and theory predicts zero contributions.
  • Early rounds: about 40–60% of the endowment is contributed (Ledyard, 1995).
  • Decay with repetition: contributions fall round by round and approach free-riding levels in the last rounds.
  • Who? About half the subjects are conditional cooperators (they want to give as much as others do); about a third are free riders (Fischbacher, Gächter and Fehr, 2001).
  • The mechanism of decay: conditional cooperators give slightly below the average, free riders pull the average down: contributions erode a little each round.
others' average contribution 01010equal contributionconditional cooperatorfree rider own contribution (stylized)

Public Goods and Punishment

Code Trace: conditional(c, rounds)

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 0: contributions[0, 10, 10, 10]payoff u = 10 − k + 0.4·ΣkP1: 22 P2: 12

Model: P1 always gives 0; P2–P4 are conditional cooperators: each round they give the others' average from the previous round.

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 0: contributions[0, 10, 10, 10]payoff u = 10 − k + 0.4·ΣkP1: 22 P2: 12

t = 0: for P2 the others' total is 20 ⇒ average 20 / 3 = 6.7 (same for P3, P4).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 1: contributions[0, 6.7, 6.7, 6.7]payoff u = 10 − k + 0.4·ΣkP1: 18 P2: 11.3

Round 1: contributions [0, 6.7, 6.7, 6.7]. The free rider P1 earns more than anyone.

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 1: contributions[0, 6.7, 6.7, 6.7]payoff u = 10 − k + 0.4·ΣkP1: 18 P2: 11.3

t = 1: for P2 the others' total is 13.3 ⇒ average 13.3 / 3 = 4.4 (same for P3, P4).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 2: contributions[0, 4.4, 4.4, 4.4]payoff u = 10 − k + 0.4·ΣkP1: 15.3 P2: 10.9

Round 2: contributions [0, 4.4, 4.4, 4.4]. Each round contributions drop to two thirds of the previous value (a one-third decline).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 2: contributions[0, 4.4, 4.4, 4.4]payoff u = 10 − k + 0.4·ΣkP1: 15.3 P2: 10.9

t = 2: for P2 the others' total is 8.9 ⇒ average 8.9 / 3 = 3.0 (same for P3, P4).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 3: contributions[0, 3.0, 3.0, 3.0]payoff u = 10 − k + 0.4·ΣkP1: 13.6 P2: 10.6

Round 3: contributions [0, 3.0, 3.0, 3.0].

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 3: contributions[0, 3.0, 3.0, 3.0]payoff u = 10 − k + 0.4·ΣkP1: 13.6 P2: 10.6

t = 3: for P2 the others' total is 5.9 ⇒ average 5.9 / 3 = 2.0 (same for P3, P4).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 4: contributions[0, 2.0, 2.0, 2.0]payoff u = 10 − k + 0.4·ΣkP1: 12.4 P2: 10.4

Round 4: contributions [0, 2.0, 2.0, 2.0].

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 4: contributions[0, 2.0, 2.0, 2.0]payoff u = 10 − k + 0.4·ΣkP1: 12.4 P2: 10.4

t = 4: for P2 the others' total is 4.0 ⇒ average 4.0 / 3 = 1.3 (same for P3, P4).

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 5: contributions[0, 1.3, 1.3, 1.3]payoff u = 10 − k + 0.4·ΣkP1: 11.6 P2: 10.3

Round 5: contributions [0, 1.3, 1.3, 1.3].

0510012345roundP2–P4 (conditional)P1 (free rider) = 0round 5: contributions[0, 1.3, 1.3, 1.3]payoff u = 10 − k + 0.4·ΣkP1: 11.6 P2: 10.3

In 5 rounds contributions fall from 10 to 1.3: a single free rider unravels the majority's cooperation (a simple model of the decay seen in the lab).

def conditional(c, rounds):
  history = [c]
  for t in range(rounds):
    new = [0]                   # P1 always free-rides
    for i in range(1, len(c)):
      others = sum(c) - c[i]
      new.append(others / (len(c) - 1))
    c = new
    history.append(c)
  return history

Public Goods and Punishment

With and Without Punishment: Fehr and Gächter (2000)

DesignGroups of 4; each token contributed returns 0.4 to the contributor. After contributions are shown, subjects may pay to punish others. 10 rounds with, 10 without; fixed or reshuffled groups.
Without punishmentContributions start high, then fall toward zero.
With punishmentContributions rise, near full cooperation in fixed groups. The furthest below average are punished most.
People punished even in reshuffled groups, where it gains the punisher nothing: altruistic punishment (Fehr and Gächter, 2002).

