Update: Reciprocity in Groups and Third Party Punishment

1
Update Reciprocity in Groups and Third Party
Punishment
Robert Kurzban
University of Pennsylvania
Hokkaido University 8 Nov 2006
2
Roadmap

• Public Goods Work
• Theories in the spotlight
• Third Party Punishment
• Directions

3
Remember this? Real Time Public Goods Game
4
50
Low Info Increase/Decrease
Low Info Increase Only
40
High Info Increase/Decrease
High Info Increase Only
30
Average Contribution (Tokens)
20
10
0
1
2
3
4
5
6
7
8
9
10
Round
5
Well, it should look familiar
6
Replication in JapanDynamics
U.S. Data
Japan Data (Ishii Kurzban)
7
Contributions by Round in the Increase Only/Low
Information Condition
8
New Questions

• Are there types? Can this explain both the
  upward and downward spirals?
• Can we get more specific about reciprocal
  players?
• Median matching?
• Minimum reciprocity?

9
Circular Game Method(Kurzban Houser, PNAS,
2005)

• Circular Public Goods game
• Players make initial contribution
• Players, in turn, observe aggregate contribution
  of other players
• After observing this value, player may update
  their own contribution
• Round ends with p .04 each update
• This allows us to plot a contribution profile
  for each player (CP)

10
Individual Differences

• This method allows us to plot a contribution
  profile for each player (CP)
• Regress contribution on information observed.
• This gives an intercept and slope.
• Intercept how much player i contributes when
  others arent contributing much
• Slope player is responsiveness to others
  contributions

11
Individual Differences

• Free Rider CP everywhere below 25 (1/2)
• 20 of sample (N 84)
• Cooperator CP everywhere above 25
• 13
• Recriprocator positive slope, and CP is both
  above and below the 50 line.
• 63
• Small percentage unclassifiable

12
Individual Differences

• We use some rounds to see if typing scheme
  captures something stable.
• If so, we should be able to predict (in a
  hold-out sample) the dynamics of play given the
  makeup of the constituted groups.
• Groups are assigned a Cooperativeness Score, 2
  for a Cooperator, 1 for a Reciprocator, 0 for a
  Free Rider

13
(No Transcript)
14
(No Transcript)
15

• Types fit reasonably into a 3-part system
• Payoffs did not vary as a function of type.
• Suggest individual differences in strategies?

16
Information Seeking

• Method
• Circular game, but allow players to observe one
  piece of information (low, median, high) before
  making their own contribution decision.
• Other parameters as in Experiment 2
• One Independent Variable This information is
  either free, or costly (2 tokens)

17
Information Seeking

• Hypotheses
• IF players know others respond to own
  contribution, costly information should decrease
  contributions.
• IF players have reciprocal (type) preferences,
  they will have systematic preferences for
  information and will pay to observe it.
• Type (reciprocator, free rider) will predict
  information-seeking preferences

18
Results
19
Results Information-Seeking
Free Info () Costly Info () (conditional on paying)
Low 35 11 (23)
Median 35 21 (46)
High 30 14 (30)
None na 54 (n/a)
20
Results Information-SeekingIndividual
Differences

• Regress subjects' contribution amounts on
  contributions seen.
• Reciprocity Index (RI) slope, how much i is
  influenced by others contributions.
• Altruism Index (AI) is the y-intercept is
  contribution when others contribution 0
• Free-riding Index (FI) is contribution when
  contribution seen equals 50 (subtracted from 50
  -- high values identify free-riding.)

21
Results Information-SeekingIndividual
Differences
(non-randomly chosen) examples of typing
regression for 3 ss,
22
Results Information-SeekingIndividual
Differences
More reciprocal players like median information
23
Results Information-SeekingIndividual
Differences
Free Riders like high information
24
Results Information-SeekingIndividual
Differences
25
Information Seeking Results

• Types max (AI, FI, RI50). In the Free
  Information condition, payoffs did not vary as a
  function of type.
• In the Costly Information condition, Free
  Riders did better than Reciprocators or
  Altruists.

26
Experiment 3 Conclusions

• There is a tendency to prefer observing the
  MEDIAN current contributor. (oops)
• People will endure costs to observe others
  decisions.
• Reciprocators tend to look at the median (Croson
  1998)
• In contrast
• Free Riding types tend to look at the high
  information.
• Altruistic types dont have clear preferences

27
Part II Third Party Punishment
28
Third Party Punishment

• If A violates a norm, for example, A cheats B,
  people (C) seem to express a preference for
  punishing A.
• There is, however, substantial debate about the
  scope of the phenomenon, as well as its
  evolutionary explanation.

29
Third Party Punishment ? Second Party Punishment

• If A cheats B, B has a preference for inflicting
  costs on A.
• Substantial evidence from field and lab
• Trivers (1971) theory of reciprocal altruism
  provides one possible explanation for this
  phenomenon.

30
Third Party Punishment

• A puzzle from either the standpoint of evolution
  or the canonical economic view.
• Letting others endure costs of punishment would
  seem to be a good strategy.
• Why pay costs of punishing is the underlying
  question.

