Public Discourse and Social Network Echo Chambers Driven by Socio-Cognitive Biases

In recent years, social media has become an important platform for political discourse, being a site of both political conversations between voters and political advertisements from campaigns. While their individual influences on public discourse are well documented, the interplay between individual-level cognitive biases, social influence processes, dueling campaign efforts, and social media platforms remains unexamined. We introduce an agent-based model that integrates these dynamics and illustrates how their combination can lead to the formation of echo chambers. We find that the range of political viewpoints that individuals are willing to consider is a key determinant in the formation of polarized networks and the emergence of echo chambers and show that aggressive political campaigns can have counterproductive outcomes by radicalizing supporters and alienating moderates. Our model results demonstrate how certain elements of public discourse and political polarization can be understood as the result of an interactive process of shifting individual opinions, evolving social networks, and political campaigns. We also introduce a dynamic empirical case, retweet networks from the final stage of the 2016 U.S. presidential election, to show how our proposed model can be calibrated with real-world behavior.

Popular Summary

Modern public opinion formation is governed by individual decision-making processes that are prone to cognitive biases, peer influence, social media architectures, and strategically executed political campaigns. While researchers have thoroughly documented how each of these players can sway political discourse, little work has examined how these factors interact in the modern media environment. To that end, we introduce a computational model that shows how psychological, social, and political processes interact and how those interactions can lead to polarization and the formation of echo chambers.

Our model simulates the evolution of individual alignment with one of two opposing political stances. Each individual in a network interacts with other individuals as well as ideological messaging from the two competing campaigns. We see that partisan polarization is more likely to occur among populations who are ideologically flexible and open minded, but not too flexible and open minded. Furthermore, we find that peer influence and campaign influence can interact in counterintuitive ways, as aggressive campaigning can create uncrossable ideological chasms in peer-discussion networks. We integrate our model with empirical data from Twitter during the 2016 U.S. presidential election to illustrate how our model can explain electoral outcomes.

Our work also sheds light on how social media campaigns and online public discourse can interact to create desired social changes on many urgent social and political issues ranging from climate inaction to vaccine refusals.

Figure 1

Evolving echo chambers in core retweet networks. The progression of echo-chamber formation in retweet networks during time frames surrounding (a) the first 2016 U.S. presidential debate and (b) the 2016 U.S. vice presidential debate. In each network subfigure on the left, the nodes represent users, node colors show the nodes cluster, the edges indicate retweets, and node size and color intensity reflect a node’s retweet volume. The bar charts on the right show the growth of support for each candidate during two of the 2016 debates. In (a), the red nodes indicate individuals who tend to support Donald Trump, while blue ones indicate Hillary Clinton supporters. The yellow cluster shows users who are mainly retweeting unaligned users or supporters of both sides. The dataset begins at 17∶45 on September 26, and the final graph includes 54840 edges. In (b), the networks are divided into two polarized echo chambers: The red cluster shows Mike Pence (Trump’s vice president) supporters, and the blue cluster shows Tim Kaine (Clinton’s vice president) supporters. Unlike the first presidential debate, there is not a cluster of undecided supporters, despite both graphs being roughly the same size in terms of nodes and edges. The dataset begins at 15∶40 on October 4, 2016, and the final graph includes 55008 edges.

Figure 2

Modeling the emergence of echo chambers. The relationship between the degree of ideological segregation and population identity scope (i.e., open-mindedness) largely influences the nature and outcome of public discourse. Shifts in identity scope are alone sufficient to drastically change the (a) breakdown of candidate support and (b) the distribution of ties between and within supporter groups. The stable network structures of simulations corresponding to three different identity scope values (c)–(e) show how open-mindedness effects the emergence of echo chambers. Parameters: Campaign effort ${\mathrm{\Omega }}_T=0.5$, position of viewpoint blue ${\theta }_1=0.2$, position of viewpoint red ${\theta }_2=0.7$, influence of candidate ${\mu }_1=0.5$, and influence of discussion partner ${\mu }_2=0.5$. All simulations begin from an Erdös–Rényi (ER) random graph; the simulations for (a) and (b) have network size $N={10}^4$ and average degree $⟨k⟩=6$. Simulation results are averaged over 100 independent runs. In the simulations that generate (c)–(e), $N={10}^3$ and $⟨k⟩=6$.