Theory: money-only players never punish, so everyone still gives 0. In the lab people punished and contributions stayed high: both predictions failed.

Public Goods and Punishment

Code Trace: Does Punishment Make Free Riding Unprofitable?

01020if all contribute: 160P1 (0)0P2 (10)0P3 (10)0P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3nothing computed yet

Call punish(c, p): P1 free-rides, P2–P4 contribute fully. Each cooperator gives P1 2 punishment points.

01020if all contribute: 1622P1 (0)12P2 (10)12P3 (10)12P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3before punishment

share = 1.6 · 30 / 4 = 12. Before punishment u = [22, 12, 12, 12]: the free rider gets 10 more than the contributors.

01020if all contribute: 1616P1 (0)10P2 (10)12P3 (10)12P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3P2 → P1: 2 points

i = 0 (P1) punishes no one. i = 1: P2 pays 2 for 2 points (12 → 10); P1 loses 2 · 3 = 6 (22 → 16).

01020if all contribute: 1610P1 (0)10P2 (10)10P3 (10)12P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3P3 → P1: 2 points

i = 2: P3 also pays 2 (12 → 10); P1 16 → 10. The free rider's advantage is gone.

01020if all contribute: 164P1 (0)10P2 (10)10P3 (10)10P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3P4 → P1: 2 points

i = 3: P4 also pays 2 (12 → 10); P1 10 → 4.

01020if all contribute: 164P1 (0)10P2 (10)10P3 (10)10P4 (10)contributions c = [0, 10, 10, 10]a punishment point costs 1, removes 3return u

return u: the free rider gets 4; by contributing (no punishment, everyone 10) he would have had 16. If the threat is credible, contributing pays, but each punisher also lost 2.

def punish(c, p):        # p[i][j]: i's points to j
    n = len(c)
    share = 1.6 * sum(c) / n
    u = [10 - c[i] + share for i in range(n)]
    for i in range(n):
        for j in range(n):
            u[i] -= p[i][j]       # costs 1 per point
            u[j] -= 3 * p[i][j]   # target loses 3
    return u

Public Goods and Punishment

Who Will Punish? Second-Order Free Riding

Punishers (2 points each)P1P2P3P4
No one22121212
P2, P3, P44101010
P2, P3 (P4 does not)10101012

Contributions [0, 10, 10, 10]; a punishment point costs 1, removes 3

  • Punishment is itself a public good: the punisher bears the cost, everyone shares the benefit (higher contributions). P4, who does not punish, earns the most.
  • Backward induction: in the last round punishing is pure cost ⇒ no one punishes ⇒ the threat is not credible.
Antisocial punishment

In public goods experiments in 16 cities, including Istanbul, high contributors were also punished in some places, which could wipe out punishment's boost to cooperation. Antisocial punishment was more common where civic norms and the rule of law were weak (Herrmann, Thöni and Gächter, 2008).

The lesson: punishment is an institution, and how it works depends on the norms around it.

Public Goods and Punishment

Back to the Game: Public Goods

Theory's answer: 0 tokens

Each token returns only 1.6/4 = 0.4 to you: whatever others do, keeping it is 0.6 better. Giving 0 is dominant; all 0 → 10 each, all 10 → 16 each.

Scenarioequilibriumper player
Basic010
Punishment (fine 2 if below 3)311.8
Optimum (everyone 10)1016

With punishment, giving 3 costs 3 · 0.6 = 1.8 but avoids the fine of 2: 3 becomes dominant.

In practice

In repeated experiments, contributions start at 40–60% of the endowment and decline (Ledyard, 1995). With costly punishment they stay high (Fehr and Gächter, 2000).

Punishment is an institution that changes incentives, like the commons tax: it moves the equilibrium from 0 to 3, not to the optimum (10).

