Herding behavior Experimental Studies --AlisonLo

My experimental studies on information cascades and herding behavior, comparing Bayesian belief updating with other behavioral heuristics that account for the observed behaviors

1. HERDING BEHAVIOR AND CONSUMER DECISION MAKING [email protected] ALISON LO

2. Imprint Uniform Social Behavior Parallel reasoning A purposive imitation act with an intention to achieve the goal the imitated act is targeting

3. Herding refers to situations in which the observation of other people’s actions contributes as an antecedent to the behavior that is not expected to occur without the observation Herding

4. Why Imitate (Herd)? Payoff externalities Economical Herding

5. Why Imitate (Herd)? Payoff externalities Economical Network externality Herding

6. Why Imitate (Herd)? Payoff externalities Economical Network externality Relative payoff contingency Herding

7. Why Imitate (Herd)? Payoff externalities Economical Network externality Relative payoff contingency Psychological Conformity preference Sanctions on deviants

8. Why Imitate (Herd)? Payoff externalities Economical Network externality Relative payoff contingency Psychological Conformity preference Sanctions on deviants Communication Coordinational

9. Why Imitate (Herd)? Payoff externalities Economical Network externality Relative payoff contingency Psychological Conformity preference Sanctions on deviants Communication Coordinational Informational

10. Examples of Herd Behavior adoption of technology (Dybvig & Spatt, 1983, Choi, 1997) agricultural innovation (Roger, 1983) employers' decisions on job candidates (Stern, 1990) banks’ lending policies (Scharfstein & Stein, 1990) choices of political candidates, selection of academic research topics, excess volatility in asset markets, and changes in fashion (Banerjee , 1992). drug, alcohol, and cigarette use, cohabitation, religious affiliation, crime, medical practice, choices of papers by editors, contagious fear of bank solvency (BHW, 1992) costly political movement (Lohman, 1994) investment decisions of fund managers (Palley, 1995) decisions of location to open a new branch (Chaudhuri, Jayaratne & Chang, 1997)

11. An informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individuals without regard to his own information. Information Cascade

12. Bikhchandani, Hirshleifer and Welch (1992), “A theory of fads, fashion, custom, and cultural change as information cascades,” Journal of Political Economy , 100 (5), 992-1026. Herding model

13. Decisions are sequential and exogenously ordered two states or nature, identical and independent payoff structure, dichotomous action space DM act on the basis of private signals and observation of the actions of previous individuals Externalities that might enforce uniformity are absent Bikhchandani, Hirshleifer and Welch (1992) Model Assumptions

14. Information Signaling Individuals update the posterior probability of the underlying state by observing the decisions made by others inferring the private signals accompanying the observed decisions Rationale Signals are imperfect and independently sampled from the underlying state Accumulation of signals should increase the posterior of the true state

15. Example

16. A B P(A) = P(B) = ½

17. BUY

18. NOT BUY

19. Normative Analysis P(A) = P(B) = ½ A B First Buyer

20. P(A) = P(B) = ½ A B NOT BUY (B) First Buyer

21. P(A) = P(B) = ½ A B Second Buyer Observes that the first buyer didn’t buy, hence he can infer that the first buyer tasted a blue bottle

22. P(A) = P(B) = ½ A B Second Buyer The relevant sample for the second buyer is 1 blue and 1 red BUY (A)

23. P(A) = P(B) = ½ A B Second Buyer The relevant sample for the second buyer is 2 blue NOT BUY (B)

24. P(A) = P(B) = ½ A B Third Buyer Observes that the first buyer didn’t buy, hence she can infer that the first buyer tasted a blue bottle

25. P(A) = P(B) = ½ A B Third Buyer Observes that the second buyer didn’t buy, hence she can infer that the second buyer tasted a blue bottle

26. P(A) = P(B) = ½ A B Third Buyer If it was red , the relevant signal to the second buyer would have been 1 blue and 1 red

27. P(A) = P(B) = ½ A B Third Buyer

28. P(A) = P(B) = ½ A B Third Buyer

29. P(A) = P(B) = ½ A B Third Buyer

30. P(A) = P(B) = ½ A B Third Buyer It is optimal for the third buyer, having observed the actions of the two buyers ahead of him, to follow their behavior without regard to his own information! Information Cascade

31. P(A) = P(B) = ½ A B Fourth Buyer First buyer Second buyer Third buyer ? His relevant sample is 2 blue plus his own private signal Information Cascade

32. P(A) = P(B) = ½ A B Fourth Buyer The cascade cannot be broken unless a stronger private signal is received Information Cascade

