News Sharing, Content Activation and Perceived Polarization on Social Media

# News Sharing, Content Activation and Perceived Polarization on Social Media
### Tiago Ventura
### CSMaP - 02/11/2022

---

## .red[About me]

### Tiago Ventura (He/Him)

### Researcher Civic Integrity at Twitter

[<svg role="img" viewBox="0 0 24 24" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <title></title>  <path d="M23.953 4.57a10 10 0 01-2.825.775 4.958 4.958 0 002.163-2.723c-.951.555-2.005.959-3.127 1.184a4.92 4.92 0 00-8.384 4.482C7.69 8.095 4.067 6.13 1.64 3.162a4.822 4.822 0 00-.666 2.475c0 1.71.87 3.213 2.188 4.096a4.904 4.904 0 01-2.228-.616v.06a4.923 4.923 0 003.946 4.827 4.996 4.996 0 01-2.212.085 4.936 4.936 0 004.604 3.417 9.867 9.867 0 01-6.102 2.105c-.39 0-.779-.023-1.17-.067a13.995 13.995 0 007.557 2.209c9.053 0 13.998-7.496 13.998-13.985 0-.21 0-.42-.015-.63A9.935 9.935 0 0024 4.59z"></path></svg> @TiagoVentura_](https://twitter.com/_Tiagoventura)
[<svg role="img" viewBox="0 0 24 24" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <title></title>  <path d="M12 .297c-6.63 0-12 5.373-12 12 0 5.303 3.438 9.8 8.205 11.385.6.113.82-.258.82-.577 0-.285-.01-1.04-.015-2.04-3.338.724-4.042-1.61-4.042-1.61C4.422 18.07 3.633 17.7 3.633 17.7c-1.087-.744.084-.729.084-.729 1.205.084 1.838 1.236 1.838 1.236 1.07 1.835 2.809 1.305 3.495.998.108-.776.417-1.305.76-1.605-2.665-.3-5.466-1.332-5.466-5.93 0-1.31.465-2.38 1.235-3.22-.135-.303-.54-1.523.105-3.176 0 0 1.005-.322 3.3 1.23.96-.267 1.98-.399 3-.405 1.02.006 2.04.138 3 .405 2.28-1.552 3.285-1.23 3.285-1.23.645 1.653.24 2.873.12 3.176.765.84 1.23 1.91 1.23 3.22 0 4.61-2.805 5.625-5.475 5.92.42.36.81 1.096.81 2.22 0 1.606-.015 2.896-.015 3.286 0 .315.21.69.825.57C20.565 22.092 24 17.592 24 12.297c0-6.627-5.373-12-12-12"></path></svg> TiagoVentura](https://github.com/TiagoVentura)
[<svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <title></title>  <path d="M424,80H88a56.06,56.06,0,0,0-56,56V376a56.06,56.06,0,0,0,56,56H424a56.06,56.06,0,0,0,56-56V136A56.06,56.06,0,0,0,424,80Zm-14.18,92.63-144,112a16,16,0,0,1-19.64,0l-144-112a16,16,0,1,1,19.64-25.26L256,251.73,390.18,147.37a16,16,0,0,1,19.64,25.26Z"></path></svg> venturat@umd.edu](venturat@umd.edu)
[<svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <title></title>  <path d="M208,352H144a96,96,0,0,1,0-192h64" style="fill:none;stroke:#000;stroke-linecap:round;stroke-linejoin:round;stroke-width:36px"></path>  <path d="M304,160h64a96,96,0,0,1,0,192H304" style="fill:none;stroke:#000;stroke-linecap:round;stroke-linejoin:round;stroke-width:36px"></path>  <line x1="163.29" y1="256" x2="350.71" y2="256" style="fill:none;stroke:#000;stroke-linecap:round;stroke-linejoin:round;stroke-width:36px"></line></svg>https://tiagoventura.rbind.io/](https://tiagoventura.rbind.io/)

