UC San Diego Electronic Theses and Dissertations

UC San Diego UC San Diego Electronic Theses and Dissertations

Title Essays on Political Economy of the Media

Permalink https://escholarship.org/uc/item/76c987rx

Author Lam, Onyi

Publication Date 2017

Peer reviewed|Thesis/dissertation

eScholarship.org Powered by the California Digital Library University of California UNIVERSITY OF CALIFORNIA, SAN DIEGO

Essays on Political Economy of the Media

A dissertation submitted in partial satisfaction of the requirements for the degree Doctor of Philosophy

Economics

Onyi Lam

Committee in charge:

Professor Roger Gordon, Chair Professor Gordon Dahl Professor James Rauch Professor Molly Roberts Professor Kenneth Wilbur

Chair

University of California, San Diego 2017

iii DEDICATION

To my parents, Heung Wah Lam and Sau Man Ho, and my city, Hong Kong

iv TABLE OF CONTENTS

Signature Page ...... iii

Dedication ...... iv

Table of Contents ...... v

List of Figures ...... viii

List of Tables ...... xi

Acknowledgements ...... xiii

Vita ...... xiv

Abstract of Dissertation ...... xv

Chapter 1 Advertisers Capture: Evidence from Hong Kong ...... 1 1.1 Introduction ...... 1 1.2 Suggestive Evidence ...... 5 1.3 Institutional Context ...... 6 1.3.1 Business and Politics ...... 7 1.3.2 Participation in Mainland Politics ...... 8 1.3.3 Newspaper Market ...... 8 1.4 Illustrative Model ...... 9 1.5 Data ...... 15 1.5.1 Advertising ...... 15 1.5.2 Politically Connected Advertisers ...... 16 1.5.3 Readers’ Demography ...... 18 1.6 Measuring Slant Gap ...... 19 1.6.1 Selecting Phrases ...... 20

v 1.6.2 Mapping Phrases to Slant ...... 21 1.7 Empirical Evidence ...... 23 1.7.1 Persistence of Slant Gap’s Effect ...... 28 1.7.2 Real Estate Industry ...... 29 1.7.3 Alternative Measure of Political Climate ...... 31 1.8 Mechanisms ...... 32 1.9 Apple Daily’s Revenue Loss ...... 33 1.10 Conclusions ...... 35 1.11 Figures and Tables ...... 40 1.12 Appendix ...... 60 1.12.1 Sample of Apple Daily and Oriental Daily Headline during Occupy Central ...... 60 1.12.2 Endogenous Readers’ Choice of Newspaper ...... 61 1.12.3 Finer Time Unit for Slant Gap ...... 64 1.12.4 Polarization of the Benchmark Language ...... 65 1.12.5 Time Series of the Number of Pro-Democracy Phrases ...... 68 1.12.6 Political Ads and Government Ads ...... 69

Chapter 2 Measuring Subjectivity in History Textbooks ...... 70 2.1 Introduction ...... 70 2.2 History Textbooks in Mainland China, Hong Kong and Taiwan ...... 74 2.3 Descriptive Statistics ...... 76 2.3.1 Major Historical Episodes ...... 77 2.4 Subjectivity Analysis ...... 78 2.4.1 Adjectives and Adverbs ...... 79 2.4.2 Political Entities ...... 80 2.4.3 Text Polarity ...... 81 2.4.4 Word Embedding ...... 82 2.5 Prior and After Textbooks Reform ...... 84 2.6 Writing Style ...... 86 2.7 Conclusion ...... 87

vi 2.8 Figures and Tables ...... 92

Chapter 3 Celebrities Capture: Evidence from Weibo in China ...... 106 3.1 Introduction ...... 106 3.2 Media Market in Mainland China ...... 109 3.3 Model ...... 110 3.4 Data ...... 114 3.4.1 Summary Statistics ...... 115 3.5 Empirical Analysis ...... 117 3.5.1 Athletes ...... 119 3.5.2 Followers’ Reactions ...... 120 3.5.3 Facebook Accounts ...... 121 3.5.4 Banned Entertainers ...... 122 3.6 Conclusion ...... 122