Threshold Projects

Threshold Projects: Contributing Becomes a Coordination Game

A threshold variant of the public goods game: 4 players, 10 tokens each; if total contributions reach 22, everyone gets 12 extra tokens, otherwise contributions are lost. ui = 10 − ki + 12·[Σk ≥ 22].
Equilibrium 1: nobody contributesOne player can give at most 10, so the threshold of 22 is out of reach: contributing only loses. Everyone gets 10, total 40.
Equilibrium 2: exactly 22Every contributor is pivotal: withdrawing 1 token misses the threshold and loses 12; anything extra is wasted. (6, 6, 5, 5) ⇒ (16, 16, 17, 17), total 66.
Which equilibrium?A contributor loses her tokens if the others do not give: giving 6 drops her to 4. Without trust, the zero equilibrium may prevail.
Who gives how much?The profile (10, 10, 2, 0) is also an equilibrium, with payoffs (12, 12, 20, 22). Threshold equilibria are efficient but may be unequal.

A threshold plus a refund (money returned when a crowdfunding goal is missed) removes the risk of "giving and losing" and makes coordination easier.

Threshold Projects

Step by Step: Which Contribution Profiles Are Equilibria?

05100P1u = 100P2u = 100P3u = 100P4u = 10total contributionthreshold 22Σk = 0 · threshold missedequilibrium: no one can reach 22 alone

(0, 0, 0, 0): if P1 gave 10 alone, the total would be 10 < 22 and her 10 tokens would be lost. No one wants to deviate: this all-zero profile is a Nash equilibrium; everyone gets 10.

05106P1u = 46P2u = 45P3u = 54P4u = 6total contributionthreshold 22Σk = 21 · threshold missednot an equilibrium: P4 adds one token

(6, 6, 5, 4): total 21, threshold missed ⇒ payoffs (4, 4, 5, 6). If P4 gives 5, the total is 22 and she gets 10 − 5 + 12 = 17 > 6: not an equilibrium.

05106P1u = 166P2u = 165P3u = 175P4u = 17total contributionthreshold 22Σk = 22 · threshold reachedequilibrium: every contributor is pivotal

(6, 6, 5, 5): total 22 ⇒ payoffs (16, 16, 17, 17). If P1 drops to 5, the total is 21 and P1 gets 5 < 16. Giving more is wasted: a Nash equilibrium, total 66.

05107P1u = 156P2u = 165P3u = 175P4u = 17total contributionthreshold 22Σk = 23 · threshold exceedednot an equilibrium: P1 withdraws a token

(7, 6, 5, 5): total 23. If P1 drops to 6, the threshold still holds and she gains 1 token: overcontributing is not an equilibrium.

051010P1u = 1210P2u = 122P3u = 200P4u = 22total contributionthreshold 22Σk = 22 · threshold reachedequilibrium, but unequal: (12, 12, 20, 22)

The profile (10, 10, 2, 0) is also an equilibrium (if P3 withdraws, the threshold is missed). There are many efficient equilibria; who gives what is a bargaining and coordination problem.

Information Aggregation

Wisdom of Crowds

  • Galton (1907): at a 1906 livestock fair, 787 guesses of an ox's weight were collected. The median guess was 1,207 lb, the true weight 1,198 lb: an error under 1%.
  • Why? Independent errors cancel out.
  • Diversity prediction theorem (Page, 2007): (average − truth)² = average squared individual error − diversity of the predictions. Diversity is the average squared deviation of the predictions from their own mean.
  • The identity always holds; for a crowd to predict well it needs independence and diversity. If everyone copies everyone else, diversity collapses.
guess xindividual error (x − 50)²diversity (x − 54)²
40100196
452581
55251
6010036
70400256
avg 54130114

Truth 50: crowd error (54 − 50)² = 16 = 130 − 114

The crowd always predicts at least as well as its average member; the greater the diversity, the larger the gain.