---
class:middle 
## Plans for the talk

#### .red[Book Project:] News sharing, content activation and perceived polarization on social media

- [News Sharing, Gatekeeping and Polarization (*Digital Journalism*)](https://www.tandfonline.com/doi/abs/10.1080/21670811.2020.1852094)

- [News by Popular Demand (*IJPP*)](https://journals.sagepub.com/doi/abs/10.1177/19401612211057068)

- [Network Activated Frames (Under Review)](https://tiagoventura.rbind.io/files/naf.pdf)

---
class:middle

## Plans for the talk

#### .red[Other projects] on my research agenda

- Political Communication

- New Methods in CSS

- Comparative Political Behavior

---
class:inverse, middle, center

## News Sharing, Content Activation and Perceived Polarization on Social Media

---
class:middle

### .center[.blue[Social media increases sorting]]

### .center[<svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M256 8C119 8 8 119 8 256s111 248 248 248 248-111 248-248S393 8 256 8zm0 448c-110.5 0-200-89.5-200-200S145.5 56 256 56s200 89.5 200 200-89.5 200-200 200zm-32-316v116h-67c-10.7 0-16 12.9-8.5 20.5l99 99c4.7 4.7 12.3 4.7 17 0l99-99c7.6-7.6 2.2-20.5-8.5-20.5h-67V140c0-6.6-5.4-12-12-12h-40c-6.6 0-12 5.4-12 12z"></path></svg>]

### .center[.blue[Echo Chambers]]

#### .center[<svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M256 8C119 8 8 119 8 256s111 248 248 248 248-111 248-248S393 8 256 8zm0 448c-110.5 0-200-89.5-200-200S145.5 56 256 56s200 89.5 200 200-89.5 200-200 200zm-32-316v116h-67c-10.7 0-16 12.9-8.5 20.5l99 99c4.7 4.7 12.3 4.7 17 0l99-99c7.6-7.6 2.2-20.5-8.5-20.5h-67V140c0-6.6-5.4-12-12-12h-40c-6.6 0-12 5.4-12 12z"></path></svg>]

### .center[.blue[Reduce cross-cutting exposure]]

### .center[.red[Polarization]]

---
class: middle

### Are social media echo-chambers real?

- Online Sorting is similar to offline levels .red[(Gentzkow and Shapiro, 2011; Wojcieszak  and  Mutz, 2009; Bisbee and Larsson, 2017)]

- Users' friendship networks are heterogeneous outside of politics .red[(Bakshy et al. 2012; Barbéra  et  al., 2 015)]

- Users’ digital media diets are balanced, and strongly influenced by big reputable outlets .red[(Guess 2021; Cardenal et. al., 2019)]

---
class:inverse, middle, center

# <svg viewBox="0 0 384 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M202.021 0C122.202 0 70.503 32.703 29.914 91.026c-7.363 10.58-5.093 25.086 5.178 32.874l43.138 32.709c10.373 7.865 25.132 6.026 33.253-4.148 25.049-31.381 43.63-49.449 82.757-49.449 30.764 0 68.816 19.799 68.816 49.631 0 22.552-18.617 34.134-48.993 51.164-35.423 19.86-82.299 44.576-82.299 106.405V320c0 13.255 10.745 24 24 24h72.471c13.255 0 24-10.745 24-24v-5.773c0-42.86 125.268-44.645 125.268-160.627C377.504 66.256 286.902 0 202.021 0zM192 373.459c-38.196 0-69.271 31.075-69.271 69.271 0 38.195 31.075 69.27 69.271 69.27s69.271-31.075 69.271-69.271-31.075-69.27-69.271-69.27z"></path></svg>

###  Most users are embedded in diverse online social networks where moderation is the norm, yet  perceptions of a highly polarized social media environment are still particularly widespread.