vii LIST OF FIGURES

Figure 1.1 Ad Share on print version of Apple Daily and Sing Tao relative to Sum of Apple, Oriental and Sing Tao...... 40 Figure 1.2 Ad Share on print version of Apple Daily and Sing Tao relative to Oriental Only...... 41 Figure 1.3 2016 Press Freedom Index issued by the Freedom House ...... 42 Figure 1.4 Hang Seng Index and relative Stock Price of the corporations that own Apple Daily (Next Digital, Ticker: 0282) and Oriental Daily (Oriental Press, Ticker: 0018) ...... 44 Figure 1.5 Total number of ads assignments and ads assignments placed by connected organizations in each industry...... 45 Figure 1.6 Ads Assignment by company characteristics ...... 45 Figure 1.7 Percentage of Connected Ad Assignment ...... 46 Figure 1.8 Percentage of Mainland Ad Assignment ...... 46 Figure 1.9 Difference in slant between the two newspapers...... 48 Figure 1.10 Industry Fixed Effects Relative to the Real Estate Industry . . . . . 55 Figure 1.11 Industry Fixed Effects Relative to the Real Estate Industry . . . . . 55 Figure 1.12 Interaction terms between Industry and Slant Gap relative to the Real Estate Industry ...... 56 Figure 1.13 Interaction terms between Industry and Slant Gap relative to the Real Estate Industry ...... 56 Figure 1.14 Interaction terms between industry fixed effects and firms’ connectivity for ads appearing on Apple only ...... 57 Figure 1.15 Interaction terms between industry fixed effects and firms’ connectivity for ads appearing on both newspapers ...... 58 Figure 1.16 Headline of Apple Daily on December 12, 2014. Translated as ”Do not forget the original intention. We will be back” ...... 60

viii Figure 1.17 Headline of Oriental Daily on December 12, 2014. Translated as ”Financier Jimmy Lai Pan-Democratic politicians accept bribes. Occupy Central Schemers All Caught” ...... 60 Figure 1.18 Slant Gap Using Monthly Data ...... 64 Figure 1.19 A measure of polarization of word choices used by the benchmark . 66

Figure 2.1 Screenshot of the dictionary application on Mac showing the word Important ...... 92 Figure 2.2 Screenshot of the dictionary application on Mac showing the word Very 92 Figure 2.3 Pro-CCP to pro-KMT entities ...... 96 Figure 2.4 Adjective Ratio ...... 96 Figure 2.5 Adverb Ratio ...... 97 Figure 2.6 Adjective Ratio of pre-1949 events ...... 97 Figure 2.7 Adjective Ratio of post-1949 events ...... 98 Figure 2.8 Adjective Ratio pre- and post- 1949 ...... 98 Figure 2.9 Adjective Ratio pre- and post- 1949 ...... 99 Figure 2.10 Positive to Negative Phrase Ratio pre- and post- 1949 ...... 99 Figure 2.11 Positive to Negative Phrase Ratio pre- and post- 1949 ...... 100 Figure 2.12 Positive to Negative Phrase Ratio in 4 major pre-1949 Historical Episodes ...... 100 Figure 2.13 Positive to Negative Phrase Ratio pre- and post- 1949 ...... 101 Figure 2.14 Adjective Ratio in Old and New Versions ...... 101 Figure 2.15 Adjective Ratio in Old and New Versions in pre- and post-1949 . . 102 Figure 2.16 Adjective Ratio in Old and New Versions in pre- and post-1949 . . 102 Figure 2.17 Positive to Negative Phrase Ratio pre- and post- 1949 ...... 103 Figure 2.18 Positive to Negative Phrase Ratio pre- and post- 1949 ...... 103 Figure 2.19 word embedding visualization of Mainland Chinese textbooks by t-SNE104 Figure 2.20 word embedding visualization of Hong Kong textbooks by t-SNE . . 104 Figure 2.21 word embedding visualization of Taiwanese textbooks by t-SNE . . . 105

Figure 3.1 A popular post after the international tribunal in The Hague delivered its judgment ...... 128

ix Figure 3.2 Number of Celebrities by Followers Group ...... 129 Figure 3.3 Number of celebrities Mentioned in Sina by Type ...... 129 Figure 3.4 Number of Celebrities (Performers only) Mentioned in Sina by Region 130 Figure 3.5 Number of celebrities Using Patriotic Phrases ...... 131 Figure 3.6 Number of celebrities Posting by Date Around 2016 July 13th . . . . 132 Figure 3.7 Number of celebrities Posting by Date Around 2016 October 1st . . 133 Figure 3.8 Number of times the celebrity is mentioned on Sina ...... 133

x LIST OF TABLES

Table 1.1 Readership (’000) in 2014 (Including online readership) ...... 43 Table 1.2 Apple Daily’s Revenue Composition ...... 44 Table 1.3 Ad Price and Readership ...... 47 Table 1.4 Readers Demography. The number indicates the fraction of readers reading the respective newspaper, except average Age, household income and totals. Source: AC Nielsen 2014...... 47 Table 1.5 Phrases with highest χ2 and used by Wenwui Daily in each year and quarter ...... 49 Table 1.6 Phrase with highest χ2 and used by pro-Democracy politicians in each year and quarter ...... 50 Table 1.7 Multinomial Logit Regression Results ...... 51 Table 1.8 Results on Slant Gap and Interactions ...... 52 Table 1.9 Results on Slant Gap Lags ...... 53 Table 1.10 List of Major Protests in Hong Kong between 2010 and 2014 . . . . . 53 Table 1.11 Results on Protesters Turnout and Interactions ...... 54 Table 1.12 Average Ad Size by Industry ...... 59 Table 1.13 Results on Polarization and Interactions ...... 67 Table 1.14 Number of Pro-Democracy Phrase over the sample period ...... 68 Table 1.15 Number of political and government ad assignments in Apple and Oriental ...... 69