Information Aggregation

Information Cascades: Everyone Can Be Rationally Wrong

  • Model (Bikhchandani, Hirshleifer and Welch, 1992): an urn is chosen by a coin flip. In turn, each person privately sees one ball, hears the earlier people's guesses (not their balls) and guesses the urn. Each ball matches its urn's color with probability 2/3.
  • Bayes: each a doubles the odds in favor of A: when a's outnumber b's by f, the odds are 2f to 1, i.e. P(A) = 2f / (2f + 1). On a tie (P(A) = 1/2) a person follows her own signal.
  • Cascade: once the inferred difference is f = 2, a person who sees b has f = 1 and P(A) = 2/3 > 1/2: she ignores her own signal and says A. Those after her learn nothing from her guess.
  • Fragility: a cascade rests on just two signals; a single piece of public information can break it. Fashion, product reviews, panics.
urn A a a burn B b b a each draw: correct signal with probability 2/3

In the lab (Anderson and Holt, 1997), cascades in this urn experiment did form in most of the situations where they could.

Information Aggregation

Code Trace: How Does a Cascade Start?

true urn: Bsignalguessperson 1person 2person 3person 4person 5person 6f = 0no cascade

Call cascade(signals); the true urn is B. The difference inferred from earlier guesses is f = 0.

true urn: Bsignalguessperson 1aAperson 2person 3person 4person 5person 6f = 1no cascade

Person 1, signal a: f = 0 ⇒ follows her signal, A. The guess reveals the signal: f = 1, P(A) = 2/3.

true urn: Bsignalguessperson 1aAperson 2aAperson 3person 4person 5person 6f = 2cascade

Person 2, signal a: f = 1 ⇒ with her own a, P(A) = 0.8, A. f = 2.

true urn: Bsignalguessperson 1aAperson 2aAperson 3bAignoredperson 4person 5person 6f = 2cascade

Person 3, signal b: f = 2 ⇒ even with her b, P(A) = 2/3 ⇒ A. The cascade has started; f no longer changes.

true urn: Bsignalguessperson 1aAperson 2aAperson 3bAignoredperson 4bAignoredperson 5person 6f = 2cascade

Person 4, signal b: same computation ⇒ A. Her guess carries no information either.

true urn: Bsignalguessperson 1aAperson 2aAperson 3bAignoredperson 4bAignoredperson 5bAignoredperson 6f = 2cascade

Person 5, signal b ⇒ A.

true urn: Bsignalguessperson 1aAperson 2aAperson 3bAignoredperson 4bAignoredperson 5bAignoredperson 6bAignoredf = 2cascade

Person 6, signal b ⇒ A.

true urn: Bsignalguessperson 1aAperson 2aAperson 3bAignoredperson 4bAignoredperson 5bAignoredperson 6bAignoredf = 2cascade

return: everyone said A, yet 4 of the 6 signals were b. Pooling the signals would give f = −2, P(A) = 0.2 ⇒ B. Everyone was rational; the information did not accumulate.

def cascade(signals):
    f = 0                  # a − b inferred from guesses
    guesses = []
    for s in signals:
        if f >= 2: t = 'A'     # A cascade
        elif f <= -2: t = 'B'  # B cascade
        else:
            t = 'A' if s == 'a' else 'B'
            f += 1 if s == 'a' else -1
        guesses.append(t)
    return guesses

Information Aggregation

The Condorcet Jury Theorem: How Smart Is the Majority?

  • Question: there is a right answer (is the defendant guilty or not?). How does majority voting change the chance of finding it? This is different from the Condorcet paradox, where preferences cycle.
  • Theorem (Condorcet, 1785): if n (odd) voters vote independently and each picks the right answer with probability p > 1/2, the probability that the majority is right increases with n and tends to 1.
  • If p < 1/2 it is the reverse: as the crowd grows, the majority is almost surely wrong.
  • Pn = Σk > n/2 C(n, k)·pk(1 − p)n − k
np = 0.6p = 0.4
10.6000.400
30.6480.352
50.6830.317
110.7530.247
510.9260.074
1010.9790.021

Probability that the majority decides correctly

Information Aggregation

Code Trace: A Five-Member Jury

0.010k = 00.077k = 10.230k = 20.346k = 30.259k = 40.078k = 5total = P(majority correct)one voter 0.6total = 0.0000

Call jury(5, 0.6): the majority is right if at least 3 of the 5 voters vote correctly. k: number of correct votes.

0.010k = 00.077k = 10.230k = 20.346k = 30.259k = 40.078k = 5total = P(majority correct)one voter 0.6total = 0.3456

k = 3: C(5, 3)·0.63·0.42 = 10·0.2160·0.1600 = 0.3456. Total 0.3456.

