Download - Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Transcript
Page 1: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Ariely, Loewenstein, Prelec (2003)

Page 2: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Anchoring and other Effects

• Anchoring: initial piece of (irrelevant) information influences

subsequent judgment/choices/valuation.

• Social Learning or Conformism:

• Initial piece of (relevant) information influences

subsequent behavior.

• E.g., what computer to buy, how much to tip.

• Can be rational (and be modeled as such).

• Attraction Effect:

– Adding an (asymmetrically) dominated product to a

choice menu changes people‘s behavior.

Page 3: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

The Attraction Effect

• Note: attraction effect is often referred to as ‚asymmetric

dominance effect‘.

Page 4: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

The Attraction Effect

• With two options, most people go for the cheaper alternative.

• But with the dominated option included, people go for the

expensive alternative.

Page 5: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Experimental Economics

Lecture 3: Bounded Rationality

Dorothea Kübler, Roel van Veldhuizen

Summer term 2015

5

Page 6: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

6

Deviations from the standard

model (Rabin 2002) Individual i at time t = 0 maximizes expected utility

subject to a probability distribution p(s) of the states of

the world s∈S:

max Σt=0

∞ δt Σst∈St p(st) U( xit | st)

xit∈Xi

Possible deviations from the standard model:

- Non-standard preferences (e.g., time, risk and social

pref.)

- Non-standard beliefs (e.g., overconfidence, non-Bayesian

updating)

- Non-standard decision making

Page 7: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Non-Standard Decision Making

• Heuristics and biases:

• Heuristic: short-cut to solving decision problem.

• Last week; Ariely, Loewenstein, Prelec (2003).

• Other examples: framing, nudging.

• System 1 vs System 2 cognition (Kahnemann, 2002).

• (Emotions).

• This week: bounded rationality

• Even when our judgment is not clouded by our biology

(e.g., heuristics), we may still not act fully rationally.

• This lecture: focus on games/strategic reasoning.

Page 8: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• Sometimes making the optimal decision is impossible

Page 9: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• We know from Nash that Chess, Hearthstone and Go have

an equilibrium.

• But no person/computer has been able to calculate it.

Page 10: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• For firms selling natural resources (oil, coal, wine?), life

can be very complicated:

• Salo and Tahvonen (2001) can solve this problem, but

most people probably cannot

Page 11: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• For firms selling natural resources (oil, coal, wine?), life

can be very complicated:

• Salo and Tahvonen (2001) can solve this problem, but

most people probably cannot.

Page 12: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• Rational agents always make the optimal (rational) decision

• Real people may not be able to make the optimal decision

• If no one/only experts can compute it

• If it takes too much time/effort to compute

• If we are inattentive/make mistakes

• Etc.

• If people are as smart as rational agents...

• Then why do you need to be taught how to solve models?

Page 13: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

But if not optimal, then what?

• Problem (for economists): there is only one way to be smart,

but there are infinitely many ways to be stupid (irrational).

• Behavioral economics attempts to identify robust,

widespread patterns of irrationality.

• And then create models incorporating these irrationalities.

• The goal of behavioral models:

• Predict/explain behavior better than rational model

• In a large number of settings

• In a parsimonious way (not just extra parameters)

• Behaviorally realistic

Page 14: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

An experiment here and now

• Let‘s play a simple game. No talking with one another.

• Your decision is to choose a whole number X between 0

and 100.

• The person whose number X is closest to 2/3 times the

mean of all the numbers submitted, will win a special

prize.

• Please write your number down on a piece of paper as

follows: X = _.

• Please put your intials on the card. Do not reveal your

choices to any others.

Page 15: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Beauty Contest Game

• N Players must choose a number between 0 and 100 inclusive.

• The person(s) whose number choice Xi is closest in absolute value

to

p/N ΣNi=1 Xi

wins a fixed prize, all others earn 0.

• p is some value different from 1, e.g. 1/2,2/3,4/3

• Dominated is any Xi > 100p when p < 1.

• If p < 1, iterated elimination of dominated strategies leads to the

choice of 0 as the unique equilibrium outcome.

• Experimental design (p ǂ1) allows for detection of individual steps of

iterated dominance.

Page 16: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Beauty Contest Game

• Example of solving for the equilibrium for p=2/3.