Table 2.1 Summary Statistics of Each Textbook Version ...... 93 Table 2.2 Character Count and Ratio of Major Historical Episodes ...... 93 Table 2.3 Most popular adjectives in each textbook ...... 93 Table 2.4 Mainland closest adjectives ...... 94 Table 2.5 HK closest adjectives ...... 94 Table 2.6 TW closest adjectives ...... 94 Table 2.7 Number of Mentions of politically-important figures and entities . . 94

xi Table 2.8 Positive and Negative Words in Each Version ...... 95 Table 2.9 positive to negative ratio of top 20 closest adjectives ...... 95 Table 2.10 Adjective ratio in major newspapers of the 3 regions ...... 95

Table 3.1 Utility Table for Weibo Users ...... 113 Table 3.2 Outcome Table for Celebrities Under Scenario 1 ...... 113 Table 3.3 Outcome Table for Celebrities Under Scenario 2 ...... 114 Table 3.4 Comparison between Mainland Performers and Athletes. The number denotes the average of the group...... 127 Table 3.5 Comparison across athletes of different sports...... 127 Table 3.6 Table of Regression Results ...... 134 Table 3.7 Logistic Regression Results on Individual Nationalistic Phrase . . . . 135 Table 3.8 Different Responses on Patriotic Post compared to Other posts . . . 135 Table 3.9 Regression Results by Sports ...... 136 Table 3.10 Logistic Regression Results on Users Engagement ...... 136 Table 3.11 Pearson Correlation Between the Number of Followers, Sina Promo- tion, and Patriotic Messages ...... 136 Table 3.12 Pearson Correlation Between the Number of Followers and Total Weibo Posts ...... 137

xii ACKNOWLEDGEMENTS

I would like to express my deepest gratitude to my advisor, Roger Gordon, who has been nothing but supportive in allowing me to pursue my own intellectual interests. He has provided invaluable guidance that pushes me to think critically on different economic mechanisms and grow as a researcher. I would also like to thank Gordon Dahl, Ruixue Jia, Kwai Ng, James Rauch, Molly Roberts, Kenneth Wilbur, for their thoughtful feedback. They have immensely enhanced the quality of the work. Sailing on this intellectual journey has been lonely at times, but it has been made easier by the support of my parents Heung Wah Lam, and Sau Man Ho. Their love, patience and understanding have enabled me to muddle through this process. I am also thankful to my brother, Wing Lam. I am grateful to the friends I have met in San Diego including but not limited to Yaein Baek, Mitch Downey, Jonathan Lam, Alvin Lee, Gene Leung, Penny Liao, Cherrie Lui, Ben Miller, John Rehbeck, Diego Vera Cossio, Michael Sharifi, Shihan Xie, Victoria Xie, Irina Zhecheva. The countless intellectual and fun conversations and outings have given color to my day-to-day life in San Diego. Chapter 1, in full, is in preparation for submission. Lam, Onyi. ”Advertisers Capture: Evidence from Hong Kong”. The dissertation author was the primary investigator and author of this material. Chapter 2, in full, is currently being prepared for submission for publication of the material. Lam, Onyi; Lin, Eddie. ”Measuring Subjectivity of History Textbooks”. The dissertation author was the primary investigator and author of this material. Chapter 3, in full, is in preparation for submission. Lam, Onyi. ”Celebrities Capture: Evidence from Weibo”. The dissertation author was the primary investigator and author of this material.

xiii VITA

2007 Bachelor of Engineering, University of California, Berkeley, USA 2012 Master of Public Policy, University of Tokyo, Japan 2017 Doctor of Philosophy, University of California, San Diego, USA

xiv ABSTRACT OF THE DISSERTATION

Essays on Political Economy of the Media

Onyi Lam

Doctor of Philosophy in Economics

University of California, San Diego, 2017

Professor Roger Gordon, Chair

My research focuses on understanding the political economy of traditional and new media. I study these issues by exploiting natural experiments, employing data techniques borrowed from machine learning and using both observational data from traditional and new sources. Chapter 1 shows that censorship often works through incentives rather than coercion by illustrating that advertisers’ decisions regarding where to advertise is affected by the opposing stances of two major newspapers. There are two mechanisms explored in this chapter. First, politically-connected firms, as well as mainland Chinese firms, aremuch more likely to avoid advertising in the pro-Democracy newspaper. Second, advertisement shares at the pro-Democracy newspaper fall sharply in politically volatile periods. Chapter 2 considers the problem of measuring subjectivity history textbooks in

xv China, Hong Kong and Taiwan. Using tools developed in the field of machine learning, we find empirical evidence that history textbooks in mainland China exhibit stronger degree of subjectivity than history textbooks used in Hong Kong and Taiwan. Specifically, the paper measures subjectivity by calculating the adjective ratio, the ratio of positive to negative phrases and employs word embedding method that measures distance from phrases of interest such as the Chinese Communist Party to other adjectives. Chapter 3 shows that popular Chinese celebrities, especially those from the entertainment industry, are more likely to express patriotic sentiment on Weibo, a Twitter-like social network tool in China than less popular ones. The tendency to post patriotic messages is also highly correlated with the intensity of government promotion of the celebrity. To explain the empirical result, I propose a clientelistic framework to show that dependency on government media to secure economic interest aligns the celebrities interest with that of the state.

