gary king, jennifer pan, molly roberts · for 2 unusual topics: constant censorship e ort 0 5 10 15...

Reverse Engineering Chinese Censorship

Gary King, Jennifer Pan, Molly Roberts

Institute for Quantitative Social ScienceHarvard University

(talk at MIT Operations Research Center, 11/21/2013)

1/29

Papers

An Observational Study:

How Censorship in China Allows Government Criticism butSilences Collective Expression (American Political Science Review, May 2013)

Experimental and Participatory Studies:

A Randomized Experimental Study of Censorship in China

Copies at GaryKing.org

2/29

Chinese Censorship

The largest selective suppression of human expression in history

:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state

2 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state

2 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state

2 Stop collective action

Right

Either or both could be right or wrong.

(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action

Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Censorship

The largest selective suppression of human expression in history:

implemented manually,

by ≈ 200, 000 workers,

located in government and inside social media firms

Theories of the Goal of Censorship

1 Stop criticism of the state Wrong

2 Stop collective action Right

Either or both could be right or wrong.

(They also censor 2 other smaller categories we’ll talk about)

Benefit

?

Huge

Cost

Huge

None

3/29

Chinese Social Media: Fractured over 1,400+ sites

4/29

Chinese Social Media: Fractured over 1,400+ sites

hi.baidu 12%

bbs.voc.com.cn 5%

bbs.m4.cn 4%tianya 3%bbs.beijingww.com 2%

Share of Posts in Our Analysis by Internet Content Provider (ICP)

(Plus 59% from Sina Blog)5/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears

(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censored

Use computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Observational Study

Collect 3,674,698 social media posts in 85 topic areas over 6 months

Random sample: 127,283

(Repeat design; Total analyzed: 11,382,221)

For each post (on a timeline in one of 85 content areas):

Download content the instant it appears(Carefully) revisit each later to determine if it was censoredUse computer-assisted methods of text analysis (some existing, somenew, all adapted to Chinese)

6/29

Censorship is not Ambiguous: Example Error Page

7/29

Censorship is not Ambiguous: BBS Error Page

8/29

The Censors are Fast; Our Automated Methods are Faster

Days Until Censorship, Shanghai Subway Analysis

Days After Post Was Written

Num

ber C

enso

red

0 1 2 3 4 5 6 7

050

100

150

200

250

9/29

The Censors are Fast; Our Automated Methods are FasterExample: Shanghai Subway Crash

Days Until Censorship, Shanghai Subway Analysis

Days After Post Was Written

Num

ber C

enso

red

0 1 2 3 4 5 6 7

050

100

150

200

250

9/29

The Censors are Fast; Our Automated Methods are FasterExample: Shanghai Subway Crash

Days Until Censorship, Shanghai Subway Analysis

Days After Post Was Written

Num

ber C

enso

red

0 1 2 3 4 5 6 7

050

100

150

200

250

9/29

For 2 Unusual Topics: Constant Censorship Effort

05

1015

Count

Jan Mar Apr May Jun

Count PublishedCount Censored

Count PublishedCount Censored

010

2030

4050

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored

Students Throw Shoes at

Fang BinXing

Pornography Criticism of the Censors

10/29

For 2 Unusual Topics: Constant Censorship Effort0

510

15

Count

Jan Mar Apr May Jun

Count PublishedCount Censored

Count PublishedCount Censored

010

2030

4050

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored

Students Throw Shoes at

Fang BinXing

Pornography Criticism of the Censors

10/29

All other topics: Censorship & Post Volume are “Bursty”

010

2030

4050

6070

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst

(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

All other topics: Censorship & Post Volume are “Bursty”0

1020

3040

5060

70

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored Riots in

Zengcheng

Unit of analysis:

volume burst(≈ 3 SDs greaterthan baseline volume)

We monitored 85 topicareas (Jan–July 2011)

Found 87 volume burstsin total

Identified real worldevents associated witheach burst

Our hypothesis: The government censors all posts in volume burstsassociated with events with collective action potential (regardless of howcritical or supportive of the state)

11/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship

-0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 1: Post Volume

Begin with our 87 volume bursts in 85 topics areas

For each burst, calculate change in % censorship inside to outsideeach volume burst within topic areas – censorship magnitude

If goal of censorship is to stop collective action, we expect:

1 On average, % censoredshould increase duringvolume bursts

2 Some bursts (associatedwith politically relevantevents) should havemuch higher censorship -0.2 0.0 0.2 0.4 0.6 0.8

01

23

45

Censorship Magnitude

Density

12/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internet

individuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;

topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

Observational Test 2: The Events Generating the Bursts

Event classification (each category can be +, −, or neutral commentsabout the state)

1 Collective Action Potential

protest or organized crowd formation outside the Internetindividuals who have organized or incited collective action on theground in the past;topics related to nationalism or nationalist sentiment that have incitedprotest or collective action in the past.

