This thesis contributes to the understanding of peer effects, both methodologically and empirically. The endogeneity of network formation has been a major obstacle to the study of peer influence. The first and the second chapters of the thesis propose a causal identification solution in the potential outcome framework. Combining results from multiple causal inference and statistical network analysis, I show that confounding can be addressed by inferring propensity scores of network link formation from the adjacency matrix. This identification strategy imposes minimum restrictions on the data-generating process and, unlike existing econometric solutions, does not rely on any parametric modelling. As an application, I estimate the effect of high school friendships on bachelor's degree attainment. While previous literature finds that exposure to more high-achieving boys makes girls less likely to obtain a bachelor's degree, I show that if the girls consider the boys as friends, their interactions induce a positive impact instead. Since friendship endogeneity has been addressed, the estimated effect is causal. The third chapter looks at the peer effects generated by group competition. It focuses on the gender differences in preference for competition in a setting where the competition does not involve face-to-face confrontation, and effort is the only determinant of the final ranking. I first develop a model of group competition with heterogeneous preference for ranking. With empirical implications generated from the theoretical model, I then test the gender difference in the preference parameter using web-scraped data from Duolingo, a free online foreign-language learning platform with over 300 million users. Every week, language learners on Duolingo are randomly allocated to groups of 30 people to compete on the number of language lessons completed during that week. The empirical results suggest in this setting, females have a stronger preference for ranking than males.

In a world where networks become a dominant form of organization, the structure of networks and the position of individuals in these networks affect individual behavior and aggregate economic outcomes. The analysis of network formation by a central planner or by individuals themselves is at the heart of this thesis on the economics of networks.Chapter 1 theoretically studies the optimal formation and protection of networks by a central planner knowing that an external agent can destroy k links. The protection of the network can be guaranteed either by densifying the links between nodes, or by protecting the links. When the cost of protection is relatively small, a minimally connected network composed of protected links guarantees the communication flow; if this cost is high, the optimal solution is to form a symmetric network where each node has at least k+1 non-protected links.Chapter 2 explores the decentralized formation of networks in the laboratory by analyzing individual linking formation decisions when one agent has a higher value than others and that the linking formation process is sequential. The results show that sequentiality facilitatesthe coordination on efficient networks but that do not correspond to the Subgame PerfectEquilibrium. The heterogeneity across agents increases the asymmetry of networks because of the polarization of links on the agent with a higher value.Chapter 3 studies the impact of the endogenous formation of networks on the importance of peer effects, applied to dishonest behavior. In order to identify the effects of social comparisons, two controlled environments are designed in the laboratory in which individuals choose or not their peers, and then observe their behavior. The results show that peer effects on dishonest behavior are significantly higher when individuals can choose their peers.

This dissertation consists of two projects, both concerned with how social networks form and how they affect economic decisions. In the first project, entitled Identifying Influential Agents in a Social Network, I define influence centrality, a new measure of influence which combines information on node and neighbor characteristics as well as information on the network structure, by building a model of friendship formation and peer effects and calculating the change in the equilibrium behavior of the network when one node changes its behavior. I apply the model to the context of smoking among middle and high school students and show that a targeted anti-smoking intervention in which a small number of influential smoking students are treated may be more efficient than a uniform program where every student in the school receives the same anti-smoking treatment. The second project, entitled Quantifying Spillovers in Network Formation Over Time, is a joint work with Sean Chu and Shankar Kalyanaraman at Facebook. The goal of the project is to separately identify and understand the relative importance of network effects and unobserved heterogeneity in a multi-period network formation model. Here, we define network effects to be the effect of the number of mutual friends on the probability of friendship and assume that unobserved heterogeneity is time-consistent. We estimate the model on a de-identified panel data set from a small group of users on Facebook and find that both network effects and unobserved heterogeneity are statistically and economically significant predictors of friendship formation on Facebook. We further re-estimate the model first, without network effects and next, without unobserved heterogeneity. Comparing the results from these specifications, we find that network effects play a more important role than unobserved heterogeneity in network formation models on Facebook.