xvi Chapter 1

Advertisers Capture: Evidence from Hong Kong

1.1 Introduction

Freedom of press is a right protected by law in many countries. Nevertheless, non- coercive political pressure on the media is prevalent in the world. In the 2016 annual report on Freedom of Press issued by Freedom House1, democratic countries such as Malaysia and Mexico2 were classified as ”not free”, and many others such as India, Japan and South Korea have ”partly free” media. How does the government influence media reporting outside of formal legislation? This paper sheds light on this question by showing that censorship often works through incentives rather than outright coercion. The 2011 earthquake and nuclear disaster in Japan and the subsequent media coverage of the event is a telling example of the mechanism illustrated in the paper. The Tokyo Electric Power Company (TEPCO), which ran the Fukushima plant, spent around 24.4 billion yen ($238 million) on advertising in a year3. The company has also maintained strong ties with the government. Even before the incident, the Japanese government has been criticized for exerting discreet political pressure on the media4. These have been

1The methodology comprises of questions that could be divided into three broad categories: the legal environment, the political environment, and the economic environment. 2The constitutions of both countries establish freedom of expression. 3Freedom House 2014 annual report 4http://www.economist.com/news/asia/ 21651295-japans-media-are-quailing-under-government-pressure-speak-no-evil

1 2 suggested as a factor to media’s conservative coverage of TEPCO’s handling of the nuclear crisis: news report on largely followed official government statements (Imtihani and Yanai, 2013), with few journalists questioning TEPCO until two weeks later after the leak. This event highlights the political function of advertisers and shows the intrigue ties between advertiser, government and the media. This paper empirically examines the role of advertisers in the political economy of media, and establishes causality of newspapers’ political stance on advertising choice. By studying the print newspaper advertisement market in Hong Kong, it shows that advertisers engage in politically-induced advertising boycott on a media that adopts a political stance which is against the government policy. Model of media political economy have attempted to explain why profit-maximizing advertisers might have a preference for biased information by incorporating how reporting affects readers response to ads (Ellman and Germano, 2009), but has fallen short in addressing the impact of state-business relations on media bias. As in the motivating example in Japan, businesses in many countries enjoy close ties with top officers (Faccio, 2006), and this tie is likely to lead firms to act as an extension ofstate. The paper quantifies the size of advertising revenue impact on the pro-Democracy newspaper due to political influence5 by exploring two mechanisms, illustrated in a simple model, that explain firms’ decision to engage in advertising boycott. The first is identified by time series variation. I show that an increase in firms’ political salience can lead to more aversion from advertising on media that is against government policy. The second is identified by cross-sectional variation in firms’ characteristics. I show that firms thatare politically connected with the mainland Chinese government and firms with headquarter located in mainland China are more likely to engage in advertising boycott. The identification strategy of the time series analysis relies on a behavioral insight that consumers’ attention tend to be drawn to an attribute that stands out among products (Schkade and Kahneman 1998; Koszegi and Szedil, 2013; Bordalo et al, 2013, 2015). The attribute is newspaper slant in our context. We exploit unexpected outbreak of political events and the variation in the intensity of these events over the sample period to obtain exogenous variation in newspaper slant. The main intuition is that news reports almost

5Work on the effect of advertising revenue on media’s independence include Petrova (2011), Qinetal (2014), Reuter and Zitzewitz (2006) and Reuter (2009) 3 always demonstrate a stronger interpretative narrative that is consistent with the newspapers’ underlying political ideology when political events occur. In other words, ideological slant implied in the news stories appearing on newspapers representing opposite political stance should diverge when political events are the central point and relatively indistin- guishable on non-political news such as traffic accidents. This divergence creates greater awareness among advertisers on newspaper’s slant and more aversion from advertising in a pro-Democracy newspaper. To operationalize this intuition, I use Gentzkow and Shapiro’s (2010) method to measure polarization of the headline news reporting made by two major newspapers sitting on opposite ends of the political spectrum6. We find evidence that firms exhibit more aversion from advertising on the pro-Democracy newspaper in periods that register large slant gap. Advertising choices of politically connected, as well as mainland Chinese companies are another focus of this paper. Following the literature on political connection (Faccio 2006, Khwaja and Milan 2005), we consider a firm as connected if at least one of its board isa member of the political organs within the mainland government. These companies are usually highly profitable. Many of them are beneficiaries of the political and economic status-quo7. In our sample period, Mainland Chinese companies are increasingly visible in the local Hong Kong economy as economic integration between mainland and Hong Kong deepens. Together, advertising spending by politically connected firms and mainland Chinese firms represent a large and growing fraction of the overall advertising revenuein the market. We find that both connected and mainland Chinese companies are more likely to avoid advertising on the pro-Democracy newspaper relative to the neutral firms. The institutional framework of Hong Kong presents several attractive features for the study. First, press freedom is protected by law. This means that the government cannot directly interfere with newspaper reporting and different voices are allowed. If freedom of press is not a legal right, newspapers can be easily regulated by the government and all newspapers will share government’s political stance. Second, the political spectrum in Hong

