Three Essays on Econometrics of Networks

Abstract

This dissertation contributes three major contents to the econometrics of networks: application of discrete choice network formation models, identification of reduced form models with measurement errors in networks, and finding efficient sampling methods of epidemics over networks. In the first chapter, we investigate the impact of the initial academic social network, formed from advisor-advisee relationships and coauthorships, for economics Ph.D. students (advisees) in the U.S. on their early stage productivity. We define the \emph{academic social network} as a union of i) an advisor-advisee network and ii) a coauthorship network. We model the advisor-advisee relationships with a preferential attachment-like process based on a discrete choice model and find that advisees show weak gender homophilic preferences when choosing advisors. We further model early stage coauthorship formation of advisees through a bipartite network setup, also based on a discrete choice model, and find that advisees prefer to choose projects that are coauthored with their advisors during their graduate studies. Given the \emph{academic social network} through the two networks, we find that the corresponding network statistics for advisees have significant positive correlation with early stage output but find weak evidence of gender causing difference. Through simulated synthetic data, we show that for advisees, in average, preference based decision making leads to individual level percentage-wise productivity gain but loss in the aggregate level, compared to random matching to advisors and projects. This implies that a preference based allocation of advisors to advisees is less efficient in the social planner's view. In the second chapter, we consider the effects of the mismeasurement of networks on reduced-form peer-effect linear-in-means and linear-in-sums estimates. Applied researchers frequently estimate network-based peer effects models using observed network data that includes only a subset of the true links. Our results require an assumption that the expected covariance of characteristics between linked agents is the same regardless of whether the link is observed or not. Analytic results show that the linear-in-means peer effects estimate is in general attenuated, and this is a special case of ``classical" measurement error. In contrast, linear-in-sums direct and peer effect estimates may be attenuated, augmented, or consistent; the inconsistency depends upon the missingness mechanism and the relationship between the network and covariates. We demonstrate the effect of mismeasured links in both models using two datasets and through simulations. These results show that the effects of mismeasured networks on subsequent estimands is quite sensitive to the parameter that is being estimated. In the last chapter, we present a feasible version of the Neyman allocation for sampling epidemics over networks. Our method requires knowledge of only the first moments of the degree-based strata and an epidemic model that captures the dynamics of the diffusion process. Through simulations on randomly generated networks, we demonstrate that the optimal Neyman allocation and our proposed methods show efficiency gains over simple random sampling, particularly in the early stages of an epidemic. The feasible method closely approximates the performance of the optimal method while being implementable in practice. Our findings can inform sampling strategies for monitoring real-world epidemics given limited resources.