2 Criticism of censors

3 Pornography

4 (Other) News

5 Government Policies

13/29

What Types of Events Are Censored?

Censorship Magnitude

Density

02

46

810

12

-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Collective ActionCriticism of CensorsPornography

PolicyNews

14/29

What Types of Events Are Censored?

Censorship Magnitude

Popular Book Published in Audio FormatDisney Announced Theme ParkEPA Issues New Rules on LeadChinese Solar Company Announces EarningsChina Puts Nuclear Program on HoldGov't Increases Power PricesJon Hunstman Steps Down as Ambassador to ChinaNews About Iran Nuclear ProgramIndoor Smoking Ban Takes EffectPopular Video Game ReleasedEducation Reform for Migrant ChildrenFood Prices RiseU.S. Military Intervention in Libya

Censorship Magnitude

New Laws on Fifty Cent PartyRush to Buy Salt After Earthquake

Students Throw Shoes at Fang BinXingFuzhou Bombing

Localized Advocacy for Environment LotteryGoogle is Hacked

Collective Anger At Lead Poisoning in JiangsuAi Weiwei Arrested

Pornography Mentioning Popular BookZengcheng Protests

Baidu Copyright LawsuitPornography Disguised as News

Protests in Inner Mongolia

PoliciesNews

Collective ActionCriticism of CensorsPornography

-0.2 0 0.1 0.3 0.5 0.7

14/29

Censoring Collective Action: Ai Weiwei’s Arrest

010

2030

40

Count

Jan Mar Apr May Jun Jul

Count PublishedCount Censored

Ai Weiwei arrested

15/29

Censoring Collective Action: Protests in Inner Mongolia

05

1015

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored

Protests in Inner Mongolia

16/29

Low Censorship on One Child Policy

010

2030

40

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored

Speculation of

Policy Reversal at NPC

17/29

Low Censorship on News: Power Prices

010

2030

4050

6070

Count

Jan Feb Mar Apr May Jun Jul

Count PublishedCount Censored

Power shortages Gov't raises power prices

to curb demand

18/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation set

ReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each category

Quantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

Observational Test 3: Content of Posts

We adapted “ReadMe” (Hopkins-King) methodology to the content ofcensored posts in Chinese

Standardize encoding of Chinese text

Eliminate punctuation and stop words

Experiment with segmenting characters into words

Calculate the quantity of interest (% censored by category)

Code a validation setReadMe estimates: % of posts in each categoryQuantity of interest: % of posts censored in a category

P(Censored |Category) =P(Category |Censored)P(Censored)

P(Category)

19/29

“ReadMe” Algorithm Validated in Chinese

20/29

“ReadMe” Algorithm Validated in ChineseExample: Labor Strikes, 2010 (Training set: 100; Test set: 900)

20/29

“ReadMe” Algorithm Validated in ChineseExample: Labor Strikes, 2010 (Training set: 100; Test set: 900)

0.0

0.2

0.4

0.6

ReadMe Results for Chinese Sampled, Not Segmented

Probability

Facts Supporting Employers

Facts Supporting Workers

Opinions Supporting Workers

Opinions Supporting Employers or Irrelevant

20/29

“ReadMe” Algorithm Validated in ChineseExample: Labor Strikes, 2010 (Training set: 100; Test set: 900)

0.0

0.2

0.4

0.6

ReadMe Results for Chinese Sampled, Not Segmented

Probability

Facts Supporting Employers

Facts Supporting Workers

Opinions Supporting Workers

Opinions Supporting Employers or Irrelevant

ReadMeTrue

20/29

Uncensored: Non-Collective Action Posts (that Support orOppose the State)

Per

cent

Cen

sore

d

0.0

0.2

0.4

0.6

0.8

1.0

Criticize Support Criticize Support Criticize Support

One Child Policy Corruption Policy Food Prices Rise

21/29

Uncensored: Non-Collective Action Posts (that Support orOppose the State)

Per

cent

Cen

sore

d

0.0

0.2

0.4

0.6

0.8

1.0

Criticize Support Criticize Support Criticize Support

One Child Policy Corruption Policy Food Prices Rise

21/29

Censored: Collective Action Posts (that Support orCriticize the State)

Per

cent

Cen

sore

d

0.0

0.2

0.4

0.6

0.8

1.0

Criticize Support Criticize Support Criticize Support

Ai Weiwei Inner MongoliaFuzhou Bombing

22/29

Censored: Collective Action Posts (that Support orCriticize the State)

Per

cent

Cen

sore

d

0.0

0.2

0.4

0.6

0.8

1.0

Criticize Support Criticize Support Criticize Support

Ai Weiwei Inner MongoliaFuzhou Bombing

22/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)

Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)

Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)

Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to type

Checked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)

Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentives

Procedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in China

Bought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselves

To learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Additional Research Designs

1 Randomized Experiment (for causal inferences)

Selected 100 top social media sites (∼87% of blogs, >500M Users,geographically diverse)Created 2 accounts on each (from inside China all over the country)Wrote 1,200 unique social media posts (CA/not CA, Pro/Anti)Submitted posts randomly assigned to typeChecked on censorship (from computers all over the world)

2 Participatory Study (for descriptive inferences)

Current method of learning how they censor: ask (carefully!)Our goal: change our sources’ incentivesProcedure: create our own social media website in ChinaBought URL; contracted with firms for servers & software; posted andcensored ourselvesTo learn: we tried every software option, read the documentation, andcalled customer support(!)

23/29

Mechanisms of Censorship

24/29

Mechanisms of Censorship

24/29

Mechanisms of Censorship

24/29

Mechanisms of Censorship

24/29

Posts For v. Against Government: Zero Causal Effect

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

PanxuProtest

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

●

YellowLightFines

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

●

YellowLightFines

●

StockMarketCrash

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

●

YellowLightFines

●

StockMarketCrash

●

Investigationof Sichuan

ViceGovernor

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

●

YellowLightFines

●

StockMarketCrash

●

Investigationof Sichuan

ViceGovernor

●

GenderImbalance

25/29

Posts For v. Against Government: Zero Causal Effect

● ●

−0.

50.

00.

51.

0

Cen

sors

hip

Diff

eren

ce (

Pro

− A

nti)

●

TibetanSelf−

Immolations

●

PanxuProtest

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

●

CorruptionPolicy

●

EliminateGoldenWeek

●

RentalTax

●

YellowLightFines

●

StockMarketCrash

●

Investigationof Sichuan

ViceGovernor

●

GenderImbalance

●

Li TianyiScandal

25/29

Collective Action Events: Large Causal Effect

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

PanxuProtest

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

●

PanxuProtest

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

●

PanxuProtest

●

TibetanSelf−

Immolations

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

26/29

Collective Action Events: Large Causal Effect

● ●0.0

0.1

0.2

0.3

0.4

0.5

Cen

sors

hip

Diff

eren

ce (

CA

Eve

nt −

Non

−C

A E

vent

)

●

PanxuProtest

●

TibetanSelf−

Immolations

●

Ai WeiweiAlbum

●

Protestsin

Xinjiang

26/29

Censorship to Preempt Collective Action: Ai Weiwei’sArrest

Mar 19 Mar 29 Apr 08 Apr 18

0.0

0.2

0.4

0.6

0.8

1.0

2011

% o

f Pos

ts C

enso

red

Actual %censorship

Predicted %censor trend basedon 3/19−3/29 data

Mar. 29,5 days prior

Apr. 3,Ai Weiwei Arrested

Placebo Test: Mostextreme of all effects

27/29

Censorship to Preempt Collective Action: Ai Weiwei’sArrest

Mar 19 Mar 29 Apr 08 Apr 18

0.0

0.2

0.4

0.6

0.8

1.0

2011

% o

f Pos

ts C

enso

red

Actual %censorship

Predicted %censor trend basedon 3/19−3/29 data

Mar. 29,5 days prior

Apr. 3,Ai Weiwei Arrested

Placebo Test: Mostextreme of all effects

27/29

Censorship to Preempt Collective Action: Ai Weiwei’sArrest

Mar 19 Mar 29 Apr 08 Apr 18

0.0

0.2

0.4

0.6

0.8

1.0

2011

% o

f Pos

ts C

enso

red

Actual %censorship

Predicted %censor trend basedon 3/19−3/29 data

Mar. 29,5 days prior

Apr. 3,Ai Weiwei Arrested

Placebo Test:

Mostextreme of all effects

27/29

Censorship to Preempt Collective Action: Ai Weiwei’sArrest

Mar 19 Mar 29 Apr 08 Apr 18

0.0

0.2

0.4

0.6

0.8

1.0

2011

% o

f Pos

ts C

enso

red

Actual %censorship

Predicted %censor trend basedon 3/19−3/29 data

Mar. 29,5 days prior

Apr. 3,Ai Weiwei Arrested

Placebo Test: Mostextreme of all effects

27/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticism

Reveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, Participatory

Automated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)

Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activities

Reveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silent

Possibly applicable to other countries

28/29

Concluding Remarks

Chinese people’s speech: Individually Free, Collectively in Chains

The Chinese Censorship Program:

Stops collective action, not criticismReveals government intentions; predicts actions

Our methodology:

Observational, Experimental, ParticipatoryAutomated text analysis methods validated in Chinese (withouttranslation)Enables continuous time measurement of Chinese governmentcensorship activitiesReveals information about state action when traditional media is silentPossibly applicable to other countries