6Eisensee and Strömberg (2007) adopted a similar identification strategy. They show that U.S. relief depends on whether the disaster occurs at the same time as other newsworthy events, such as the Olympic Games, which are unrelated to need for relief. 7Fernandez and Rodrik (1991) consider a model that explains aversion from political reform. In the model, firms expect uncertainty over individual outcome. While firms might believe that democracycan bring net gain to the economy, uncertainty over individual gain/loss can lead firms to prefer the status-quo. 4

Kong spans from pro-Democracy to pro-Beijing. This allows us to follow the methodology of Gentzkow-Shaprio (2010) to quantify polarization8. Third, several incidents that are highly political in nature - including the Umbrella Movement - erupted in our sample period. This provides sources of variation in newspaper slant. I focus on the print version of the 2 paid-for newspapers with the largest readership in Hong Kong: Apple Daily and Oriental Daily. As shown in table 1.1, readership of each of the top 2 newspapers is about 4-5 times higher than the third most popular newspaper - Ming Pao. Together, readers of the two newspapers occupy approximately 75% of the total readership (including online readership). Apple Daily is the widely-recognized pro- Democracy paper whereas Oriental is pro-Beijing. The owner of Apple Daily, Jimmy Lai, is known to be a fiercely anti-Communist tycoon and frequently attacked by pro-Beijing critics. In contrary, owner of the Oriental Daily, Ma Ching Kwan, was appointed to be the member of the 10th and 11th session of PCC, an important political organ of the central PRC government (more in section 3). Our thesis is most comparable to Di Tella and Franceschelli (2009), which documents government’s intention to capture the media9. They find that a one standard- deviation increase in government advertisement in Argentina is associated with a reduction in corruption coverage by almost half of a front page per month. While their focus is on government ads, we examine ad patterns by all advertisers. This paper is also closely related to DellaVigna et al (2015) in estimating the effect of changes in political environment on advertising behavior. This paper differs in an important dimension that we emphasize the effect of news content on advertisers’ choices rather than the media10 ownership . The content is arguably a more important channel because of its direct tie to the freedom of press. While existing literature has emphasized that a large advertising market can promote media independence (Petrova, 2011; Gehlbach and Sonin, 2011), this paper suggests that

8Alternatively, Qin et al (2014) used principal component analysis to quantify media bias. They consider 9 content categories: mention of political leaders, citation of Xinhua, controversial issues intensively covered by oppositional overseas Chinese media, corruption, disasters, accidents, sports, crimes, and entertainment. They collapse these 9 dimensions of content differentiation into a single dimension. Groseclose and Milyo (2005) count the times that a particular media outlet cites various think tanks and policy groups, and then compare this with the times that members of Congress cite the same groups 9McMillan and Zoido (2004) provided the most direct evidence of government capture. They show the monetary transfer from the Peruvian government to the media companies during Fujimoto’s presidency 10Work on the effects of news contents includes Stromberg (2004); Gentzkow and Shapiro, 2004; Gentzkow , 2006; DellaVigna and Kaplan, 2007; Gerber,Karlan, and Bergan, 2009; Knight and Chiang, 2011; Enikolopov et al, 2011; Snyder and Stromberg, 2010 5 this view warrants caution because the advertisers can be captured by government also. This research also contributes to the literature on the relevance of firms’ political connections (Faccio, 2006; Khwaja and Mian, 2005; Knight, 2007; Cingano and Pinotti, 2013; Coulomb and Sangnier, 2012; Luechinger and Moser, 2012). While much of the emphasis of this strand of literature emphasizes the benefit of political connections, this paper extends this literature by showing that politically connected firms are likely to inherit government’s political preference as well. Finally, the suggested mechanism is broadly related to the literature of salience in consumers’ choice (Schkade and Kahneman 1998; Koszegi and Szedil, 2013; Bordalo et al, 2013, 2015). A number of empirical studies have documented the importance of salience and limited attention in a variety of economic contexts, and attention is drawn to the attribute that stands out from that of the rest of the choices11. This paper provides new empirical evidence to this literature by considering a novel context, namely political salience increases with news polarization.

1.2 Suggestive Evidence

There has been speculation that Apple Daily suffers from advertising boycott. In fact, from an interview as early as 2007, Jimmy Lai said,

There’s a very well organised boycott here. We have almost no real estate advertisement because real estate companies are mostly big companies with business in China. We have 30-35% less ads than we should have. And now the boycott has become very permanent and very organised.