33. P(A) = P(B) = ½ A B Fifth Buyer Is allowed to take 2 samples (with replacement)

34. P(A) = P(B) = ½ A B Fifth Buyer

35. Why Imitate (Herd)? Payoff externalities Economical Psychological Information cascades Herding Imitation can be sophisticated, being based upon rational weighing of pros and cons Irrational Imitation

36. Rationale Signals are imperfect and independently sampled from the underlying state Accumulation of signals should increase the posterior of the true state Problem Sampling error Path dependency in signal accumulation via observing decisions

37. “On the empirical side, we must point out that most models so far rely primarily on anecdotal observations. Thus, the foremost challenge to the herding literature may be the current scarcity of rigorous empirical evidence.” (p. 612) Devenow and Welch (1996), “Rational herding in financial economics,” European Economics Review , 40, 603-615.

38. Anderson and Holt (1997), “Information cascade in the laboratory,” The American Economic Review , 87(5), 847-862. Huck and Oechssler (1999), “Informational cascades in the laboratory: Do they occur for the right reasons?” (Working paper, Humboldt University)

39. “Laboratory experiments have been performed. These experiments tests have confirmed that people do fall into cascades. Further, people imitated in these experiments primarily when it was sensible to do so based upon the circumstances, rather than irrationally following predecessors out of a taste for the status quo or for conformity.” (Hirshleifer, 1995)

40. Combining sources of information Order effect Anchoring and adjustment hypothesis (Slovic and Lichtenstein, 1971) Belief updating Conservative bias (Edward, 1968)

41. Edward’s (1968) explanations to conservative bias in belief updating: Mis aggregation Misp e rc e ption Response Bias

42. Combining sources of information of different forms Base rate studies: Base rate and case specific data (Tversky and Kahneman, 1973)

43. Base rate studies: The cab problem (Tversky and Kahneman, 1980) A cab was involved in a hit-and-run accident at night. Two cab companies, the green and the blue, operate in the city. A witness reports that the offending cab was blue. On a test of ability to identify cabs under appropriate visibility conditions, the witness is correct on 80% of the identifications. The court learns that 85% of the city’s cab are green and 15% are blue.

44. Base rate studies: Base rate and case specific data (Tversky and Kahneman, 1980)

45. Combining sources of information of different forms Base rate studies: Base rate and case specific data (Tversky and Kahneman, 1973) Consumer decision making: Consensus and product attribute information (Aaker and Maheswaran, 1997)

46. Combining sources of information of different forms Base rate studies: Base rate and case specific data (Tversky and Kahneman, 1973) Consumer decision making: Consensus and product attribute information (Aaker and Maheswaran, 1997)

47. w w b Normative integration of signal information Public Private

48. w w b Normative integration of signal information irrespective of signal order and form Public Private

49. Objectives Do people intuitively realize the existence of information cascades and discount the observed decisions of those under the cascade? Does Bayesian reasoning explain people’s behavior in this case? If not Where does it fail (computational, common knowledge of rationality) Any alternative models?

50. Method Experiment 1 (Group setting) Subjects: 96 Task: Guess which one of two urn has generated the signal samples for the group Incentive: $20 for a correct guess for 4 randomly selected problems

51. Urn A 3 white 2 black Urn B 1 white 2 black

52. Urn A 3 white 2 black Urn B 1 white 2 black

53. Urn A 3 white 2 black Urn B 1 white 2 black

54. Urn A 3 white 2 black Urn B 1 white 2 black

55. Urn A 3 white 2 black Urn B 1 white 2 black

56. Urn A 3 white 2 black Urn B 1 white 2 black

57. Urn A 3 white 2 black Urn B 1 white 2 black Which is the target urn for this trial? A B

58. Urn A 3 white 2 black Urn B 1 white 2 black Which is the target urn for this trial? A B

59. Method Experiment 1 (Group setting) Subjects: 96 Task: Guess which one of two urn has generated the signal samples for the group Incentive: $20 for a correct guess for 4 randomly selected problems Experiment 2 (Individual problem solving) Subjects: 80 Task: Solve equivalent logical reasoning problem (All previous decisions are correct) Incentive: HK$60 if 80% or more correct

60. Urn A 2 white 1 black Urn B 3 white 5 black Which urn is more likely to be the target urn? A B

61. Results

62. Urn A 2 white 1 black Urn B 3 white 5 black Types of decisions: First position First position : only private signal

63. Urn A 2 white 1 black Urn B 3 white 5 black Types of decisions: First one under cascade First one under cascade First one under second cascade

64. Urn A 2 white 1 black Urn B 3 white 5 black Types of decisions: Later ones under cascade Later ones under cascade Later ones under cascade

