Confronting Theory with Experimental Data and vice versa. Lecture VII Social learning. The Norwegian School of Economics Nov 7-11, 2011

Transcription

1 Confronting Theory with Experimental Data and vice versa Lecture VII Social learning The Norwegian School of Economics Nov 7-11, 2011

2 Quantal response equilibrium (QRE) Players do not choose best response with probability one (as in Nash equilibrium). Players choose responses with higher expected payoffs with higher probability better response instead of best responses. Players have rational expectations and use the true mean error rate when interpreting others actions.

3 Modify Nash equilibrium to incorporate realistic limitations to rational choice modeling of games. Provide a statistical framework (structural econometric approach) to analyze game theoretic data (field and laboratory). If Nash had been a statistician, he might have discovered QRE rather then Nash equilibrium Colin Camerer

4 In practice, QRE often uses a logit or exponentiation payoff response function: Pr(a i )= exp[λ P a i A i Pr(a i )u i (a i,a i )] Pa 0 i A i exp[λ P a i A i Pr(a i )u i (a 0 i,a i)]. The choice of action becomes purely random as λ 0, whereas the action with the higher expected payoff is chosen for sure as λ.

5 QRE does not abandon the notion of equilibrium, but instead replaces perfectly with imperfectly, or noisy, rational expectations. Players estimate expected payoffs in an unbiased way(expectations are correct, on average). As such, QRE provides a convenient statistical structure for estimation using either field or experimental data.

6 Example I Consider the game where A>0 and B>0. L M R U 1, 1 0, 0 1, 1 M 0, 0 0, 0 0,B D 1, 1 A, 0 1, 1 The game has a unique THP (D, R), and the NE consists of all mixtures between U and D (resp. L and R) forplayer1 (resp. 2). The limit logit equilibrium selects p =( 1 2, 0, 1 2 ) and q =(1 2, 0, 1 2 ) as the limit point.

7 QRE for example I with A=B=5

8 QRE for example I with A=B=100

9 Example II Consider the game R L T x, 1 1, 2 B 1, 2 2, 1 All limit points are Nash equilibria but not all Nash equilibria are limit points (refinement). Computable in small finite games (Gambit).

10 QRE for example II Properties of the QRE correspondence

11 QRE for example II Own-payoff Effects

12 Data Lieberman (1960) B 1 B 2 B 3 A A A Ochs (1995) B 1 B 2 A 1 1, 0 0, 1 A 2 0, 1 1, 0 B 1 B 2 A 1 9, 0 0, 1 A 2 0, 1 1, 0 B 1 B 2 A 1 4, 0 0, 1 A 2 0, 1 1, 0 Game 1 Game 2 Game 3

13 QRE for Lieberman (1960)

14 QRE for Ochs (1995) Game 2

15 QRE for Ochs (1995) Game 3

16 Localized conformity One of the most striking regularities of (human) society is localized conformity. In many social and economic situations, individuals are influenced by the decisions of others. The commonest examples occur in everyday life, as in choosing a fashionable restaurant or a popular movie. Similar influences also affect technology adoption and asset market decisions.

17 Why should individuals behave in this way? N - finite set of agents Ω - set of states of nature σ i (ω) -privatesignal A - finite set of actions U(a, ω) -commonpayoff function

18 Incomplete and asymmetric information Agentsareuncertainabouttheeventandtheinformationaboutitis shared asymmetrically among them. Pure information externality Each agent s payoff depends only on the state of nature and his own action.

19 Social learning Agents use their information to identify a payoff-maximizing action so the choice of action reflects that information. By observing an agent s action, it is possible to learn something about his information and make a better decision. In social settings, where agents can observe one another s actions, it is rational for them to learn from one another. Social learning occurs when individuals learn by observing the behavior of others.