12. The quote motivates our analysis on how firms’ characteristics affect their advertising decision, with a specific emphasis on the real estate industry. Another way to assess whether Apple Daily has suffered financially relative to Oriental Daily is to study the stock price of the corporations. Figure 1.4 shows the stock prices of the two corporations that own Apple Daily (Next Digital) and Oriental Daily (Oriental Press), as well as the Heng Sang Index over the sample period. Stock price of Next Digital registers much larger volatility in this period. While both corporations under-

11See DellaVigna(2009) for a detailed review. 12https://www.theguardian.com/media/2007/may/14/mondaymediasection8 6 perform relative to the market, Next Digit’s decline was much more apparent after the 1st quarter of 2011. Next Digital continued to under-perform relative to Oriental in most of 2013 and 2014. We also present the revenue composition between subscription and advertising for Apple Daily in table 1.2, which is printed in Apple Daily’s annual investor report. The advertising revenue in 2014 reduced to approximately 1/2 of that in 2010 whereas subscription revenue in 2014 was 68% of that in 2010. Fraction of advertising revenue to total revenue fell significantly between 2013 and 2014 from 0.621 to 0.566. These aggregate time trends, though short, suggest that the reasons to fall in demand for Apple’s advertising space might not be fully accountable by economic factors.

1.3 Institutional Context

When the sovereignty of Hong Kong was returned to China in 1997, the rights of the Hong Kong residents -including freedom of speech, freedom of the press, and freedom of assembly- were to be carried over and protected by the Basic Law. By law, neither the local and PRC government can directly interfere with newspapers’ reporting. However, there were signs that suggest press freedom in Hong Kong had gone under increasing pressure. The most telling example is the physical attack on a journalist of the liberal media in 201413, which highlighted the safety concern for liberal journalist. Local journalists saw the attack as effort of the PRC government in seeking to rein in Hong Kong’s vibrant and still free press while the pro-Beijing camp dismissed any link between the attack and press freedom. Another example is the repeated effort by the exclusive free TV channel inHong Kong (TVB) to downgrade politically sensitive material in news air time14. According to the annual report by Reporters Without Borders, Hong Kong’s decline in the press freedom index 15 ranking is drastic despite no change in law; the city was ranked 18 in 2002, 58 in 2013, 61 in 2014 and 70 in 2015. In parallel to an increasing pressure on media, scholars have noted that politics had

13Kevin Lau, the editor of Ming Pao, was physically stabbed and suffered from serious injuries 14The news department of TVB downgraded coverage of the 20th anniversary of the 4 June Incident. 15The index ranks the performance of 180 countries according to a range of criteria that include media pluralism and independence, respect for the safety and freedom of journalists, and the legislative, institutional and infrastructural environment in which the media operate. 7 become gradually polarized over the same period (Cheng, 2014). One piece of evidence is the increasing occurrence of filibusters in the Legislative Council. Our sample period isalso marked by increasing occurrences of political events such as large-scale protests. Several important events follow. In the second half of 2012, the leading political controversy was the government-backed Moral and National Education (MNE). Opponents believed that the goal of MNE was to brainwash the students to love China and launched large-scale protests with more than 100,000 protesters. In the last quarter of 2013, HKTV, a TV company started by a self-made billionaire with no political connection, was denied a TV license while other two politically connected companies were granted. The decision spurred suspicion that HKTV was denied for political considerations and tens of thousands protested on the street. In the second quarter of 2014, Beijing released the ”White Papers”. It was seen as a reneger on Beijing’s pledges for a democratic, autonomous Hong Kong under Beijing’s rule in 1997. In the last quarter of 2014, Umbrella Movement took place and marked the new height of confrontation.

1.3.1 Business and Politics

The business sector has been the chief ally of the mainland government in fending off the demands for progressive democratization well before the handover in 1997. Dur- ing the Basic Law drafting period from 1985 to 1990, local powerful capitalists such as Li Ka-Shing, formed the biggest group among the 23 Hong Kong members in the Basic Law Drafting Committee. The resultant Basic Law guaranteed business representation in the Legislative Council by allocating half of the seats to ”Functional Constituency” - seats allocated to professional interest group with very narrow electoral base. The chief executive is also to be elected by an election committee, of which the composition is weighted in favor of the business committee (Ma, 2007). Similarly, Cheung and Wong (2004) showed that business and professionals made up a large percentage of the major advisory committees in Hong Kong16. Most of the business and professional elite in these advisory committees had their seats in the united front17 organizations appointed by Beijing. Government in-

16A review by the Legislative Council in 1997 identified more than 300 advisory committees, including statutory advisory boards and committees (65), non- statutory boards and committees (116), non-advisory statutory bodies (73), statutory charitable bodies (4), and statutory bodies dealing with appeals (45) 17The united front tactic is a PRC initiative in bringing different individuals and groups together to unite support for the authority. 8 formation showed that about half of the 800 members of the EC in 2000–05 were members or chairmen of advisory committees ((Ma, 2007). The heavy business representation in the government binds government’s interest with the business interest, and secures support of business to government policies.

