Using Facebook Data to Predict 2016 US Presidential Election

Keng-Chi Chang Chun-Fang Chiang Ming-Jen Lin

Department of Economics National Taiwan University

2018-05-29

Prepared for Innovations in Political Methodology and China Study International Conference

0 / 32 ▸ But people also consume news and process info. through posts • We use 19B “likes” on posts of 2K US fan pages to scale ideology ▸ Also account for media, interest groups, parties, etc, and users ▸ Pages share similar ideology should share “likes” from similar users ▸ Adds time, post content, and region (guessed states) dimensions • We predict 2016 US presidential election using this measure ▸ Derive state level FB support rates based on spatial model ▸ Compare with actual vote shares and state polls • We nd under minimal assumptions, Facebook support rates: ▸ Predicts election quite well and shares similar trends with polls ▸ Overestimates winner’s vote share, but may enhance prediction

In This Paper

• Previous social media ideology measures ▸ Are mostly for elites, and uses “following” of a fan page

1 / 32 • We use 19B “likes” on posts of 2K US fan pages to scale ideology ▸ Also account for media, interest groups, parties, etc, and users ▸ Pages share similar ideology should share “likes” from similar users ▸ Adds time, post content, and region (guessed states) dimensions • We predict 2016 US presidential election using this measure ▸ Derive state level FB support rates based on spatial model ▸ Compare with actual vote shares and state polls • We nd under minimal assumptions, Facebook support rates: ▸ Predicts election quite well and shares similar trends with polls ▸ Overestimates winner’s vote share, but may enhance prediction

In This Paper

• Previous social media ideology measures ▸ Are mostly for elites, and uses “following” of a fan page ▸ But people also consume news and process info. through posts

1 / 32 • We predict 2016 US presidential election using this measure ▸ Derive state level FB support rates based on spatial model ▸ Compare with actual vote shares and state polls • We nd under minimal assumptions, Facebook support rates: ▸ Predicts election quite well and shares similar trends with polls ▸ Overestimates winner’s vote share, but may enhance prediction

In This Paper

1 / 32 • We nd under minimal assumptions, Facebook support rates: ▸ Predicts election quite well and shares similar trends with polls ▸ Overestimates winner’s vote share, but may enhance prediction

In This Paper

1 / 32 In This Paper

• Previous social media ideology measures ▸ Are mostly for elites, and uses “following” of a fan page ▸ But people also consume news and process info. through posts • We use 19B “likes” on posts of 2K US fan pages to scale ideology ▸ Also account for media, interest groups, parties, etc, and users ▸ Pages share similar ideology should share “likes” from similar users ▸ Adds time, post content, and region (guessed states) dimensions • We predict 2016 US presidential election using this measure ▸ Derive state level FB support rates based on spatial model ▸ Compare with actual vote shares and state polls • We nd under minimal assumptions, Facebook support rates: ▸ Predicts election quite well and shares similar trends with polls ▸ Overestimates winner’s vote share, but may enhance prediction 1 / 32 • Specify fan page ideological universe ▸ 1475 fan pages of national politicians ↝ Members and candidates of Senate, House, and Governors ▸ Top 1000 pages related to 2016 presidential election ↝ In Aug 2016, nd all pages mentioned “Trump” and “Clinton” ↝ Weight by likes, comments, shares, nd top 1000 pages ↝ Includes all major news outlets, interest groups, parties, etc ↝ NYT, Fox News, NRA, RNC, Occupy Wall St, Tea Party, 9GAG, ... • Collect all 24M posts in 2015 and 2016 on these pages • And user’s 19B reactions (mostly likes) to these posts

Facebook Data

• Facebook provides fan page data through Graph API

2 / 32 ▸ Top 1000 pages related to 2016 presidential election ↝ In Aug 2016, nd all pages mentioned “Trump” and “Clinton” ↝ Weight by likes, comments, shares, nd top 1000 pages ↝ Includes all major news outlets, interest groups, parties, etc ↝ NYT, Fox News, NRA, RNC, Occupy Wall St, Tea Party, 9GAG, ... • Collect all 24M posts in 2015 and 2016 on these pages • And user’s 19B reactions (mostly likes) to these posts