0.010k = 00.077k = 10.230k = 20.346k = 30.259k = 40.078k = 5total = P(majority correct)one voter 0.6total = 0.6048

k = 4: C(5, 4)·0.64·0.41 = 5·0.1296·0.4000 = 0.2592. Total 0.6048.

0.010k = 00.077k = 10.230k = 20.346k = 30.259k = 40.078k = 5total = P(majority correct)one voter 0.6total = 0.6826

k = 5: C(5, 5)·0.65·0.40 = 1·0.0778·1 = 0.0778. Total 0.6826.

0.010k = 00.077k = 10.230k = 20.346k = 30.259k = 40.078k = 5total = P(majority correct)one voter 0.6total = 0.6826

return: P5 = 0.6826 > 0.6. Five independent, moderately competent voters decide better than one.

from math import comb
def jury(n, p):
    total = 0
    for k in range(n // 2 + 1, n + 1):
        term = comb(n, k) * p**k * (1 - p)**(n - k)
        total += term
    return total
print(round(jury(5, 0.6), 4))

Information Aggregation

When Does the Jury Theorem Break Down?

DependenceIf 4 jurors on a 5-member jury follow an opinion leader, the majority is that one person's vote: P = 0.6. With independent votes it was 0.683. Cascades and shared information sources destroy independence.
CompetenceOn a question where p < 1/2 (a misleading clue), a larger crowd makes the error certain: with p = 0.4 and 101 voters, the majority is right with probability only 2.1%.
Strategic votingA voter who reasons that her vote matters only when it is pivotal may rationally vote against her own signal (Austen-Smith and Banks, 1996). Under unanimity this can raise the chance of convicting the innocent (Feddersen and Pesendorfer, 1998).
The wisdom of crowds, cascades and the jury theorem teach the same lesson: a group knows only as much as it can pool of its members' independent information.

Summary

Common Mistakes: Collective Decisions

"Crowds are always wise."Independence and p > 1/2 are needed. In the cascade, all six people were wrong.
"Those who follow a cascade are irrational."Each acted correctly by Bayes' rule; the problem is that information does not accumulate.
"Punishment always increases cooperation."Punishment is costly and invites second-order free riding; antisocial punishment can cancel its effect.
"A threshold project has a single equilibrium."Both zero contribution and every profile summing to exactly 22 are equilibria.
"The commons needs either the state or the market."Ostrom's local institutions are a third way.
"The jury theorem and the Condorcet paradox are the same."The theorem is about finding the truth, the paradox about cyclic preferences.

Summary

Summary

Prospect theoryReference point, λ ≈ 2.25, probability weighting: the ±100 TL bet is rejected, the acceptance threshold is ≈ 251 TL; the endowment effect and framing share the same root.
Boundedly rational equilibriumQuantal response: better options are chosen more often; in asymmetric matching pennies, play shifts toward the big payoff instead of Nash's 1/2.
Social preferencesFehr–Schmidt: a responder with α = 1 rejects offers below 10/3 out of 10; reciprocity in the trust game.
The commonsWith 4 herders, the equilibrium total is 1,600 vs an optimum of 2,500; taxes, quotas and Ostrom's local institutions.
Public goods and punishmentContributions start at 40–60% and erode; punishment sustains cooperation but can breed second-order free riding and antisocial punishment.
Information aggregationThe diversity theorem, information failing to accumulate in a cascade, the jury theorem (5 voters, p = 0.6 ⇒ 0.683).

Summary

Sample Problems

  1. With the prospect theory parameters, is a +300 / −150 coin-flip bet accepted?
  2. In the commons model with n = 9 herders, what are the equilibrium G* and the total payoff? What tax achieves the optimum?
  3. In the punishment variant of the public goods game, what would the equilibrium contribution be if the fine were 1 token instead of 2?
  4. In the Fehr–Schmidt model, at least how many of 10 tokens does a responder with α = 2 demand?
  5. In the Condorcet jury theorem with p = 0.7, with what probability is the majority of a 3-member jury right?
Show answers
  1. ½·3000.88 − ½·2.25·1500.88 ≈ −16.8 < 0 ⇒ rejected (threshold ≈ 377 TL).
  2. g* = 10, G* = 90, total 90·10 = 900; tax: (100 − t)/10 = 50/9 ⇒ t = 44.4.
  3. Contributing 3 tokens costs 1.8 but the fine is only 1: whatever the others give, not contributing is better by 0.8; the equilibrium contribution is 0.
  4. s ≥ 10·2/(1 + 4) = 4 ⇒ at least 4 tokens.
  5. P3 = 0.73 + 3·0.72·0.3 = 0.343 + 0.441 = 0.784.