1.3.2 Participation in Mainland Politics

Very few selected Hong Kong residents can participate in mainland politics. The two main government bodies that enable limited involvement are the National People Congress (”NPC”) and the Chinese People’s Political Consultative Conference (PCC, ). The NPC is the official legislative body of the PRC government consists of about 3000del- egates. Hong Kong NPC delegates are elected by a pre-selected group of members. There are 36 delegates from Hong Kong for each 5-year term. The PCC is a political advisory body in the PRC government. This organization consists of delegates from a range of political parties and organizations, as well as independent members. Unlike the NPC, the PCC does not have actual legislative power but it is an important political organization with the stated goal to ”brings together people and groups from different sectors in the society toa common struggle against anti-communism threat”. In essence, it aims to unite individuals with different objectives to secure the continuing ruling of the Chinese Communist Party. Hong Kong delegates of the PCC have varied backgrounds and come from different sectors in the society: high school principal, lawyers, architects, business owners...etc. To become a PCC members, one needs to be ”invited” or ”recommended” by ”related political groups”. Neither the selection process of NPC and PCC is open nor transparent to the public.

1.3.3 Newspaper Market

The print news market in Hong Kong is vibrant. Despite the ubiquity of the Internet and the growing popularity of online news sites, the traditional paid-for newspapers remained a major source of information for people in Hong Kong in the sample period. There are 7 major paid-for newspapers to serve the 7 million population in Hong Kong. Aside from Apple Daily and Oriental Daily, Economic Times, Economic Journal, Ming Pao, Sing Tao and the Sun 18 are newspapers that have a non-trivial fraction of readership

18the Sun has ceased publication in April, 2006. 9 but their market share is significantly less than that of Apple Daily and Oriental Daily. While Ming Pao moderately leans toward pro-Democracy, others are known to adopt a more pro-Beijing stance. Turning to the readers, unlike the US, subscription of print copies of the newspapers is not common. Most readers purchase a print copy from local news vendor on the street every day. Table 1.1 shows that a non-trivial fraction of readers read more than one newspapers. For example, about 19% of Apple’s readers also read Oriental but it is unclear how many of them purchase both. Based on a survey19, Apple Daily’s credibility as a newspaper was ranked 15 out of the 17 newspapers in 2010 and Oriental Daily ranked 9, which suggests that Apple’s popularity is not driven by an objective news stance. Most paid newspapers sell at a cover price of $7HKD (less than $1 USD) in 2014. The price was raised from $6HKD to $7HKD in 2013 by Apple, Oriental Daily and the Sun. The decision was soon followed by other newspapers. This suggests that price is not the distinguishing factor that newspapers use to compete for readership. According to Sing Tao Daily’s online rate card 20, the newspaper requires at least one week notice in order to cancel a previously ordered advertisement. This means that an advertiser cannot easily cancel a previously ad reservation in response to the news report published on the day before the advertisement. However, an advertiser can respond to the news report today when trying to decide where to put an ad tomorrow. This means that for ads that require longer planning time (for example, a large real estate project, or grocery ads that take more than a week to plan), it is difficult to for advertiser to respond immediately to news report. A key determinant of advertisers’ newspaper choice is ad price set by newspapers. Ad price reflects newspapers’ perceived effect of political climate on the advertising demand, among other economic factors. In Hong Kong, ad prices are usually set in the beginning of the year and are committed by the newspapers throughout the year.

1.4 Illustrative Model

This section considers a very simple supply and demand model of newspaper advertising to illustrate the effects of political salience and political connection on firms’ adloca-

19http://www.com.cuhk.edu.hk/ccpos/en/research/Credibility_Survey%20Results_2010_ENG.pdf 20http://std.stheadline.com/daily/upload/singtao.pdf 10 tion decision. The demand side depicts firms’ choice between the two newspapers: Apple (A) and Oriental(O). The supply side assumes Bertrand-Nash competition in ad prices between the two newspapers. Firms’ choice is affected by their salience to newspapers’ political stance and their relation to the PRC government. The economy is populated by a large number of firms normalized to 1. Firms, indexed by i, can be classified as either Beijing-friendly (B) or neutral(N). Let r denote the fraction of firms that are Beijing-friendly, which includes politically connected firms and mainland firms in the empirical section21. Each firm puts only one ad on the chosen newspaper. We assume that advertising is always more beneficial than not advertising.