Facebook Data

• Facebook provides fan page data through Graph API • Specify fan page ideological universe ▸ 1475 fan pages of national politicians ↝ Members and candidates of Senate, House, and Governors

2 / 32 ↝ Includes all major news outlets, interest groups, parties, etc ↝ NYT, Fox News, NRA, RNC, Occupy Wall St, Tea Party, 9GAG, ... • Collect all 24M posts in 2015 and 2016 on these pages • And user’s 19B reactions (mostly likes) to these posts

Facebook Data

2 / 32 • Collect all 24M posts in 2015 and 2016 on these pages • And user’s 19B reactions (mostly likes) to these posts

Facebook Data

• Facebook provides fan page data through Graph API • Specify fan page ideological universe ▸ 1475 fan pages of national politicians ↝ Members and candidates of Senate, House, and Governors ▸ Top 1000 pages related to 2016 presidential election ↝ In Aug 2016, nd all pages mentioned “Trump” and “Clinton” ↝ Weight by likes, comments, shares, nd top 1000 pages ↝ Includes all major news outlets, interest groups, parties, etc ↝ NYT, Fox News, NRA, RNC, Occupy Wall St, Tea Party, 9GAG, ...

2 / 32 • And user’s 19B reactions (mostly likes) to these posts

Facebook Data

2 / 32 Facebook Data

2 / 32 Data Summary

Time Period 2015-01-01 to 2016-11-30 Total Reactions 19,085,783,534 US Political User Likes 16,180,488,916 Total Users 366,840,068 US Political Users 29,412,610 Total Posts 24,788,093 Total Pages 2132 Politicians 1225 News Outlets 560 Political Groups 211 Other Public Figures 93 Others 43

3 / 32 • First build the page by page afliation matrix A ↝ Number of shared users (based on likes) between pages

Trump FoxNews TeaParty Clinton CNN NYTimes

Trump 2243216 1078513 128225 32731 120963 25842 FoxNews 1078513 2449174 148016 87084 186850 63401 TeaParty 128225 148016 242089 1528 10738 2162 Clinton 32731 87084 1528 1768980 351210 367021 CNN 120963 186850 10738 351210 1201156 216163 NYTimes 25842 63401 2162 367021 216163 986613

Estimation: Shared Users Matrix

• Measure ideology of pages, then measure those of users ↝ Similar to Bond and Messing (2015, APSR)

4 / 32 Estimation: Shared Users Matrix

• Measure ideology of pages, then measure those of users ↝ Similar to Bond and Messing (2015, APSR) • First build the page by page afliation matrix A ↝ Number of shared users (based on likes) between pages

Trump FoxNews TeaParty Clinton CNN NYTimes

4 / 32 Estimation: Transform to Ratios

= / • Transform A to matrix of ratios G, where gi j ai j aii Pr (Trump ∩ FoxNews) ↝ 0.44 = = Pr (Trump ∣ FoxNews) Pr (FoxNews) • Can interpret columns as features and rows as observations ↝ Col 1 is how each row similar to “Trump” feature

Trump FoxNews TeaParty Clinton CNN NYTimes

Trump 1.00 0.48 0.06 0.01 0.05 0.01 FoxNews 0.44 1.00 0.06 0.04 0.08 0.03 TeaParty 0.53 0.61 1.00 0.01 0.04 0.01 Clinton 0.02 0.05 0.00 1.00 0.20 0.21 CNN 0.10 0.16 0.01 0.29 1.00 0.18 NYTimes 0.03 0.06 0.00 0.37 0.22 1.00

5 / 32 • PC1 is the dimension explains the largest variation ↝ Unsupervised ⇒ Guess and verify PC1 is related to “ideology” • User ideology = mean ideology of pages user liked • Guess user’s state residence by their likes on national politicians ↝ Like more politicians from NY ⇒ More likely from NY

Estimation: Dimension Reduction

• Compute the principal components of G after standardizing

6 / 32 • User ideology = mean ideology of pages user liked • Guess user’s state residence by their likes on national politicians ↝ Like more politicians from NY ⇒ More likely from NY