31
Punishment ( negative reciprocity)
Cooperation in Groups

• Cultural group selection (Boyd et al., 2003,
  PNAS)
• Groups with those with such a taste do better
  because they give incentives to others in the
  group to be pro-social.
• Strong Reciprocity (Gintis, 2000, JTB)
• Groups with punishers to better than those
  without.
• Inequity aversion driven by reduction of fitness
  differentials (Price et al., 2003, EHB).

32
3rd Party Punishment

• On some (recent) models, signaling that one
  punishes norm violators or, more narrowly, those
  who defect, leads to benefits through
  reputational processes.
• e.g., Indirect reciprocity (Panchanathan and
  Boyd, 2004, Nature).
• Signaling models (E A Smith, etc.)

33
Comparing models

• So.
• Some models dont specifically predict
  sensitivity to audience effects (though such
  effects dont rule out MLS)
• To the extent that 3rd party punishment is
  sensitive to cues to the presence of an audience,
  this implies a history of selection associated
  with reputation effects.

34
Experiments showing effects of blinding and
social distance

• Dictator Games
• Dictator game as social distance decreases,
  altruism increases. (Hoffman et al., 1996)
• Public goods games
• Buchan et al. Personal communication
• Ultimatum games
• Bolton Zwick. Anonymity has VERY LIMITED
  effects on rejecting unequal offers.

35
Method (cont)

• Part I
• Trust game each DM1 plays 5 games, paired with
  a different DM2
• Part II
• New Ss can punish (bad) DM2s
• Part III
• Participants from Part I return to collect their
  money.

36
Current Study Method

• Part I
• Trust game each DM1 plays 5 games, paired with
  a different DM2

37
Part I Stimuli
Values 1 / 39 3 / 37 6 / 34 9 / 31 12 / 28
38
Method (part II)

• Players given 3 show-up payment
• Players given 7 to punish DM2s in the game in
  which result was 1/39
• Two conditions
• Anonymous elaborate envelope technique
• Non-anonymous one experimenter knows how much
  of 7 used to punish DM2

39
Conditions

• Anonymous condition
• Measures punishment due to tastes
• Non-zero punishment implies some taste for
  punishment.
• Non-anonymous condition
• Measures punishment due to tastes PLUS
  punishment due to knowledge that punishment is
  observed.
• Significantly greater punishment implies
  computation associated with others knowledge.

40
Results, part I Trust Games
One untrustworthy DM2 at 1/39
Game Tree Proportion of DM1s Moving Down (trust) Proportion of DM2s moving Down (trustworthy)
1/39 1/7 0/1
3/37 1/7 0/1
6/34 1/7 0/1
9/31 4/7 1/4
12/28 6/7 3/6
N 14. All remaining DM1 moves (22) were 10,10
41
Results, Part II Punishment
Subjects do the funniest things 1
Subject changed Treatment
t(41) 2.87, p lt .01. means anonymous .58,
observed 2.42. Better test
Kolmogorov-Smirnov, J 1.37, p lt .05
42
Results

• People in the observed condition punished (four
  times) more than those in the anonymous
  condition.
• Punishment in the anonymous condition was small,
  0.58/7.00.

43
Experiment 3b

• Like Experiment 3a, only PD with labeled
  extensive form game.

44
Method (cont)

• Participants received a game piece from Stage 1
  in which DM1 had played C and DM2 played D
• Participants could pay 0-10 to deduct a tripled
  amount from that DM2.

45
Method IVs

• 3 conditions
• Anonymous elaborate envelope technique
• Non-anonymous one experimenter knows how much
  of 7 used to punish DM2
• Participants punishment decisions were revealed
  to both the experimenters and all other
  participants.

46
Results Stage 1
47
Results Stage 2
Subjects do the funniest things 2
Some subjects announced Cooperate, Cooperate.
ns

48
Research Agenda
3PP to 4PO Emotions
Cross-cultural Replications
Developmental
Vectors Strategy Method in a PGG
Consensus on Punishment
49
Acknowledgements

• Collaborators
• Alex Chavez
• Peter DeScioli
• Dan Houser
• Keiko Ishii
• Kevin McCabe
• Erin OBrien
• Vernon Smith
• Bart Wilson

• Funding
• University of Pennsylvania Research Foundation
• University of Pennsylvania University Scholars
• MacArthur Foundation
• Japan Society for the Promotion of Science

50
Thank You
51
NOTE The issue is the design of the mechanism,
the cues that people respond to.

• In a game in which people have a decision to
  cooperate (or not) (e.g., Burnham Hare)
• The test treatment adds a pair of human eyes to
  the control environmentthe evolutionary legacy
  hypothesis suggests that the test treatment,
  although actually still private with regard to
  other subjects, will be perceived as public
  (by the modular system in question)

52
(No Transcript)
53
(No Transcript)
54
Results Stage 2
55
(No Transcript)
56
(No Transcript)
57
Fessler Haley
Only the desktop background on participants
computers varied In the Eyespots condition,
players used computers displaying two stylized
eye-like shapes along with familiar desktop
icons In the Control condition, the word
CASSEL was displayed across the same portion of
the screen
58
(No Transcript)