Figure 3

Confirmation bias and attitude evolution. (a),(b) Generally, moderate discussion partner influence (peer influence) and weak candidate influence lead to larger echo chambers, in this case a significantly higher proportion of red supporters and a lower proportion of undecided individuals. (c) Strong levels of candidate influence may increase the belief distance between supporters and undecided people and accelerate opinion polarization, which hinders sufficient social discourse and results in smaller echo chambers around candidates. (d) When candidate influence is fixed, moderate levels of discussion partner influence promote more social discourse, which reduces the size of the undecided population and effectively prevents the formation of small clusters that hold extreme opinions [also see (f)]. (e),(f) Distribution of individual beliefs in steady states, corresponding to the various situations in (c) and (d). Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\theta }_1=0.2$, ${\theta }_2=0.7$, and population open-mindedness ${\delta }_1=0.18$. In addition, (c) ${\mu }_2=0.5$ and (d) ${\mu }_1=0.5$. Simulation results are averaged over 100 independent runs.

Figure 4

Evolution of individual beliefs under different degrees of confirmation bias. (a)–(c) Strong positive feedback between candidates and voters accelerates the opinion polarizing process and leads to insufficient social discourse where more undecided individuals exist. (d)–(f) Intermediate levels of discussion partner influence increase interactions between individuals of different beliefs, which reduces the undecided people and effectively prevents the formation of small clusters of extremists. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\theta }_1=0.2$, ${\theta }_2=0.7$, and ${\delta }_1=0.18$.

Figure 5

Competing campaign ideologies within social networks. We can also use our model to see how campaigns should ideologically position themselves to maximize voter appeal. We assume that, in the short run, the open-mindedness (identity scope) of voters and levels of discussion partner and candidate influence on voters remains constant, so we fix ${\delta }_1=0.18$ and ${\mu }_1={\mu }_2=0.5$. (a)–(c) show how vote shares for blue, red, and undecided varies as a function of the candidate’s ideological positions. Both candidates maximize vote share by picking opposing positions in the ranges of [0.2, 0.3] and [0.7, 0.8]. In (d),(e), the ideological viewpoint of one campaign is fixed [(d) ${\theta }_1=0.2$ and (e) ${\theta }_2=0.7$] and the other is varied. In (f), the difference between the two campaigns is fixed (${\delta }_0=0.4$), and maximum vote share for both candidates occurs when the campaigns are equidistant from 0.5 (${\theta }_1=0.3$ and ${\theta }_2=0.7$). All simulations begin with an ER graph with $N={10}^4$, $⟨k⟩=6$, and ${\mathrm{\Omega }}_T=0.5$. Simulation results are averaged over 100 independent runs.

Figure 6

Impact of campaign effort on winning majority support. (a)–(e) As the relative attraction of the viewpoints (defined as identity scope divided by candidate spread) in a campaign competition increases ($\rho =0.257$, 0.3, 0.36, 0.56, and 0.76, respectively), the relative efforts of campaigns become more important. In some cases, the effects of campaign effort become counterproductive [see (c)]. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, and ${\mu }_1={\mu }_2=0.5$. Simulation results are averaged over 100 independent runs.

Figure 7

Model-data integration provides insights into the 2016 U.S. presidential election. (a) Retweet networks of the first, second, and third presidential debates in the 2016 U.S. presidential election show the rise of echo chambers that gradually absorb undecided voters. (b) Sankey diagram of individual-level opinion evolution for Twitter users between the first and second debates and the second and third debates. Most Donald Trump or Hillary Clinton supporters are “loyal followers who would not change their political tendency from beginning to end.” (c) Reproduction of opinion evolution processes in the 2016 U.S. presidential election using an agent-based model. We take the Real Clear Politics polling average on September 15, 2016, as an initial estimate of predebate support, while, from the first to third debates, we use the same Twitter data as in (a). Fixing a middle level of confirmation bias (${\mu }_1={\mu }_2=0.5$), we find a group of best-fitting parameters for the ideological positions of campaigns, the identity scope (open-mindedness) of voters, and relative levels of campaign effort: ${\theta }_1=0.32$, ${\theta }_2=0.7$, ${\delta }_1=0.32$, ${\mathrm{\Omega }}_1={\mathrm{\Omega }}_2=0.5$, ${\mathrm{\Omega }}_3=0.8$, ${\mathrm{\Omega }}_4=0.4$, and ${\mathrm{\Omega }}_5=0.1$ (we let ${\mathrm{\Omega }}_6={\mathrm{\Omega }}_5=0.1$ to model one extra time step). Each result in the figure is the average of 100 simulations. (d) Illustration of the potential mechanisms that lead to side switching within this model (in this case, from blue supporter to red supporter). Two possible ways are shown: (1) a blue supporter within ideological range of both candidates switches via one single update as a result of campaigning; (2) switch via multiple updates, a voter is not persuaded by the candidate initially but instead talks to a discussion partner within their range who is a red supporter and then updates their beliefs. In the next time step, the voter is reached by the red campaign again and is close enough to the red candidate to be persuaded.