Estimation: Dimension Reduction

• Compute the principal components of G after standardizing

• PC1 is the dimension explains the largest variation ↝ Unsupervised ⇒ Guess and verify PC1 is related to “ideology”

6 / 32 Estimation: Dimension Reduction

• Compute the principal components of G after standardizing

• PC1 is the dimension explains the largest variation ↝ Unsupervised ⇒ Guess and verify PC1 is related to “ideology” • User ideology = mean ideology of pages user liked • Guess user’s state residence by their likes on national politicians ↝ Like more politicians from NY ⇒ More likely from NY

6 / 32 Scree Plot for Principal Component Analysis xplained .04 0.06 Proportion of Variance E of Variance Proportion 0.00 0.02 0

0 5 10 15 20

k-th Principal Component

7 / 32 Density 0.00 0.25 0.50 0.75 2-1 -2 Public Figure Estimated Facebook Ideo

Clinton

NYTimes Political Groups 1 0 Fox News

logy Scorelogy Trump News OutletsNews 2 32 / 8 Density 0.0 0.3 0.6 0.9 15-1.0 -1.5

The New York Times Politics and Washington

The New York Times Opinion Section of Density PC1 Newspap PC1 (First Principal Com

Washington Post

050.0 -0.5 The New York Times

Chicago Tribune The Wall Street Journal

USA TODAY

ponent) Boston Herald er Pages

. 1.0 0.5 The Christian Post

The Washington Times 32 / 9 Density 0.0 0.5 1.0 1.5 2.0 2-1 -2 PC1 Density of Density PC1 TV, Radio

The Rachel Maddow Fan Page. PC1 (First Principal Com MSNBC PBS

CNN

1 0 ABC News

Fox News , Website Pages ponent) Breitbart The Federalist Papers

NRA News Fox News Opinion 2 type_sub website tv radio 0/32 / 10 Validation for Congressional Politicians

1.0 Cruz Democratic Party Independent Rubio Republican Party 0.5 14th Congress McCain McConnell Ryan 0.0

Schumer

-0.5 Pelosi ρ = 0.92 Sanders ρR = 0.50

DW-Nominate Score of 1 of Score DW-Nominate ρ = 0.22 Warren Booker D -1.0

-2 -1 0 1 2 Estimated Facebook Page Ideology Score, 2015-01 to 2016-11 Using politician and top 1000 page matrix 11 / 32 Validation for Media

Magazine Newspaper Radio TV Website

Fox News The Blaze Breitbart

1 WSJ Opinion National Review AWM

liated User liated USA Today The Hill

0.5 WSJ ABC News MSNBC CNN

Politico WashPost BuzzFeed

0 Share ofRepublican-Aﬀi Share New Republic NYTimes New Yorker

-2 -1 0 1 2 Estimated Facebook Ideology Score 12 / 32 User Ideology Density by States

Massachusetts Washington 2.0 1.00 1.5 0.75 1.0 0.50

0.5 0.25

0.0 0.00 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Michigan Pennsylvania

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Texas Wyoming 1.2 1.00

0.9 0.75

0.6 0.50

0.3 0.25

0.0 0.00 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

13 / 32 User Ideology Density by States

Massachusetts Washington 30 2.5

2.0 20 1.5

1.0 10 0.5

0 0.0 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Michigan Pennsylvania 1.5

2 1.0

1 0.5

0 0.0 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Texas Wyoming 8 6

6 4 4

2 2

0 0 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Politician-Only Method (Bond and Messing 2015) 14 / 32 Media Ideology Dynamics

2 CNSNews.com NRA News

Federalist Papers

1 Breitbart logy Score

Fox News

0 ABC News WSJ CNN Bloomberg New York Times

Estimated Facebook Ideo Facebook Estimated Washington Post -1 MSNBC

ThinkProgress.com

06-01 08-01 10-01 12-01 02-01 04-01 06-01 08-01 10-01 12-01 02-01 04-01 2015 2015 2015 2015 2016 2016 2016 2016 2016 2016 2017 2017