---
class: middle

## Our Contributtions

--
#### <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M504 256C504 119 393 8 256 8S8 119 8 256s111 248 248 248 248-111 248-248zm-448 0c0-110.5 89.5-200 200-200s200 89.5 200 200-89.5 200-200 200S56 366.5 56 256zm72 20v-40c0-6.6 5.4-12 12-12h116v-67c0-10.7 12.9-16 20.5-8.5l99 99c4.7 4.7 4.7 12.3 0 17l-99 99c-7.6 7.6-20.5 2.2-20.5-8.5v-67H140c-6.6 0-12-5.4-12-12z"></path></svg> Activation vs Sorting: .red[*Propagation on social media depends fundamentally on the users’ sharing decision*]

#### <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M504 256C504 119 393 8 256 8S8 119 8 256s111 248 248 248 248-111 248-248zm-448 0c0-110.5 89.5-200 200-200s200 89.5 200 200-89.5 200-200 200S56 366.5 56 256zm72 20v-40c0-6.6 5.4-12 12-12h116v-67c0-10.7 12.9-16 20.5-8.5l99 99c4.7 4.7 4.7 12.3 0 17l-99 99c-7.6 7.6-20.5 2.2-20.5-8.5v-67H140c-6.6 0-12-5.4-12-12z"></path></svg> .red[*Partisan users*] are more issue-motivated <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M504 256C504 119 393 8 256 8S8 119 8 256s111 248 248 248 248-111 248-248zm-448 0c0-110.5 89.5-200 200-200s200 89.5 200 200-89.5 200-200 200S56 366.5 56 256zm72 20v-40c0-6.6 5.4-12 12-12h116v-67c0-10.7 12.9-16 20.5-8.5l99 99c4.7 4.7 4.7 12.3 0 17l-99 99c-7.6 7.6-20.5 2.2-20.5-8.5v-67H140c-6.6 0-12-5.4-12-12z"></path></svg>  .red[*partisan content*] will be over-represented <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;" xmlns="http://www.w3.org/2000/svg">  <path d="M504 256C504 119 393 8 256 8S8 119 8 256s111 248 248 248 248-111 248-248zm-448 0c0-110.5 89.5-200 200-200s200 89.5 200 200-89.5 200-200 200S56 366.5 56 256zm72 20v-40c0-6.6 5.4-12 12-12h116v-67c0-10.7 12.9-16 20.5-8.5l99 99c4.7 4.7 4.7 12.3 0 17l-99 99c-7.6 7.6-20.5 2.2-20.5-8.5v-67H140c-6.6 0-12-5.4-12-12z"></path></svg>  the average user experiences .red[*more partisan content*] than the prevalence of partisans.

---
class: middle, inverse, center

## News Sharing Model: Modeling Sharing Behavior on Social Media Data

---
class: middle
## News Sharing: Data Collection

**Cases**:  .red[Brazil, Argentina, and the United States]

**Source**: Twitter APIs, Forward stream and the backward search (Using Twarc)

**Pre-Processing**:

- Built a network of retweets with the largest connected clusters.

- Draw users’ [x,y] coordinates implementing the Fruchterman-Reingold algorithm

- Walk-trap community detection algorithm.

- Regex query search for embedded hyperlinks (DV).

#### Close to 10 millions retweets + 500 thousand accounts.

---
### Activation Bubbles in Social Media

]

---
## News Sharing: Model

---

## News Sharing: Estimation

`\(\mu_i = \exp(\alpha_{i[q]}\left(x_{i}-L_{j}\right)^2 + A_{[q]} + R_{[j]} + \gamma_{[i]})\)`
]

---
## Results: Ideology

---
## Results: Attention

---
## Media Equilibrium Positions

**Using Adams, Merill, and Grofman's (2012) solution:**

---

## Conclusion: Attention and Ideology

- More partisan users share more content

- Partisan content will be over-represented.

- Social media bubbles will emerge from activation, even if networks are heterogenous.

]

---

## .red[*Observational data cannot fully separate activation from sorting.]*

**Solution**: Image-Based Conjoint Experiments

**Advantages**:

- Flexible: Ideal for social media experiments.

