社会的交換における 評判情報の信頼性

Reputational Information in “Noisy” Interactions:
Experimental Studies
Motoki Watabe (Waseda University, JAPAN)
Contributors: Yoshitsugu Yamamoto and Sho Tsuboi
What I want to say
 Unconditional Cooperation strategy can produce more
reliable reputational information than reciprocal strategies in
noisy exchange.
A Model of Reputational Information
B
A
EXCHANGE
Observe
Person A and B exchange their
resources repeatedly.
≒ They play PD repeatedly.
Impression
X
Person A’s reputation
1. Source 1: A’s evaluation on him/herself
2. Source 2: B’s evaluation on A
3. Source 3: X’s evaluation on A
Reliability of Reputational Information
Person A’s reputation
B
A
EXCHANGE
Observe
Impressi
on
X
1. Source 1: A’s evaluation on him/herself
2. Source 2: B’s evaluation on A
3. Source 3: X’s evaluation on A
•Past research assume that reputational
information is always objective and reliable.
•Source 1 = Source 2 = Source 3
•Intuitively,
•Source 1 < Source 2 < Source 3
•Source 3 seems to be the most objective and reliable.
•However, the reliability of source 3 is not always guaranteed.
Outline of the presentation
People have a bias of reputational information
processing.
1.

Experiment 1
Reciprocal strategies reduce reliability of
reputational in social exchange.
3. Unconditional strategies can produce reliable
reputational information.
2.