20 What have we learned from Social Learning? The striking uniformity of social behavior is an implication of social learning: Despite the asymmetry of information, agents rationally ignore their own information and follow the herd. Despite the available information, so-called herd behavior and informational cascades often result in an inefficient choice. Mass behavior is fragile, in the sense that small shocks may cause behavior to shift suddenly and dramatically.

21 The standard social-learning model The standard model has special features that are restrictive and deserve further examination: Perfect information Once-in-a-lifetime decisions The dynamics of social learning depend crucially on the extensive form of the game.

22 The canonical model of social learning A set of players N, afinite set of actions A, asetofstatesofnatureω, and a common payoff function U(a, ω) where a A is the action chosen and ω Ω is the state of nature. Player i receives a private signal σ i (ω), a function of the state of nature ω, and uses this private information to identify a payoff-maximizing action.

23 The canonical assumptions Bayes-rational behavior Incomplete and asymmetric information Pure information externality Once-in-a-lifetime decisions Exogenous sequencing Perfect information

24 Direct methodological extensions Caplin & Leahy (AER 1994), Chamley & Gale (ECM 1994) Avery & Zemsky (AER 1999), Chari & Kehoe (JET 2004) Çelen & Kariv (GEB 2004), Smith & Sørensen (2008) Bala & Goyal (RES 1998), Gale & Kariv (GEB 2004), Acemoglu et al. (2008)

25 The model of BHW (JPE 1992) There are two decision-relevant events, say A and B, equally likely to occur ex ante and two corresponding signals a and b. Signals are informative in the sense that there is a probability higher than 1/2 that a signal matches the label of the realized event. The decision to be made is a prediction of which of the events takes place, basing the forecast on a private signal and the history of past decisions.

26 Whenever two consecutive decisions coincide, say both predict A, the subsequent player should also choose A even if his signal is different b. Despite the asymmetry of private information, eventually every player imitates her predecessor. Since actions aggregate information poorly, despite the available information, such herds / cascades often adopt a suboptimal action.

27 Anderson and Holt (AER 1997) investigate the social learning model of BHW experimentally. They report that rational herds / cascades formed in most rounds and that about half of the cascades were incorrect. Extensions: Hung and Plott (AER 2001), Kübler and Weizsäcker (RES 2004), Goeree, Palfrey, Rogers and McKelvey (RES 2007).

28 The model of Smith and Sørensen (ECM 2000) Two phenomena that have elicited particular interest are informational cascades and herd behavior. Cascade: players ignore their private information when choosing an action. Herd: players choose the same action, not necessarily ignoring their private information. Smith and Sørensen (2000) show that with a continuous signal space herd behavior arises, yet there need be no informational cascade.

29 The model of Çelen and Kariv (GEB 2004) Signals Each player n {1,...,N} receives a signal θ n that is private information. For simplicity, {θ n } are independent and uniformly distributed on [ 1, 1]. Actions Sequentially, each player n has to make a binary irreversible decision x n {0, 1}.

30 Payoffs x =1is profitable if and only if P n N θ n 0, andx =0is profitable otherwise. Information Perfect information I n = {θ n, (x 1,...,x n 1 )} Imperfect information I n = {θ n,x n 1 }

31 The decision problem The optimal decision rule is given by x n =1if and only if E h P Ni=1 θ i I n i 0. Since I n does not provide any information about the content of successors signals, we obtain x n =1if and only if θ n E h P n 1 i=1 θ i I n i.

32 The cutoff process For any n, theoptimalstrategyisthecutoff strategy ( 1 if θn ˆθ x n = n 0 if θ n < ˆθ n where Xn 1 ˆθ n = E i=1 θ i I n is the optimal history-contingent cutoff. ˆθ n is sufficient to characterize the individual behavior, and {ˆθ n } characterizes the social behavior of the economy.