This thesis consists of three papers in the field of social and economic networks. In the first, called Peer Effects in Endogenous Networks, I build a model of endogenous net- work formation in the presence of peer effects, which play an important role for decisions concerning educational attainment, criminal activity, labor market participation and R&D expenditures of firms. The class of payoff functions assumed induce local complementar- ities in effort levels and positive local externalities. Links are one-sided and agents move simultaneously in links and effort levels. I find that equilibrium networks display - other than the complete and the empty network - a core-periphery structure, which is commonly observed in empirical studies. Ex-ante homogenous agents may obtain very different ex-post outcomes, depending on the network that arises in equilibrium. Multiplicity of equilibria serves as an explanation for large differences in behaviour across otherwise identical groups. The second paper, titled Alliance Formation and Coercion in Networks, presents a game-theoretic model of network formation, which allows agents to enter bilateral alliances and to extract payoffs from enemies. Each pair of agents creates a surplus of one, which allies divide in equal parts. If agents are enemies, then the agent with more allies obtains a larger share of the surplus. I show that Nash equilibria are of two types. First, a state of utopia, where all agents are allies. Second, asymmetric equilibria, such that agents can be partitioned into sets of different size, where agents within the same set are allies and agents in different sets are enemies. These results stand in contrast to coalition formation games in the economics of conflict literature, where stable group structures are generally symmetric. The model also provides a game-theoretic foundation for structural balance, a long-standing notion in social psychology, which has been fruitfully applied to the study of alliance formation in international relations. The third paper, A Note on Stochastically Stable States for Alliance Formation and Coercion in Networks, introduces dynamics into the model of the second paper and provides a conjecture for stochastically stable states. At every time period t and with fixed probability p, each agent adjusts his strategy myopically, while with small probability E chooses his strategy at random. The configuration where all agents sustain only positive links is shown to not be stochastically stable. Stochastically stable state are thought to be such that the number of cliques is maximal, under a restriction on the relative size of groups.

This dissertation focuses on the social interaction with economic perspectives. This first essay tests peer effects in the workplace with piece-rate compensation. Reference groups are defined as the geographical peers in undirected networks. A series of spatial econometric models are employed to investigate the social effects. We identify and present evidence for endogenous effects (production) while find no evidence for exogenous effects (characteristics). We also find that the heterogeneity of endogenous effects depends on workers and their peers' characteristics, which is defined as conditional endogenous effects in this paper. Our results suggest that rearranging workers' seats according to their personal characteristics could lead to changes in overall productivity. From a field experiment design, the second essay studies the relation between social distance and training outcomes. We test our hypothesis through two measures: the tips shared with trainees by trainers, and the exist test results of trainees. We find that trainers share more tips to socially closer trainees, and the communications between trainers and trainees have a significant indirect effects on the number of tips shared. The productivity of trainees are also higher when they are socially closer to their trainers. The third essay discusses the identification problem in social peers effect studies. By considering the canonical linear in means model with the rank condition in simultaneous equations model, it suggests that the group structures determines the identifiablity of the desired social effects estimates. Transitive networks are not identified unless there are more information contained in the between group structures. Modifications to the conventional model are also suggested with respect to the recovery of transitive networks and potentially incomplete networks.

Peer pressure and social networks are powerful influences on behaviour. The focus of this thesis is studying the channels through which social networks impact individuals' choices and outcomes in three different contexts.The first paper of this thesis (Chapter 2) develops a theoretical network-based model of Twitter, formulating individual interaction as a dynamic game in which heterogenous agents choose a 'niche' to tweet in, and whom to follow. By characterizing the stable networks that the dynamic Markov process converges to, we show that information does not diffuse as widely as one might expect: although many agents are directly or indirectly connected to each other, agents strategically filter information in accordance with their niche.The second paper of this thesis (Chapter 3) presents a social network model of criminal activity, where agents' payoffs depend on the structure of their connections with each other. The Nash equilibria in crime activity are characterized, and the theoretical results are used to identify the optimal network, which maximize the sum of agents' payoffs, by searching over all possible non-isomorphic graphs of given size. In addition, the effects of different anti-crime policies on the optimal crime network structure and the overall crime level are analyzed and presented.The third paper of this thesis (Chapter 4) studies the direct and spillover effects of social interactions on fundraising and engagement activities in a network of volunteers from Engineers Without Borders, Canada. The network effects are modelled through two separate channels: a strategic interaction term which affects the marginal benefit from supplying effort and a direct spillover term affecting the level of payoff. This model is estimated using several online and offline networks via instrumental variables and system GMM. The results always present large significant levels of strategic complementarities in fundraising activities. However, in engagement activities, strategic complementarities are only significant in online networks. Additionally, engagement activities exhibit positive significant levels of direct spillovers for all networks. In contrast, in fundraising campaigns, the direct spillover effect is only significant in large offline networks.