65. Urn A 2 white 1 black Urn B 3 white 5 black Types of decisions: Break cascade Break cascade

66. Definitions of the four types of decisions in terms of posteriors

67. % Bayesians: EXP2, EXP1 For both experiments, Bayesian % are higher for consensual decisions than for private- or public-dominant decisions Findings:

68. % Bayesians: EXP2, EXP1 For both experiments, Bayesian % are higher for consensual decisions than for private- or public-dominant decisions Findings:

69. % Bayesians: EXP2, EXP1 For both experiments, Bayesian % are higher for consensual decisions than for private- or public-dominant decisions Findings:

70. % Bayesians: EXP2, EXP1 For both experiments, Subjects are more likely to correctly decide against their private signal if they are in the later positions under cascade than in the first position under cascade Findings: Within Cascade Effect

71. % Bayesians: EXP2, EXP1 When the private and public information is in conflict , subjects in EXP1 exhibited a biased tendency to follow their private signals . They are more likely to correctly dissolve a cascade than to initiate or follow a cascade Findings:

72. % Bayesians: EXP2, EXP1 Subjects in EXP2 did not demonstrate a general tendency to overweigh or underweigh their private signals (against public information). Findings:

73. w w b Normative integration of signal information Public Private

74. Smart Counting rule (Anderson and Holt, 1997) Extract signals from observed decision not under cascade Combine it with private signal If there is more white than black balls in the sample, pick the urn with a higher white to black ratio Unsophisticated Counting rule Count decisions heuristic Extract signals from observed decision whether or not they are under cascade Combine it with private signal If there is more white than black balls in the sample, pick the urn with a higher white to black ratio

75. w w b Public Private Overweighing Private Signal 2 w w b w w

76. w w b Public Private Overweighing Public Signal 2 w w b b

77. Alternative Behavioral heuristics ( EXP1 , EXP2) If the decisions are consensual , all except the follow public signal fits the data well

78. Alternative Behavioral heuristics ( EXP1 , EXP2) The follow public heuristic is dominated by the Bayes rule in both experiments.

79. Alternative Behavioral heuristics ( EXP1 , EXP2) Smart counting rule is dominated by the follow private rule in Experiment 1.

80. Alternative Behavioral heuristics ( EXP1 , EXP2) Evidence against the empirical use of Bayes rule include the differential Bayesian rates for starting vs. following positions of cascade.

81. Alternative Behavioral heuristics ( EXP1 , EXP2) For later positions, follow private and count signals heuristics are both more descriptive than Bayes rule

82. Percentage of subjects whose decisions have the highest coincidence with the following behavioral rules in Exp 1

83. Percentage of subjects whose decisions have the highest coincidence with the following behavioral rules in Exp 1

84. Comparison of the two experimental conditions suggest that when subjects are not assured of the rationality of others they overweigh their private signal The within cascade effect is inconsistent with the use of Bayes rule. Instead, the count-signals heuristic can explain why latter members are more likely to herd than the first one under cascade Conclusion

85. Major findings Higher Bayesian rate for private-dominant than public-dominant decisions Explanations: High salience for private information Uncertainty of rationality assumption

86. Major findings Following a cascade is more likely than starting one Explanations: Attentional Illusional in perceived diagnosticity

87. Major findings Overweight private signal + Within Cascade effect (exaggerate the impact of decision under cascade) =? Simultaneous rather than sequential decisions False consensus bias (Ross, 1977) People have a general tendency to infer that others will behave as they do themselves.

88. Characteristics of the Count-Signals heuristic does not consider the order in which the observed decisions are made does not recognize information cascade does not consider asymmetric diagnosticity of signals overlooks indeterminacy of multiple signals

89. Extended predictions of the Unsophisticated Counting heuristic a cascade should increase (decrease) the observed Bayesian percentage for positions after the cascade if the Bayesian decision for that position is consistent (inconsistent) with the decision under cascade a very long cascade is almost impossible to break

90. Further Experiments Generalized (facilitative + debilitating) cascade effect Putting Bayesian subjects in a group Association between Bayesian performance and overweighing of private signal

91. Marketing Implicatons: introductory offer Herding or Hedging? Choice of dining dishes in a group (Ariely and Levav, 1999) Traders in a market (Camerer, 1987) Two types of traders differ in the dividends they received in the two states of nature, voluntarily exchanged assets in a double bid auction

92. Comparative Ignorance hypothesis: (Fox and Tversky, 1995) Individual: Ambiguity aversion, present in a comparative context in which a person evaluates both clear and vague prospects, seems to disappear in a noncomparative context in which a person evaluates only one of these prospects in isolation. Interpersonal: People were reluctant to bet on unfamiliar events only when they were reminded of other people who were more knowledgeable concerning those events.