15 / 32 Politician Ideology Dynamics

Cruz Rubio 1.0 Cruz Rubio Ryan Ryan Trump 0.5 Trump logy Score

0.0 Johnson Johnson

-0.5 Sanders Clinton Clinton

Estimated Facebook Ideo Facebook Estimated Sanders Warren Warren -1.0

05-01 07-01 09-01 11-01 01-01 03-01 05-01 07-01 09-01 11-01 01-01 2015 2015 2015 2015 2016 2016 2016 2016 2016 2016 2017

16 / 32 • In each state, we compare: ▸ FB support rate: Share of user’s ideology closer to Trump or Clinton ↝ May not be a precise estimator for vote shares ↝ Since turnouts may not be the same across states ↝ But adding assumptions may look like tting the data ▸ Polls: State polling averages calculated by FiveThirtyEight ▸ Actual vote shares in 2016 election

FB Support Rates, Polls, and Vote Shares

• Apply the Hotelling-Downs spatial model for voting: Voters support candidates closer to their own ideological location

17 / 32 ↝ May not be a precise estimator for vote shares ↝ Since turnouts may not be the same across states ↝ But adding assumptions may look like tting the data ▸ Polls: State polling averages calculated by FiveThirtyEight ▸ Actual vote shares in 2016 election

FB Support Rates, Polls, and Vote Shares

17 / 32 ▸ Polls: State polling averages calculated by FiveThirtyEight ▸ Actual vote shares in 2016 election

FB Support Rates, Polls, and Vote Shares

• Apply the Hotelling-Downs spatial model for voting: Voters support candidates closer to their own ideological location • In each state, we compare: ▸ FB support rate: Share of user’s ideology closer to Trump or Clinton ↝ May not be a precise estimator for vote shares ↝ Since turnouts may not be the same across states ↝ But adding assumptions may look like tting the data

17 / 32 ▸ Actual vote shares in 2016 election

FB Support Rates, Polls, and Vote Shares

17 / 32 FB Support Rates, Polls, and Vote Shares

17 / 32 Predicting Vote Shares and Outcomes

HI CA Rep wins 2016 & 2012 MD VT 60 Swings from Obama to Trump NY MA

Dem wins 2016 & 2012 RI IL WA NJ CT DE OR VA 50 NM ME PA NV WI CO MN NC NH GA AZ FL MI TX OH IA SC MS 40 IN LA MO AK KS MT TN NE AR AL SD 2016 Clinton Vote Share Vote 2016 Clinton KY 30 OK UT ID ND WV ρ = 0.73

WY 95% CI [0.56, 0.84]

25 50 75 Share of Facebook User Closer to Clinton (10-01 to 11-07) 18 / 32 Florida 29 Trump ○ × × × Pennsylvania 20 Trump ○ × × × Wisconsin 10 Trump ○ × × × Michigan 16 Trump × × × × Ohio 18 Trump ○ ○ ○ ○ Iowa 6 Trump ○ ○ ○ ○ Montana 3 Trump × ○ ○ ○ Alaska 3 Clinton × ○ ○ ○ Maine 2 Clinton × ○ ○ ○ Trump’s Electoral Vote 306 292 235 216 215

† Electoral Votes. * Princeton Election Consortium.

Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

19 / 32 Michigan 16 Trump × × × × Ohio 18 Trump ○ ○ ○ ○ Iowa 6 Trump ○ ○ ○ ○ Montana 3 Trump × ○ ○ ○ Alaska 3 Clinton × ○ ○ ○ Maine 2 Clinton × ○ ○ ○ Trump’s Electoral Vote 306 292 235 216 215

† Electoral Votes. * Princeton Election Consortium.

Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

Florida 29 Trump ○ × × × Pennsylvania 20 Trump ○ × × × Wisconsin 10 Trump ○ × × ×

19 / 32 Ohio 18 Trump ○ ○ ○ ○ Iowa 6 Trump ○ ○ ○ ○ Montana 3 Trump × ○ ○ ○ Alaska 3 Clinton × ○ ○ ○ Maine 2 Clinton × ○ ○ ○ Trump’s Electoral Vote 306 292 235 216 215

† Electoral Votes. * Princeton Election Consortium.

Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

Florida 29 Trump ○ × × × Pennsylvania 20 Trump ○ × × × Wisconsin 10 Trump ○ × × × Michigan 16 Trump × × × ×

19 / 32 Montana 3 Trump × ○ ○ ○ Alaska 3 Clinton × ○ ○ ○ Maine 2 Clinton × ○ ○ ○ Trump’s Electoral Vote 306 292 235 216 215

† Electoral Votes. * Princeton Election Consortium.

Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

Florida 29 Trump ○ × × × Pennsylvania 20 Trump ○ × × × Wisconsin 10 Trump ○ × × × Michigan 16 Trump × × × × Ohio 18 Trump ○ ○ ○ ○ Iowa 6 Trump ○ ○ ○ ○

19 / 32 Trump’s Electoral Vote 306 292 235 216 215

† Electoral Votes. * Princeton Election Consortium.

Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

Florida 29 Trump ○ × × × Pennsylvania 20 Trump ○ × × × Wisconsin 10 Trump ○ × × × Michigan 16 Trump × × × × Ohio 18 Trump ○ ○ ○ ○ Iowa 6 Trump ○ ○ ○ ○ Montana 3 Trump × ○ ○ ○ Alaska 3 Clinton × ○ ○ ○ Maine 2 Clinton × ○ ○ ○

19 / 32 Compare with Major Forecasters

Battleground States E.V.† Winner FB 538 NYT PEC*

† Electoral Votes. * Princeton Election Consortium. 19 / 32 Predicting Electoral Votes

Predicted Electoral Votes for Trump 350

Actual 300

270 Even 250 Facebook

200

FiveThirtyEight

150

03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016

20 / 32 Trump: FB (Dotted), Polls, and Vote Shares

Pennsylvania Ohio Nevada 55 55 50 50 50 45 45 45

40 40 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

Maine New Hampshire Florida 50.0 60 47.5 60 45.0 55 50 42.5 50 40 40.0 45

03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

Michigan Wisconsin Iowa 80

48 70 70

44 60 60 50 40 50

40 40 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

21 / 32 Clinton: FB (Dotted), Polls, and Vote Shares

Alabama West Virginia Kentucky 40 40 35 35 35 30 30

30 25 25 20 25 20 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

Tennessee Kansas Arkansas 40 40 40 35 35 35 30 30 30 25 25 25 20 20 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

Idaho Missouri North Dakota 45 60 36 42 50 33 39 40 30 36 30 33 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 03-01 05-01 07-01 09-01 11-01 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016 2016

22 / 32 Polls Overestimates Clinton in Red and Swing States

ID Dem wins 2016 & 2012 WV TN ND Rep wins 2016 & 2012 SD Swings from Obama to Trump KY WY 5 OK MO SC AK WI IA KS Actual Vote Share Vote Actual NE IN OH AR MN MS ME PA AL MI MT NC LA FL CO GA AZ NH VA 0 MD DE TX NV OR UT WA VT NJ RI IL CT NY NM MA CA

Clinton: 538 Vote Share - Share 538 Vote Clinton: HI -5 -20 0 20 40 Clinton: Facebook Support - Actual Vote Share

23 / 32 FB Overestimates Trump in Red and Swing States

HI Dem wins 2016 & 2012 CA Rep wins 2016 & 2012 WA IL MA Swings from Obama to Trump OR VT NY NV NJ NM 0 VA CO CT RI MD DE FL TX Actual Vote Share Vote Actual GA MN LA AZ NH ME NC PA WI ID IA MI AR MS SC MT OH AK AL IN -5 MO KS NE OK KY ND WY TN SD UT Trump: 538 Vote Share - Share 538Vote Trump: WV

-40 -20 0 20 Trump: Facebook Support - Actual Vote Share

24 / 32 • Weaknesses, compared to polls or surveys: ▸ Not representative ↝ Can reweight if more social-demographic information is known ▸ Hard to link with ofine behaviors ↝ Ex. “Strong supporter” vs. “Likely voter” • Can complement each other if more research try to link the two