Figure 8

Data-processing method of retweet networks. Define the weight of each node as the total number of times it is retweeted. In the largest connected component, we get rid of those “pure followers” who are not being retweeted (in degree equal to 0) and retweet others only once (out degree equal to 1) and plot the remaining retweet network with all nodes’ weights recorded. The simplified retweet network has all the important information and better represents the core structural characters.

Figure 9

Emergence of echo chambers with heterogenous levels of population open-mindedness. We relax the assumption of homogenous population open-mindedness (${\delta }_1$) and instead allow a truncated normal distribution of open-mindedness (${\delta }_{1,i}$). As ${\delta }_1^{\prime }$ increases, we observe (similar to the main model) that there are three phase stages divided by two transition points. Two opposing echo chambers are more likely to occur when individual open-mindedness ${\delta }_{1,i}$ surrounds a moderate ${\delta }_1^{\prime }$. Compared to mild heterogeneity (a),(b), strong heterogeneity allows for wavering individuals or independent clusters for slightly higher levels of mean open-mindedness (c),(d). Simulation results are averaged over 30 independent runs. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\theta }_1=0.2$, ${\theta }_2=0.7$, ${\mu }_1=0.5$, and ${\mu }_2=0.5$.

Figure 10

Effect of candidate influence and peer influence on polarization under heterogenous population open-mindedness. We run simulations with mild heterogeneity in open-mindedness (a)–(c) and strong heterogeneity in open-mindedness (d)–(f). Higher levels of heterogeneity change the relationship between candidate influence, peer influence, and polarization. In higher heterogeneity and high-peer-influence conditions, candidate influence has less of a dampening effect on polarization. Furthermore, higher levels of heterogeneity generally lead to less polarization across all combinations of influence parameters. Simulation results are averaged over 30 independent runs. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\theta }_1=0.2$, and ${\theta }_2=0.7$.

Figure 11

Candidate influence and peer influence under polarized initial beliefs. We relax the initial condition of individual beliefs from uniform distribution to (a)–(c) symmetrical bimodal distribution and (d)–(f) asymmetrical bimodal distribution. Simulation results are averaged over 30 independent runs. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\theta }_1=0.2$, ${\theta }_2=0.7$, and ${\delta }_1=0.18$.

Figure 12

Competing campaign ideologies under polarized initial beliefs. The initial condition of individual beliefs is relaxed from a uniform distribution to a symmetrical bimodal distribution (a)–(c) and an asymmetrical bimodal distribution (d)–(f). Results largely resemble findings from the original model: Candidates collectively capture more voters and minimize undecideds by having opposing views. Simulation results are averaged over 30 independent runs. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\mathrm{\Omega }}_T=0.5$, ${\delta }_1=0.18$, ${\mu }_1=0.5$, and ${\mu }_2=0.5$.

Figure 13

Impact of campaign effort on winning majority support under polarized initial beliefs. The initial distribution of individual beliefs is relaxed from uniform to a symmetrical bimodal distribution (a)–(c) and an asymmetrical bimodal distribution (d)–(f). The obvious finding that campaigns can suffer if they exert no effort, along with the counterintuitive finding that campaigning slightly too much can also slightly hurt their campaign, both replicate from the original model. Simulation results are averaged over 30 independent runs. Parameters: All simulations begin from an ER graph with $N={10}^4$, $⟨k⟩=6$, ${\theta }_1=0.2$, ${\theta }_2=0.7$, ${\delta }_1=0.18$, ${\mu }_1=0.5$, and ${\mu }_2=0.5$.

Figure 14

Emergence of undecided individuals. (a),(b) The undecided individual $i$ updates with viewpoint blue (${\theta }_1$) that is beyond its identity scope; therefore, $i$ seeks friend $j$ for interpersonal interaction. (b),(c) The undecided individual $i$ updates with viewpoint red (${\theta }_2$) that is also beyond its identity scope and further discusses with friend $k$.

Figure 15

Emergence of wavering supporters. (a),(b) The undecided individual $i$ updates with viewpoint blue and directly becomes a blue supporter with its belief $w_i$ converging to ${\theta }_1$. (b),(c) Blue supporter $i$ updates with viewpoint red and directly becomes a red supporter with its belief $w_i$ converging to ${\theta }_2$.