Network Spillovers and Social Interactions
Paper Session
Monday, Jan. 4, 2021 3:45 PM - 5:45 PM (EST)
- Chair: Alan Griffith, University of Washington
Binary Outcomes and Linear Peer Effects in Networks
Abstract
We show that the linear-in-means model of peer effects in networks, introduced to study continuous outcomes, can be used to analyze binary outcomes. Error terms must have a specific structure, which we characterize. We analyze the microfoundations of this model. We show that any binary action game of complete information consistent with the econometric model can be modified in a simple way to guarantee equilibrium uniqueness in dominant strategies. We propose two simple consistent estimators and investigate their small sample properties through Monte-Carlo simulations. We apply this linear framework to revisit the empirical analysis of teenage smoking of Lee, Li, and Lin (2014). Overall, we show that peer effects on binary outcomes can be analyzed in a linear framework. A main advantage is that multiplicity, at the heart of the existing literature on the topic, can be avoided.Identification and Estimation of Social Interactions in Endogenous Peer Groups
Abstract
This paper studies a linear-in-means social interaction model with endogenously formed groups. The endogeneity of peer groups results from the correlation between the unobservables that affect the group formation and the unobservable that affects the outcome. We characterize the group formation using a many-to-one matching model, where each individual joins one group and each group contains many individuals. In this model, each individual chooses a group among the groups that she is eligible for. The groups formed in equilibrium depend on the observed and unobserved characteristics of all the individuals in the market, making it difficult to correct for the selection bias without simplification. Following the idea in the matching literature, we approximate the equilibrium in the finite model using the equilibrium in a limiting model, where the number of individuals in the market approaches infinity. Given the limiting equilibrium, each individual's group only depends on her own characteristics. We then show that the formation of the groups follows a multivariate selection rule and the selection bias is a nonparametric function of the preference and qualification indices in the group formation. We provide constructive identification results for the parameters in social interactions and group formation under the assumption that the unobservables in the group formation are nonparametrically distributed.Adjusting for Peer-Influence in Propensity Scoring When Estimating Treatment Effects
Abstract
Analyses of treatments, experiments, policies, and observational data, are confounded if people's treatment decisions and outcomes are influenced by the behavior of their friends and acquaintances. We show how to account for this interference by explicitly modeling peer interaction in treatment participation decisions, as well as how to incorporate these interaction effects into the most common matching technique used to evaluate treatment effects: propensity score matching. As an illustration, we show how peer-influenced propensity score matching gives different results compared to the use of standard propensity score matching in the estimation of the impact of exercise participation on depression.An Analysis of Network Peer Effects without Network Data: The Application of Graphical Network-Detection to Peer Effects
Abstract
The past couple decades have seen an explosion of research on peer effects in networks. However, such analyses require data on peer groups, which is often unavailable and costly to collect. New methods (e.g., de Paula et al. 2018; Manresa, 2016) have been developed to infer the network structure from panel data on outcomes, but these methods have yet to be applied in peer effects settings. Here, we take the first steps in assessing the performance of Lasso, elastic net, and graphical lasso (glasso) network detection methods in finding relevant peer networks. We perform analysis using data from AddHealth and the Tennessee STAR experiment. In the former, we compare detected networks to those networks identified in the data. In both datasets, we assess the algorithms' success in recovering plausible networks by comparing detected networks to the data's group structure. The analysis here is a first step in bringing these new methods to bear in studying the effects of peers in settings in which network data is unavailable.JEL Classifications
- C1 - Econometric and Statistical Methods and Methodology: General
- C3 - Multiple or Simultaneous Equation Models; Multiple Variables