- Equal probability for all the frames

- High ecological validity

---
## Image-Based Conjoints

]

]
]

---
class: middle

#### .red[Hypothesis]

]

- *H1: Partisan users will be unconditionally more issue motivated than non-partisan voters to share political content.*

- *H2: Users will share congruent content that aligns politically with the preferences of their co-partisans (in-group cognitive congruence)*

]

---
## H1: Sharing-Rate

---
## H2: Framing Activation

---
class: middle
## Main Take Aways

---
class:center, middle, inverse

## Overall Research Agenda

#### *.red[And its CSS applications]*

---
## Overall Research Agenda

<br>

.panelset.sideways[

**Streaming Chats**:

- [Political Effects of Streaming Chats](https://osf.io/r6muc/): Field Digital Experiment

- [Toxicity and Streaming Chats on Facebook](https://journalqd.org/article/view/2573): Develop scrappers to collect more 100k chatboxes' comments from Facebook

**Misinformation**

- [Truth be Told: Cognitive moderators of selective sharing of fact-checks on social media](https://tiagoventura.rbind.io/files/ABCVV_truth_be_told.pdf): Collaboration with Chequeado + Large Scale Data Collection + Survey Experiment

- [Doubt the Messenger: The reputation cost of fact-checking](https://mfr.osf.io/render?url=https://osf.io/ya274/?direct%26mode=render%26action=download%26mode=render)

]

- Strategies to work with big social media network data: [The Path-Weighted Regression Model](https://methods.sagepub.com/book/research-methods-in-political-science-and-international-relations/i4823.xml)

- Image-Based Behavioral Experiments: [social media activation](https://www.dropbox.com/s/r23wwboo6mirpq0/naf.pdf?dl=0)

- Natural Language Processing: Bert-Based Topic Models using pre-trained political contextualized embeddings.

]

As a Comparativist, I also do some work focused on issues of **criminal violence and security policies** in Latin America.

- [A Network Approach to Crime Victimization](https://tiagoventura.rbind.io/files/net.pdf): Novel network strategies to model overdispersion of crime.

- [Who owns the issue of security in fragmented democracies?](https://www.dropbox.com/s/e7p9v4ezmpf4ppt/Legislating_for_Violence.pdf?dl=0): 140k Legislative Speeches (Audio + Text) - Topic Models + NLP to detect emotions.

- [Criminal Organizations and provision of public goods during Covid-19 pandemic](https://egap.org/project/criminal-governance-amid-the-covid-19-pandemic-explaining-violence-and-goods-provision-as-public-health-responses-by-organized-crime-groups-in-mexico/): Facebook Ads to target hard-to-reach population

]
]

---
class:inverse, middle

---
class:inverse, center, middle

## Thank you

---
class:inverse, center, middle

## Extra Slides

---
class: middle, center
## Policy Implications

#### Slowing Sharing Behavior: Double-Click Retweet.

---
#### Path-Weighted Regression Model (With J. Timoneda and E. Calvo)

- Most network models are to computationally intensive for large networks.

- Solution: Model Locally + Weighted by Closest Paths

- Extension of Geographically Weighted Regression Model

- Use Cross-Validation to Determine Optimal Number of Paths

- Allows to model heterogeneity on networks and easy to paralelize.
]

CB Ford’s testimony. time to retweet ~ In-degree

]
]
]

---
class:middle, center, inverse

## NAF Extra
---

## Brazil Conjoint

---
## Framing Activation (Partisans vs Non-Partisans)

---
## Expectations in the Network

---

class:middle, center, inverse

## Argentina Conjoint

---
## Framing Activation (Partisans vs Non-Partisans)

---
## Expectations in the Network

---
class:inverse, middle, center

## Mexico Conjoint

---

## Framing Activation (Partisans vs Non-Partisans)

---
## Expectations in the Network

---

class:inverse, middle, center

## Streaming Chats: Extra

---

## New Technologies: Streaming Chats

**Video Feed + Social Chat: All in one screen.**

Popular among the younger generations.

