We explore an equilibrium model of games where players' choice behavior is given by logit response functions, but their payoff responsiveness and beliefs about other players' payoff responsiveness is heterogeneous. We extend the definition of quantal response equilibrium to this setting, calling it subjective quantal response equilibrium, and study two empirically relevant special cases: (1) Heterogeneus quantal response equilibrium (HQRE), where players share the same, correct beliefs, about the distribution of payoff responsiveness; and (2) Truncated quantal response equilibrium (TQRE), where players systematically underestimate the distribution of the other players' responsiveness. We identify a formal connection between TQRE, and the Cognitive Hierarchy (CH) model, and prove that CH can be approximated arbitrarily closely by TQRE. We conduct a series of experiments designed to di®erentiate these models, including simple matrix games and a dominance- solvable game of incomplete information. TQRE fits best. However, CH and QRE models fit almost equally well, and share the property that large predicted deviations from optimization are rarer than small deviations.

We present a dynamic model of network formation where nodes find other nodes with whom to form links in two ways: some are found uniformly at random, while others are found by searching locally through the current structure of the network (e.g., meeting friends of friends). This combination of meeting processes results in a spectrum of features exhibited by large social networks, including the presence of more high- and low-degree nodes than when links are formed independently at random, having low distances between nodes in the network, and having high clustering of links on a local level. We fit the model to data from six networks and impute the relative ratio of random to network-based meetings in link formation, which turns out to vary dramatically across applications. We show that as the random/network-based meeting ratio varies, the resulting degree distributions can be ordered in the sense of stochastic dominance, which allows us to infer how the formation process affects average utility in the network.

We examine the spread of a disease or behavior through a social network. In particular, we analyze how infection rates depend on the distribution of degrees (numbers of links) among the nodes in the network. We introduce new techniques using first- and second order stochastic dominance relationships of the degree distribution in order to compare infection rates across different social networks.

In laboratory experiments, information cascades are ephemeral phenomena, collapsing soon after they form, and then reforming again. These formation/collapse/formation cycles occur fre- quently and repeatedly. Cascades may be reversed (collapse followed by a cascade on a di®erent state) and more often than not, such a reversal is self-correcting: the cascade switches from the incorrect to the correct state. With a long enough horizon, full information aggregation may therefore occur in an environment where Nash equilibrium predicts learning to be incomplete. Past experimental work focused on relatively short horizons, where these interesting dynam- ics are rarely observed. We present experiments with a longer horizon, and also investigate the e®ect of signal informativeness. We propose a theoretical model, based on quantal response equilibrium, where temporary and self-correcting cascades arise as equilibrium phenomena. The model predicts that learning will be complete and also predicts the systematic di®erences we observe experimentally in the dynamics, as a function of signal informativeness. We extend the basic model to include a parameter measuring base rate neglect and ¯nd it to be a statistically signi¯cant factor in the dynamics, resulting in somewhat faster rates of social learning.

We consider an environment where individuals sequentially choose among several actions. The payoff to an individual depends on her action choice, the state of the world, and an idiosyncratic, privately observed preference shock. Under weak conditions, as the number of individuals increases, the sequence of choices always reveals the state of the world. This contrasts with the familiar result for pure common-value environments where the state is never learned, resulting in herds or informational cascades. The medium run dynamics to convergence can be very complex and non-monotone: posterior beliefs may be concentrated on a wrong state for a long time, shifting suddenly to the correct state.

We examine a simple economic model of network formation where agents benefit from indirect relationships. We show that small-world features—short path lengths between nodes together with highly clustered link structures—necessarily emerge for a wide set of parameters.

Homophily, the tendency of linked agents to have similar characteristics, is an important feature of social networks. We present a model of network formation that allows the meeting process to depend on individuals types and study the impact of such a bias on the network structure. Our main results fall into three categories: (i) we show how, at the group level, homophily depends on the groups size and the parameters of the formation process, (ii) we understand precisely the determinants of local homophily at the individual level, and (iii) we compare the distributions of intra- and inter-group links in terms of stochastic dominance. Especially, we find that popular individuals have more diverse social networks. Our results are supported empirically in the AddHealth data looking at networks of social connections between boys and girls.

This paper studies environments where individuals allocate resources across relationships with others, creating a weighted, directed network. Value is achieved both through an exogenous factor and maintaining close connections to high-value individuals. We consider two cases corresponding to the direction benefits flow along links. In Model T (for "taking") agents receive benefits through the links they create, whereas in Model G (for "giving") the reverse is true: agents pass value along their links. Equilibrium and socially e±cient networks are characterized. In Model G equilibrium networks do not necessarily maximize group welfare, but in Model T e±cient networks constitute equilibria despite extensive network externalities.

Previous experimental work has demonstrated the power of the classical theory of economic dynamics. In particular, the models have proved to be accurate in predicting the principle directions of movement and orbit-like behavior in general equilibrium systems. Questions left open and addressed in this study are (i) do the markets necessarily converge to a unique interior equilibrium or can markets exhibit the “explosive property” of instability and (ii) among the several variations of the classical model, which, if any, is most accurate in predicting what is actually observed in experiments? Markets were created and studied in the extreme environments identified by Scarf and Hirota. Such environments allow us to study features of market adjustments that are obscured by the complexity of naturally occurring markets. Two fundamental results are reported. First, the instability phenomenon of “expanding orbits” predicted by theory does actually exist in the markets and exists in much the form that theory suggests. That is, prices spiral outwardly around the equilibrium prices and do so in directions predicted by theory. This type of disequilibrium behavior is observed for the first time in actual market behavior. Thus, the principles governing market adjustment are not among those that guarantee convergence to a unique interior equilibrium. Second, the best dynamic model from among those studied is of the form [NOTATION] where the diagonal elements are positive and the E_i(P) are excess demands.

We consider a pure informational externalities environment in which agents make binary decisions. The agents are asymmetrically informed about a payoff-relevant state variable, and choose the timing of their decisions strategically. There is a delay cost that must be balanced against the possibility of learning from the announcements of predecessors. For two player games in discrete time, we show that in the unique equilibrium the game ends in finite time and the agent with better information decides earlier. As the time intervals become vanishingly short, all announcements occur immediately, no delay costs are incurred, and the equilibrium outcome approaches the first best. We show that for games with many players and short time intervals, the true state is revealed immediately and thus an e±cient herd arises in which almost all agents announce correctly. For any time interval, welfare is higher under strategic timing relative to an exogenous sequence of decisions.