• Winning number X* equals 2/3 times the average Xav

• Step 1: X*<66.67. Even if Xav is 100 (the maximum), the best

response would be to select 66.67. Therefore no one should ever

choose X>66.67

• Step 2: Given Xav<66.67, X*<44.44

• Step 3: Given Xav<44.44, X*<29.63

• (...)

• Step ...: Given Xav<.00001, X*=0

Page 17: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Beauty Contest Game

• N Players must choose a number between 0 and 100 inclusive.

• The person(s) whose number choice Xi is closest in absolute value

to

p/N ΣNi=1 Xi

wins a fixed prize, all others earn 0.

• p is some value different from 1, e.g. 1/2,2/3,4/3

• Dominated is any Xi > 100p when p < 1.

• If p < 1, iterated elimination of dominated strategies leads to the

choice of 0 as the unique equilibrium outcome.

• Experimental design (p ǂ1) allows for detection of individual steps of

iterated dominance.

Page 18: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Nagel AER 1995

• N = 15 - 18.

• Prize to winner is 20 DM.

• p = 1/2, 2/3 or 4/3.

• Nash Equilibrium is 0 in first two cases, 100 in the

second following an infinite number of steps of iterated

elimination of dominant strategies (fixed point solution).

Page 19: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Nagel‘s findings

• Clearly, the Nash Equilibrium (0) is a poor predictor of behavior.

Page 20: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Level-K model

• Non-equilibrium behavioral model first proposed by Stahl and Wilson (1995)

and Nagel (1995) with further extensions by Ho, Camerer and Weigelt

(1998), Costa-Gomes, Crawford and Broseta (2001) and Costa-Gomes and

Crawford (2006).

• The general idea is that people differ in their level of strategic sophistication

ranging from level 0 (clueless) to level ∞ (supersmart).

• Level 0, corresponds to non-strategic behavior wherein strategies are

selected at random.

• Level 1 players, L1, believe that all their opponents are L0 and play a best

response to this belief.

• Level 2 players, L2, play a best response to the belief that all their

opponents are L1, etc.

Page 21: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Beauty Contest: Level-K

• Example of Level-K model for p=2/3.

• Winning number X* equals 2/3 times the average Xav

• Level 0: randomizes uniformly on the interval between 0 and 100.

Hence the average Level 0 quantity (X0) will be 50.

• Level 1: assumes everyone else level 0, plays best response.

X1=2/3*50=33.33

• Level 2: assumes everyone else level 1, plays best response.

• X2=2/3*33.33=22.22

• Level 3: X3=2/3*22.22=14.81

Page 22: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Nagel‘s findings

• The Level-K model seems to do better (though...)

• And how plausible is level-K really?

Page 23: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Cognitive Hierarchy Model

• Extension of Level-K model by Camerer, Ho and Chong (2004).

• Level 0 behaves randomly (as in level K) – X0=50

• Level 1 best responds to level 0 (as in level K) – X1=50*2/3=33.33

• Level 2 best responds to a mix of level 1 and level 0 – Suppose L2 thinks 70% is L1 and 30% is L0

– Then X2=2/3*(.7*33.33+.3*50)=25.55

• Level 3 best responds to a mix of level 0, 1 and 2 – Suppose L3 thinks 35% is L1 and 15% is L0 and 50% is L2

– Then X3=2/3*(.35*33.33+.15*50+.5*25.55)=21.29

Page 24: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Comparison of Models

• Nash equilibrium most parsimonious model (point prediction)

– But the prediction is not borne out by the data

• Level-K fits/predicts better, but maybe not very realistic

• Cognitive hierarchy is the most realistic model

– But has additional free parameter(s): the distribution of k-levels in the population.

– And even for given level, the prediction depends on (assumed) distribution.

Level/Step Nash Eq. Dom. Solv Level-K Cognitive Hierarchy

0 50 50

1 67 33 33

2 44 22 26 (?)

3 29 15 21 (?)

...

∞ 0 0 0 16 (?)

Page 25: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

More models of irrationality

• The level-K and cognitive hierarchy model assume that:

– People have non-equilibrium beliefs (e.g., all others are level 2)

– But they then best respond, given those beliefs.

• Another approach:

– Assume people have equilibrium beliefs

– But may not always best respond (exactly).