- Twitch (Amazon)
- YouTube/YouTube Gaming (Google)
- Mixer (Microsoft)
- Facebook Lives
]
.pull-right[
<img src="./figs/genelectionabc.png" width="80%" />
]
]

We conducted a large scale ''field" experiment that assigns would-be debate viewers to watch on different platforms the October 2019 Democratic Debate.

Two-Wave On-line Survey in September 2019 through MTurk (following Gross, Porter and Wood, 2019).

**Three main experimental conditions**

- Control (standard NBC broadcast)
- Expert chat (538 website)
- Streaming chat (Facebook)

We then surveyed the respondents after the debate and measured the effects of streaming chats. 
]

#### Text Analysis

To complement our analysis, we built automatic scrapper (Selenium) to collect comments on facebook pages

- One Hundred Thousand Comments  from Facebook pages.

- Several Expert Chats

- Performed Dictionary Methods Sentiment Analysis.

- Deep Learning Models to identify toxicity.

]

#### Grant under review to:

- Deploy a similar experiment in Brazil, during the Presidential Election

- Combine the field experiments with collection of web-browsing data

- Measure behavioral mechanisms with a laboratory experiment

]
]
---
class: middle, center, inverse

## Results

---
### High Frequency

.center[
<img src="figs/proportion_toxicity.png" width="80%" />
].footnote[Deep Learning Models from Google Perspective API]

---
### Contains Mostly Negative Primes

---
#### Mostly Negative Polarity about the Candidates

---
class:inverse, middle, center

## Experimental Results

---
## Frequency and Toxicity Hypotheses
.center[
Facebook chat somewhat less informative, enjoyable, and engaging
<img src="figs/frequencyplot.png" width="80%" />
]

---
## Feeling Thermomethers about the Candidates

---
## Content Effect
.center[
<img src="figs/ft_on_negative_comments.png" width="80%" />
]

---
## Poll Performance

---
## Context Effect
.center[
<img src="figs/polls_on_positive_comments.png" width="80%" />
]

---
## Summary

- Creates worse viewing experience.

- May disproportionately negatively affect certain candidates subject to toxic, negative comments.

- May distort inferences about candidate viability.

]

- More research on the mechanism (laboraty experiments).

- More comparative evidence (Does it replicate to other plataforms or other countries?).

- More descriptive evidence about these news technologies.

]

---
class:middle, center, inverse

### Comparative Politics, Violence and CSS

---

## CSS and Political Violence

140 thousand legislative speeches in Brazil (text + audio)

- Topic modeling to detect speeches about violence + ML Models to detect Issue Ownership.

**Next Steps**: Modern NLP Techniques (Word Embeddings)

- Emotions in context (word2vec + cossine similarity + dictionary)

- Scaling of policy preferences (doc2vec)

]

]
]

Novel network models to estimate contextual victimization

- How many friends do you know that had a kid last year?

- How many Silvia do you know?

- How many friend do you know that work as teachers?

- .red[How many friends do you know that were victims of crime?]

]

**EGAP COVID-19 grant**

<img src="./figs/ads.jpeg" width="60%" />
]
]
]
]

---

## Statistical Model for News Sharing

The estimation of the utility function for sharing uses a multilevel overdispersed poisson model.

`\(\mu_i = \exp(\alpha_{i[q]}\left(x_{i}-L_{j}\right)^2 + A_{[q]} + R_{[j]} + \gamma_{[i]})\)`
]

Where:

* `\(y_{i}\)` = Number of links embedded by user x media

* `\(\alpha_{q}\)` = Random Slope by quantile

* `\(A_{q}\)` = Random intercept by quantile

* `\(R_{j}\)` = Random intercept by media outlet

* `\(\gamma_{[i]}\)` = Overdispersion parameter

---
# Model Winbugs