33 A three-agent example ˆ ˆ θ1 θ2 3 θˆ

34 A three-agent example 1 x =0 1/2 0 x =1-1/2-1 ˆ ˆ θ1 θ2 3 θˆ

35 A three-agent example under perfect information 1 x =0 1/2 3/4 1/4 0 x =1-1/ /2 ˆ - 3/4 ˆ θ1 θ2 3 θˆ

36 A three-agent example under imperfect information 1 1/2 5/ /2 ˆ - 5/8 ˆ θ1 θ2 3 θˆ

37 The case of perfect information The cutoff dynamics follows the cutoff process where ˆθ 1 =0. ˆθ n = 1+ˆθ n 1 2 if x n 1 =1 1+ˆθ n 1 2 if x n 1 =0

38 A sequence of cutoffs under perfect information Cutoff Agent

39 A sequence of cutoffs under perfect information Cutoff Agents

40 Informational cascades 1 < ˆθ n < 1 n so any player takes his private signal into account in a non-trivial way. Herd behavior {ˆθ n } has the martingale property by the Martingale Convergence Theorem a limit-cascade implies a herd.

41 Thecaseofimperfectinformation The cutoff dynamics follows the cutoff process where ˆθ 1 =0. ˆθ n = 1+ˆθ 2 n 1 2 if x n 1 =1 1+ˆθ 2 n 1 2 if x n 1 =0

42 A sequence of cutoffs under imperfect and perfect information Cutoff Imperfect Perfect Agent

43 A sequence of cutoffs under imperfect and perfect information Cutoff Imperfect Perfect Agent

44 Informational cascades 1 < ˆθ n < 1 n so any player takes his private signal into account in a non-trivial way. Herd behavior {ˆθ n } is not convergent and the divergence of cutoffs implies divergence of actions. Behavior exhibits periods of uniform behavior, punctuated by increasingly rare switches.

45 Takeaways The dynamics of social learning depend crucially on the extensive form of the game. Longer and longer periods of uniform behavior, punctuated by (increasingly rare) switches. A succession of fads: starting suddenly, expiring easily, each replaced by another fad. Why do markets move from boom to crash without settling down?

46 The experiments of Çelen and Kariv (AER 2004, ET 2005) In market settings, we observe behavior but not beliefs or private information. In the laboratory, we can elicit subjects beliefs and control their private information. Test the model s predictions and study the effects of variables about which our existing theory has little to say.

47 Cutoff Perfect information (Mean cutoffs when all predecessors acted alike) Theory Actual Subject

48 Percentage of concurring, neutral and contrary decision points 50% 45% 40% 35% 30% 25% 20% 15% 10% 5% 0% Concurring Neutral Contrary

49 The distribution of contrary subjects 50% 45% 40% 35% 30% 25% 20% 15% 10% 5% 0%

50 Cutoff Imperfect information (Mean cutoffs in concurring decisions) Theory Actual Subject

51 Cutoff Imperfect information (Mean cutoffs) Actual Theory Subject

52 The econometric analysis At each decision turn n, with probability p n a player is rationally, and with probability 1 p n he is noisy. The cutoff of a noisy player is a random draw from a distribution function G n with support [ 1, 1] and mean θ n. Others cannot observe whether a player behavior is noisy, but the sequences {p n } and {G n } are common knowledge.

53 The estimated cutoff process The cutoff dynamics of rational players follow the process ˆθ n = ˆθ n 1 where ˆθ 1 =0. 10+(1 p n 1 ) θ n 1 +p n 1ˆθ n 1 2 if x n 1 = A, 10+(1 p n 1 ) θ n 1 +p n 1ˆθ n 1 2 if x n 1 = B, The estimated parameters for the first decision-turn are employed in estimating the parameters for the second turn, and so on.

54 Cutoff Sequences of cutoffs under perfect information (Theory and estimated) Theory Estimated Subject

55 Cutoff Sequences of cutoffs under perfect information (Theory and estimated) Theory Estimated Subject

56 Sequential social-learning model: Well heck, if all you smart cookies agree, who am I to dissent?

57 Imperfect information: Which way is the wind blowing?!