28/29

For more information

GaryKing.org

j.mp/JenPan

j.mp/MollyRoberts

29/29

Appendix

30/29

Predicting the South China Sea Peace Agreement

Jun 12 Jun 22 Jul 02

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2011

% o

f Pos

ts C

enso

red

Predicted % censortrend basedon 6/10−6/20 data

Actual %censorship

Jun. 20,5 days prior

Jun. 25,PeaceAgreement

Placebo Test: Mostextreme of all effects

31/29

Predicting the South China Sea Peace Agreement

Jun 12 Jun 22 Jul 02

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2011

% o

f Pos

ts C

enso

red

Predicted % censortrend basedon 6/10−6/20 data

Actual %censorship

Jun. 20,5 days prior

Jun. 25,PeaceAgreement

Placebo Test: Mostextreme of all effects

31/29

Predicting the South China Sea Peace Agreement

Jun 12 Jun 22 Jul 02

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2011

% o

f Pos

ts C

enso

red

Predicted % censortrend basedon 6/10−6/20 data

Actual %censorship

Jun. 20,5 days prior

Jun. 25,PeaceAgreement

Placebo Test:

Mostextreme of all effects

31/29

Predicting the South China Sea Peace Agreement

Jun 12 Jun 22 Jul 02

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2011

% o

f Pos

ts C

enso

red

Predicted % censortrend basedon 6/10−6/20 data

Actual %censorship

Jun. 20,5 days prior

Jun. 25,PeaceAgreement

Placebo Test: Mostextreme of all effects

31/29

Censorship Pre-empting Collective Action: Wang Lijun’sDemotion

Jan 23 Jan 30 Feb 06 Feb 13

−0.

20.

00.

20.

40.

60.

81.

0

2012

% o

f Pos

ts C

enso

red

Jan. 28,5 days prior

Feb. 2,Wang Lijundemoted

Actual %censorship

Predicted %censor trend basedon 1/18−1/28 data

Placebo Test: Mostextreme of all effects

32/29

Censorship Pre-empting Collective Action: Wang Lijun’sDemotion

Jan 23 Jan 30 Feb 06 Feb 13

−0.

20.

00.

20.

40.

60.

81.

0

2012

% o

f Pos

ts C

enso

red

Jan. 28,5 days prior

Feb. 2,Wang Lijundemoted

Actual %censorship

Predicted %censor trend basedon 1/18−1/28 data

Placebo Test: Mostextreme of all effects

32/29

Censorship Pre-empting Collective Action: Wang Lijun’sDemotion

Jan 23 Jan 30 Feb 06 Feb 13

−0.

20.

00.

20.

40.

60.

81.

0

2012

% o

f Pos

ts C

enso

red

Jan. 28,5 days prior

Feb. 2,Wang Lijundemoted

Actual %censorship

Predicted %censor trend basedon 1/18−1/28 data

Placebo Test:

Mostextreme of all effects

32/29

Censorship Pre-empting Collective Action: Wang Lijun’sDemotion

Jan 23 Jan 30 Feb 06 Feb 13

−0.

20.

00.

20.

40.

60.

81.

0

2012

% o

f Pos

ts C

enso

red

Jan. 28,5 days prior

Feb. 2,Wang Lijundemoted

Actual %censorship

Predicted %censor trend basedon 1/18−1/28 data

Placebo Test: Mostextreme of all effects

32/29

Uncensored Posts (w/o Collective Action Potential)Critical of the State

33/29

Uncensored Posts (w/o Collective Action Potential)Critical of the State

33/29

Uncensored Posts (w/o Collective Action Potential)Critical of the State

This is a city government [Yulin City,Shaanxi] that treats life with contempt, thisis government officials run amuck, a citygovernment without justice, a city govern-ment that delights in that which is vul-gar, a place where officials all have mis-tresses, a city government that is shamelesswith greed, a government that trades dig-nity for power, a government without hu-manity, a government that has no limits onimmorality, a government that goes back onits word, a government that treats kindnesswith ingratitude, a government that caresnothing for posterity. . .

33/29

Censored Post (with Collective Action Potential)Supporting the State

34/29

Censored Post (with Collective Action Potential)Supporting the State

34/29

Censored Post (with Collective Action Potential)Supporting the State

The bombing led not only to the tragedy ofhis death but the death of many governmentworkers. Even if we can verify what QianMingqi said on Weibo that the building de-molition caused a great deal of personal dam-age, we should still condemn his extreme actof retribution. . . . The government has con-tinually put forth measures and laws to pro-tect the interests of citizens in building de-molition. And the media has called attentionto the plight of those experiencing housingdemolition. The rate at which compensationfor housing demolition has increased exceedsinflation. In many places, this compensationcan change the fate of an entire family.

34/29