Summary

References

  • H. A. Simon, “A behavioral model of rational choice”, Quarterly J. Economics 69, 1955.
  • D. Kahneman, A. Tversky, “Prospect theory: An analysis of decision under risk”, Econometrica 47(2), 1979.
  • A. Tversky, D. Kahneman, “The framing of decisions and the psychology of choice”, Science 211, 1981; “Advances in prospect theory”, J. Risk and Uncertainty 5, 1992.
  • D. Kahneman, J. L. Knetsch, R. H. Thaler, “Experimental tests of the endowment effect and the Coase theorem”, J. Political Economy 98(6), 1990.
  • R. D. McKelvey, T. R. Palfrey, “Quantal response equilibria for normal form games”, Games and Economic Behavior 10(1), 1995; J. K. Goeree, C. A. Holt, “Ten little treasures of game theory and ten intuitive contradictions”, American Economic Review 91(5), 2001.
  • E. Fehr, K. M. Schmidt, “A theory of fairness, competition, and cooperation”, Quarterly J. Economics 114(3), 1999; C. Engel, “Dictator games: A meta study”, Experimental Economics 14(4), 2011.
  • J. Berg, J. Dickhaut, K. McCabe, “Trust, reciprocity, and social history”, Games and Economic Behavior 10(1), 1995; M. Rabin, “Incorporating fairness into game theory and economics”, American Economic Review 83(5), 1993.
  • G. Hardin, “The tragedy of the commons”, Science 162, 1968; E. Ostrom, Governing the Commons, Cambridge University Press, 1990.

Summary

References (2/2)

  • J. O. Ledyard, “Public goods: A survey of experimental research”, in J. H. Kagel, A. E. Roth (eds.), The Handbook of Experimental Economics, Princeton University Press, 1995.
  • U. Fischbacher, S. Gächter, E. Fehr, “Are people conditionally cooperative? Evidence from a public goods experiment”, Economics Letters 71(3), 2001.
  • E. Fehr, S. Gächter, “Cooperation and punishment in public goods experiments”, American Economic Review 90(4), 2000; “Altruistic punishment in humans”, Nature 415, 2002.
  • B. Herrmann, C. Thöni, S. Gächter, “Antisocial punishment across societies”, Science 319, 2008.
  • F. Galton, “Vox populi”, Nature 75, 1907; S. E. Page, The Difference, Princeton University Press, 2007.
  • S. Bikhchandani, D. Hirshleifer, I. Welch, “A theory of fads, fashion, custom, and cultural change as informational cascades”, J. Political Economy 100(5), 1992; L. R. Anderson, C. A. Holt, “Information cascades in the laboratory”, American Economic Review 87(5), 1997.
  • Marquis de Condorcet, Essai sur l'application de l'analyse à la probabilité des décisions rendues à la pluralité des voix, 1785; D. Austen-Smith, J. S. Banks, “Information aggregation, rationality, and the Condorcet jury theorem”, American Political Science Review 90(1), 1996; T. Feddersen, W. Pesendorfer, “Convicting the innocent”, American Political Science Review 92(1), 1998.
  • R. Nagel, “Unraveling in guessing games: An experimental study”, American Economic Review 85(5), 1995.
  • J. C. Harsanyi, “Cardinal welfare, individualistic ethics, and interpersonal comparisons of utility”, J. Political Economy 63, 1955; J. Rawls, A Theory of Justice, Harvard University Press, 1971.

The End

Chapter 12: Behavioral and Collective Decisions
Previous: Chapter 11: Algorithmic Fairness and Ethics · Next: Chapter 13: Learning and Adaptation
1 / 1 Sercan KÜLCÜ. All rights reserved.