Discussions

25 / 32 • Can complement each other if more research try to link the two

Discussions

• Strengths of Facebook based prediction: ▸ Revealed preference instead of self-report ▸ Low cost and almost in real time ▸ Trace individuals repeatedly over time ▸ Overestimation for winners can help to make predictions • Weaknesses, compared to polls or surveys: ▸ Not representative ↝ Can reweight if more social-demographic information is known ▸ Hard to link with ofine behaviors ↝ Ex. “Strong supporter” vs. “Likely voter”

25 / 32 Discussions

25 / 32 • If so, what kind of fake stories have larger effect, and why? • Fake news pool on Facebook: ▸ Top 40 fake stories, 536 posts, 130 pages ▸ Posts link to fake domains, 139,074 posts, 177 pages

Working on: Effect of Fake News

• Joint with Chun-Fang Chiang, Brian Knight, and Ming-Jen Lin • Would consuming fake news change people’s ideology or information consumption?

26 / 32 • Fake news pool on Facebook: ▸ Top 40 fake stories, 536 posts, 130 pages ▸ Posts link to fake domains, 139,074 posts, 177 pages

Working on: Effect of Fake News

26 / 32 Working on: Effect of Fake News

• Joint with Chun-Fang Chiang, Brian Knight, and Ming-Jen Lin • Would consuming fake news change people’s ideology or information consumption? • If so, what kind of fake stories have larger effect, and why? • Fake news pool on Facebook: ▸ Top 40 fake stories, 536 posts, 130 pages ▸ Posts link to fake domains, 139,074 posts, 177 pages

26 / 32 Individual Ideology Users Do Not Like Fake Post Users Like Fake Post

2 Density

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Ideology (2016 Jan to Apr)

27 / 32 Individual Ideology Diﬀernece Users Do Not Like Fake Post Users Like Fake Post

4 Density

-0.5 -0.25 0 0.25 0.5 Ideology Diﬀerence (2016 Jul to Nov - 2016 Jan to Apr)

28 / 32 • For each fake post, we: ▸ Find nonfake pages very similar to fake page through different matching methods as control ▸ Find potential followers of these pages, instead of “likes” ▸ Compare the ideology of these fake and nonfake followers before and after fake page unexpectedly started posting fake story = 1( ) + 1( ) Ideologyit α Aftert γ FollowFakei

+ β 1(FollowFakei)1(Aftert) + εit

Strategies for Identication

29 / 32 ▸ Find potential followers of these pages, instead of “likes” ▸ Compare the ideology of these fake and nonfake followers before and after fake page unexpectedly started posting fake story = 1( ) + 1( ) Ideologyit α Aftert γ FollowFakei

+ β 1(FollowFakei)1(Aftert) + εit

Strategies for Identication

29 / 32 ▸ Compare the ideology of these fake and nonfake followers before and after fake page unexpectedly started posting fake story = 1( ) + 1( ) Ideologyit α Aftert γ FollowFakei

+ β 1(FollowFakei)1(Aftert) + εit

Strategies for Identication

29 / 32 = 1( ) + 1( ) Ideologyit α Aftert γ FollowFakei

+ β 1(FollowFakei)1(Aftert) + εit

Strategies for Identication

• Challenges: ▸ People “like” fake post may be very different ▸ Pages posting fake posts may attract very different users ▸ Some stories may be “too fake” for people to believe, even backre • For each fake post, we: ▸ Find nonfake pages very similar to fake page through different matching methods as control ▸ Find potential followers of these pages, instead of “likes” ▸ Compare the ideology of these fake and nonfake followers before and after fake page unexpectedly started posting fake story

29 / 32 Strategies for Identication

+ β 1(FollowFakei)1(Aftert) + εit

29 / 32 "BREAKING: Oﬀicial Set to Testify Against Hillary Found Dead" by Western Journalism 0.85

Users Follow Fake Page 0.80 (Treatment) Parallel Line

0.75

Mean Follower Ideology Follower Mean 0.70

Users Not Follow Fake Page (Control) 0.65

-2 0 2 Week A�er Post 30 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Trump Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

