Survey_HK_2019

Abstract

We build a simulated model of network dynamics and limited information. We remove the possibility of fake news from the model and demonstrate that, given the search and matching facility that social media o§er for connecting with new online friends, two social elements alone, are su¢ cient for producing, (a) networks that gradually exhibit more homophily and polarization over time, and, (b) a gradual downgrading of expert opinion on issues for which knowledge is limited. The two social elements that are su¢ cient for producing these dynamics are, (i) individual biases, such as biased assimilation and conÖrmation bias, and (ii) the tendency that people have for socially aligning their actions with actions of network friends. For distinguishing the two eras of social networks, the pre-social-media era and the postsocial-media era, the key is to introduce a search-and-matching mechanism that can bring together new friends. Our search-and-matching process involves features of coordination games with incomplete information. In these games, players need to form beliefs about a fundamental value. In our framework, there is a public noisy signal that captures the role of expert opinion on this fundamental value. In addition, players have access to private signals and also try to coordinate with network friends. In order to take actions (e.g., immunizations, political votes, etc.) related to this fundamental value (e.g., the risk of a disease, the risk of a Öscalcrisis, etc.), players form beliefs on what other players believe, i.e., they form higher-orderbeliefs. In this environment, fundamental (structural) biases of players, more related to their education level or culture, such as conÖrmation biases, cause a preference for choosing internet social media friends with similar biases, the network feature known as homophily. The key assumption we make is that, in each period t 2 f0, 1, :::g, there is a new task carrying a new fundamental value, that is unknown and needs to be learned through signals available in period t. Therefore, the time horizon available for learning about parameter is one period only. Despite that the fundamental value to be learned is new in every period, we assume, for simplicity, that the stochastic structure underlying the signals that guide learning of t, is the same in every period. Since our goal is to produce an algorithm for running network simulations, the datagenerating process of Ii,t = (yt, xi,t) in every period needs a ìtrueîparameter, unknown to players in the model, to be used by a modeler.