• Quantal Response Equilibrium

– Assumes that people sometimes make small mistakes in best responding

– Costly mistakes less likely than non costly mistakes.

– People have correct beliefs about (noisy) behavior of others.

– Equilibrium model: players (noisily) best respond to other players‘ (noisy)

behavior.

– Goeree and Holt (2001): Ten little treasures.

Page 26: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

More models of irrationality

• People differ in important ways, they are heterogeneous: – Different levels of strategic reasoning?

– More or less likely to make mistakes?

– Or even use different strategies altogether?

• Can be modeled and tested experimentally (as in Nagel, 95).

• Problem: same outcome can often be result of different choice

processes

– E.g., in beauty contest 32 may be L0 or L2 in level K, or may be a

higher level in Cognitive Hierarchy model.

• Use experiments that can directly obtain strategies.

Page 27: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Linde, Sonnemans, Tuinstra (2013)

• Experiment on the minority game.

• Each player is asked to choose either Red or Blue.

• Groups of five.

• You win if you are part of the minority.

• For example, in RRBBB, the R-players are the winners.

Page 28: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Linde, Sonnemans, Tuinstra (2013)

• Experiment on the minority game.

• Each player is asked to choose either Red or Blue.

• Groups of five.

• You win if you are part of the minority.

• For example, in RRBBB, the R-players are the winners.

• Mixed equilibrium: randomize with equal probability.

• Repeated game: more involved, but still suboptimal equilibria.

• Experimental results suggest people behave close to mixed equilibrium.

• But do people truly randomize, or do they use more sophisticated strategies

that resemble the mixed equilibrium in the average?

Page 29: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Strategy Method

• Strategy method: rather than choosing actions (B or R), people choose a

strategy, conditional on previous events.

Page 30: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Overview of the Experiment

• 42 Students from multidisciplinary bachelor in Amsterdam.

• Five week course on evolutionary dynamics and game theory.

• Week 1: students come to the lab and program their simulations.

• Computer simulations for all possible combinations of strategies.

• Top 5 get paid 150,120,90,60 and 30 Euros respectively.

• Week 2-5: students submit new strategies online.

• Before submitting, could try out strategies against previous week.

Page 31: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Results

• Average behavior and outcomes very close to Mixed Equilibrium.

• But actual strategies followed very different.

Page 32: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Winning Strategies

• Winning strategies differ strongly from the Mixed Equilibrium.

Page 33: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Evolutionary Simulations

• In the long run, only 4/107 strategies survive.

Page 34: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Evolutionary Simulations

• The surviving strategies switch less often and score more points.

Page 35: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Evolutionary Simulations

• The four surviving long run strategies are very different from Mixed Equilibrium.

• First strategy forms the largest part of the population.

Page 36: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

How important is bounded

rationality in practice?

• Bounded rationality/mistakes may cancel out on average: – But even if they do, is it not still important/interesting to investigate?

– And: systematic deviations may affect the market equilibrium.

• Irrational individuals may learn from their mistakes:

– But: people are not always very good at learning.

– Learning/evolution of strategies does not always lead to equilibrium.

• Irrational individuals will perform worse and be driven from the market:

– Insufficient feedback to change behavior.

– Non-rational behavior may be (ex post) very profitable.

• The influence of irrationality may depend on structure of the market – Substitutability (e.g., Cournot): irrational mistake compensated by rational agent

– Complementarity (e.g., Bertrand): irrational mistake amplified by rational agent

– Fehr and Tyran (2005)

Page 37: Ariely, Loewenstein, Prelec (2003) - TU Berlin · • Last week; Ariely, Loewenstein, Prelec (2003). • Other examples: framing, nudging. • System 1 vs System 2 cognition (Kahnemann,

Bounded Rationality

• Rational agents always make the optimal decision.

• Real world agents may not:

– Heuristics and biases.

– Limited reasoning (Level-K/Cognitive Hierarchy).

– Mistakes.

• Real world agents are also heterogeneous:

– In their level of strategic reasoning (beauty contest game).

– In the type or frequency of mistakes. (Goeree and Holt, 2001)

– In the types and sophistication of the strategies they use (minority game).

• Good models in behavioral economics:

– Predict/explain behavior well in large number of settings

– Are parsimonious

– Are behaviorally realistic