Pentagon furious Clinton nuclear response time ● Clinton financial connection to Saudi Arabia ● Wikileaks: Clinton sold weapons to ISIS ● Pope Francis endorses Trump ● Trump sends own plane to transport marines ● Obama refuses to leave office if Trump elected ● Clinton HIV secret revealed ● Clinton goes to Texas Muslim fundraiser ● Associate to testify against Clinton dead ● Stanford University: Dem election fraud ● Trump protester: I was paid to protest ● Official to testifiy against Clinton dead ● Uncounted Sanders ballots on Clinton server ● Clinton ISIS email leaked ● ISIS leader calls voters support Clinton ● Clinton disqualified holding Federal office ● Clinton tells nuclear launch response time ● Bill Clinton 2000 sex partners, Hillary lesbian ● Billy Graham STUNNING statement on Trump ● Putin: Emails reveal Clinton threatens Sanders ● Graham: Christians must support Trump ● Clinton email reopens, Comey asks immunity ● Clinton to be indicted, prayers answered ●

−0.05 −0.025 0 0.025 0.05 0.075

31 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Trump Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

−0.05 −0.025 0 0.025 0.05 0.075

31 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Trump Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

−0.05 −0.025 0 0.025 0.05 0.075

31 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Trump Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

−0.05 −0.025 0 0.025 0.05 0.075

31 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Clinton Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

Palin to become Trump VP ●

Rupaul: Trump touched me inappropriately ●

Trump critical condition choking own bullshit ●

Ireland accepts Trump refugees ●

Rage Against the Machine anti Trump album ●

Trump: Giving Canada independence mistake ●

Trump U offers Palin honorary climate degree ●

Mexico will close border if Trump elected ●

Trump: I will overtern shocking gay marriage ●

Trump picks Stacey Dash as VP ●

Pence: Michelle Obama most vulgar FLOTUS ●

Palin endorses Cruz ●

Sauron endorses Trump ●

Trump sues Chicago after forced to cancel rally ●

−0.075 −0.05 −0.025 0 0.025

32 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Clinton Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

Palin to become Trump VP ●

Rupaul: Trump touched me inappropriately ●

Trump critical condition choking own bullshit ●

Ireland accepts Trump refugees ●

Rage Against the Machine anti Trump album ●

Trump: Giving Canada independence mistake ●

Trump U offers Palin honorary climate degree ●

Mexico will close border if Trump elected ●

Trump: I will overtern shocking gay marriage ●

Trump picks Stacey Dash as VP ●

Pence: Michelle Obama most vulgar FLOTUS ●

Palin endorses Cruz ●

Sauron endorses Trump ●

Trump sues Chicago after forced to cancel rally ●

−0.075 −0.05 −0.025 0 0.025

32 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Clinton Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

Palin to become Trump VP ●

Rupaul: Trump touched me inappropriately ●

Trump critical condition choking own bullshit ●

Ireland accepts Trump refugees ●

Rage Against the Machine anti Trump album ●

Trump: Giving Canada independence mistake ●

Trump U offers Palin honorary climate degree ●

Mexico will close border if Trump elected ●

Trump: I will overtern shocking gay marriage ●

Trump picks Stacey Dash as VP ●

Pence: Michelle Obama most vulgar FLOTUS ●

Palin endorses Cruz ●

Sauron endorses Trump ●

Trump sues Chicago after forced to cancel rally ●

−0.075 −0.05 −0.025 0 0.025

32 / 32 Story Level Ideology Change for Following Pages Sharing Pro−Clinton Fake News Week +1 to −1, DiD Estimates with Individual Fixed Effects and 99.9% CI

5−NN Matching ● 10−NN Matching 5−Nearest PS Matching 10−Nearest PS Matching

Palin to become Trump VP ●

Rupaul: Trump touched me inappropriately ●

Trump critical condition choking own bullshit ●

Ireland accepts Trump refugees ●

Rage Against the Machine anti Trump album ●

Trump: Giving Canada independence mistake ●

Trump U offers Palin honorary climate degree ●

Mexico will close border if Trump elected ●

Trump: I will overtern shocking gay marriage ●

Trump picks Stacey Dash as VP ●

Pence: Michelle Obama most vulgar FLOTUS ●

Palin endorses Cruz ●

Sauron endorses Trump ●

Trump sues Chicago after forced to cancel rally ●

−0.075 −0.05 −0.025 0 0.025

32 / 32