Experiment 2 & 3
A Model of Reputational Information
B
Person A and B exchange their
resources repeatedly.
≒ They play PD repeatedly.
A
EXCHANGE
Observe
Impression
X
X’s impressions on A and B
are reputational information
to distribute other people.
A is a person like @@@@@
B seems to be*****
Two functions of reputation(Raub 2002,
Yamagishi 1998)
Control: In order to get future exchange partners, you
need to have good reputation
2. Learning: In order to get a good exchange partner,
you need to find a person with good reputation
 Given these functions, people always have motivations
to establish good reputation.
 Prediction
1.
 Negative reputational information is more important
to predict others’ behavior than positive reputational
information.
People have a bias of reputational information processing.
Number of
Good Reputation
Number of
Bad Reputation
Total Score
Experiment 1
 Hypothesis: People take NEGATIVE reputation
more seriously than POSITIVE reputation.
 Participants : Thirty seven undergraduate students
 Procedure
They were in a hypothetical auction situation
2. They were assigned as “buyer” trying to seek a good
deal for a $200-value product.
3. They looked at reputation of 64 sellers.
4. They evaluate degree of each seller’s trustworthiness as
a potential partner in the hypothetical auctions.
1.
Experimental Design : 64 different set of reputation
Num of BAD info
0,1,2,3,8,12,24,48
この人は
%くらい信頼できると思う(0から100までの数字でお答えください)
この人は
%くらい信頼できると思う(0から100までの数字でお答えください)
8x8=64 combinations
Num of GOOD info
0,1,2,3,8,12,24,48
How much impact for
one increment of information
Analysis
15
Take absolute Values
10
5
Negative
0
0
1
2
3
Positive
-5
-10
14
-15
12
10
8
Negative
Positive
6
4
2
0
0
1
2
3
70
60
Impact of Information
50
Negative information has more impact than
Positive information, especially when the number of
information is small
40
Negative(All)
30
Positive(All)
20
10
0
0
1
2
3
8
number of information
12
24
48
70
60
Impact of Information
50
40
Negative (experienced)
Positive(experienced)
Negative(All)
30
Positive(All)
20
10
0
0
1
2
3
8
number of information
12
24
48
Reciprocal Strategies and
Misunderstandings
B
A
EXCHANGE
Observe
Ex.: B takes Tit-for-Tat and A
takes unconditional defection.
Both continue to defect except
the first move.
Impression
X
Both A and B are not
Cooperative persons!
☹
Amplified misunderstanding by
Noise
B
Noise
A
EXCHANGE
Observe
Ex:Both A and C
take TFT.
They are actually
cooperative, but it often
turns out mutual defection.
Impression
X
Uhmmm… Both A and B
are not cooperative!
☹
Social exchange are noisier than economic exchange (Kollock 1993)
Strategies producing reliable
reputational information
 Reciprocal strategies not good ones in terms of
producing reliable reputational information
 Repeated
 Noisy
 Question: What strategy is good for production
of reliable reputational information ?
 Always consistent behavior regardless others’ behavior
 Unconditional Defection (UD)
 Unconditional Cooperation (UC)
 UC has two advantages for creating
reliable reputation.
No.1: UC is more likely to be recognized as UC
than reciprocal strategies
UC
Noisy
situation
Noise
I don’t know exactly,
but Pink Person
seems to be GOOD!
A
EXCHANGE
Strategy
A
Noise
Evaluate Target Person
Experiment 2
 Participant looked at the
Participant
Noise
Evaluation
No Noise
Evaluation
Difference
same set of players
exchanging with and without
noise
 Hypothesis: The evaluation
on UC is more consistent
across with or without
NOISE than the evaluation
on Reciprocal strategies
(TFT)
Experiment 2
 Participants looked at actors
playing repeated PD games
(actually programmed, but they
were told that the actors were
real persons).
 They then evaluate target
actor’s trustworthiness with 7point schale. (Beggan, Messick, &
Allison, 1988)
Exchange
コンピューター同士の交換
by two programmed actors
Repeated PD
(10 times per set)
Four strategies and noise
Name Strategy 4x4=16 combinations
Always cooperate
UC
TFT
Cooperate at the first game, cooperate if and only
if the other cooperated in the previous game.
TFTB
Basically same as TFT, but when the other have
cooperated two times in the raw, this chooses
defection with 67% of prob. in the current game.
UD
Always defect
Exchange
by two simulated actors
With-Noise Condition: Each player misunderstands other’s
choice with 10% of probability.
Without-Noise Condition: No misunderstanding occurs
Experiment 2 : results
 Participants: Forty seven undergraduate students
 Results
Difference of trustworthiness between
with-noise and without-noise conditions
 UC is the most
robustly evaluated
strategy
 F(3,135)=8.85,
p<.001
1.7
1.6
1.5
1.4
 UC is the least
affected by noise.
 UC can produce
robust reputation.
1.3
1.2
1.1
TFT
TFTB
UD
UC
Situation 1
X is reciprocal strategy
STR
X
STR
A
X
A
EXCHANGE
? ? ?
?
Observers
No.2:UC is able to declare the other
Situation
2
strategy’s
characteristics.
UC is
STR
A
UC
UC
A
EXCHANGE
!
!
!
Cooperative
!
AVulnerable
seems to be*
@xw&
Easy to be
recognized
Basic Idea of Expt3
Target strategies (Programmed)
A
②
Very exploitive
B ①
Exploitive
C ③
Less exploitive
 Difference between
Well..
B is the Best
I guess
C is the worst!
Participant
She wrongly evaluates
A and B
One error occurs.
Participant’s evaluation and
actual ranking of
exploitiveness.
 The more error, the less
reliable reputational
information.
Independent Variables
Target strategies (Programmed)
Three kinds of
strategies
A
Very exploitive
B
Exploitive
C
Not exploitive
LessUC
Error
TF2T: Defects if and only if the partner defected two
times in a raw (Generous)
More Error
Tit-for-Tat
Experiment
3
(Design)
Target strategies (Programmed)
Name
A
Very exploitive
B
Exploitive
C
Less exploitive
Description of strategy
Less Exploitive
Basically same as TFT. It defects just one time after the
partner cooperated two games in the row.
Exploitive
Basically same as the above “Less Exploitive,” and it may
defects two times in a raw with 67% of probability.
Very Exploitive
Basically same as the above “Less Exploitive,” and it may
defects two times in a raw with 80% of probability.
Procedure
 Participants: Eighty six undergraduate students.
 They looked at video clips that three pairs playing repeated PD
game (15 times per pair)
 They then prioritized the three BROWN strategies as potential
partners.
Target strategies (Programmed)
Observe
&
evaluate
Participant
UC, TF2T, or TFT
A
Very exploitive
B
Exploitive
C
Less exploitive
Results of Experiment 3
Mean score of Error
2.5
 F(2,81)=9.81,
2
p<.001
1.5
1
0.5
TFT
TF2T
UC
Pink strategy
UC produces more reliable reputational
Information than reciprocal strategies
Summary
People have a bias in reputational information
processing.
1.

Negative reputation > Positive Reputation (Expt 1)
Reciprocal strategies reduce reliability of
reputational in social exchange. Unconditional
2.
strategy(UC) can produce reliable reputational
information.


UC is easy to be recognized as GOOD even in noisy
situation.(Expt 2)
UC makes it clear the exploitiveness of other
strategies(Expt 3).
To do list…
 How about Unconditional Defection (UD)?
 My prediction
 Probably, UD also has the ability to produce reliable reputational
information as well as UC.
 However, this ability is VERY BAD for UD to survive.
 Is this ability adaptive for UC ? Useful to survive?
 Already conduced a computer simulation study to find
the conditions under which UC is adaptive for this bility.
 Preliminary results say that UC needs help from TFT,
and TFT also needs help from UC in noisy conditions.
Thank you
☺