Firm i receives a random economic benefit eij for reaching the readers of newspaper j. The economic benefit includes benefit from reaching the target audience as well as the thesizeof readership. We assume that economic benefit differs across firms because firms can target different population segment. Thus, the quality of the match between firms’ target audience and newspapers readers can vary across firms. Furthermore, daily readership varies and is unknown to both firms and the newspapers, and firms may only care about the expected readership of a certain day in the year. 22 Firms pay the fixed price pj for putting an advertisement on newspaper j . This price, as described below, is endogenously determined by market conditions and is committed by the newspapers in the beginning of the period. Beijing-friendly firms benefit from the political and economic status-quo and therefore favor Oriental; we model this preference for Oriental by considering a cost ν on Beijing-friendly firms when they advertise on Apple. This cost could be interpreted as a signal needed to demonstrate loyalty to Beijing (Spence, 1973). Signaling loyalty can be understood as business’s need to invest in Guanxi (good relation with government) to ensure economic benefit such as ”access to profit-enhancing information and asset” (Wank, 1996). Della Vigna et al. (2015) showed that advertising on Berlusconi’s TV station increased when Berlusconi came to power, and their interpretation was that firms received

21The fraction of connected firms is treated as exogenous in the model. In reality, government needsto decide whether the company is ”worthwhile” in connecting. There is a potential trade-off in exchanging economic benefit for the firm with loyalty for the government. A direct implication of endogenizing political connection is that government can easily influence firms’ behavior and more effectively apply pressure on the media in an economy where few large firms dominate. 22I assume that prices are the same for all firms and that there are no quantity discounts. 11 a political benefit to advertise on Berlusconi’s TV station. By similar logic, the Beijing- friendly firms could potentially enjoy a political benefit by advertising on Oriental and the interpretation of ν would be quite different. However, there are two reasons we prefer the ”political cost” interpretation in this setting. First, the context in Della Vigna et al (2015) has a single beneficiary: Berlusconi’s TV station. In our setting, there are several newspapers that adopt a pro-Beijing stance while Apple Daily is the only newspaper that adopts a strong pro-Democracy stance. As a result, political gains are diversified, while political cost is concentrated in one newspaper. Second, none of the pro-Beijing newspapers is actually owned by government officials in Beijing. In other words, advertising onthem do not buy advertisers political favors directly. As discussed in section 3, firms in Hong Kong - albeit the politically unconnected ones - are generally conservative in their political view. To reflect firms’ political conser- vatism, we assume that firms derive disutility from advertising on Apple Daily, and thelevel of discontinuity depends on the firms’ political salience and newspapers’ ”political image”. Political salience is determined by the news report slant in each newspaper. The ”political image” is the newspapers’ political stance that is rooted in the public’s perception. In other words, firms still have a preference for Oriental Daily even in periods with no political event because Apple is well known for its pro-Democracy stance. Denote the news report slant of newspaper j as θj and we assume that the larger the θ the more pro-Democracy the news report slants toward. We represent the ”political image” of Apple and Oriental by θA and

θO. Since Apple has a pro-Democracy image, θA > θO. Given all of these, the neutral and Beijing-friendly firms’ indifference condition between the two newspapers can be described as follows:

eiA − pA − θA − θA = eiO − pO − θO − θO (1.1)

eiA − pA − ν − θA − θA = eiO − pO − θO − θO (1.2)

The first equality implies that even when the news report slant is the same between two newspaper, Oriental will still be more attractive than Apple because of its pro-Beijing image. The second equality implies that it takes a larger economic benefit from advertising on Apple for Beijing-friendly firms to advertise on Apple Daily relative to the neutral firms. 12

In this setup, demand for Oriental will also suffer if their reporting becomes more pro-

Democracy. The equalities also imply that the larger the slant gap (△θ ≡ θA − θO), the less attractive is Apple to firms holding other parameters constant. This suggests that while political salience increases with news report’s slant, increase in salience alone does not affect newspapers choice if both newspapers slant in the same direction. Increasing contrast in news report slant, however, sway firms away from Apple Daily. That salience increases with contrasts in options has been studied in the literature (Bordalo et al, 2013, 2015; Koszegi and Szedil, 2013). A remark on the assumption of exogenous newspapers’ reporting follows. One might reasonably suspect that the slant of news report is endogenous to expected advertising demand since newspapers can choose to tune up or down the level of slant in their reporting. In particular, given the political environment at the time, a pro-Democracy newspaper might engage in self-censorship and report in milder languages, while the pro-Beijing paper would use more provocative pro-Beijing languages to win ads. While this is entirely plausible, our identification assumption does not require the newspapers to report inways as if politically-induced advertising pressure is absent. All that is needed for identification is that newspapers, despite their intent to slant one way or other, to report with more slant in politically turbulent periods that are consistent with their political image to the public. And since the occurrences of these politically polarizing events are exogenous, the slant gap over time is also exogenous. This is a reasonable assumption since newspapers, despite its political ideology, still has to convey informative news of the local happenings to its readers. This means that Oriental Daily cannot always squeeze pro-Beijing sentiment in its news report, and likewise for Apple Daily.

Denote △ei ≡ eiA−eiO . For tractability, we assume △e follows a uniform distribution ∗ with support [−δ, δ]. Given a threshold △e of which firms are indifferent between the ∗ newspapers, the firm i will place the ad on Apple Daily if eiA > △e . Below the threshold, the firm will place the an Oriental Daily. Hence the probabilities that the neutral and 13

Beijing-friendly firms will advertise on Oriental are given as

pA − pO + △θ + △θ + δ S ≡ (1.3) NO 2δ

pA − pO + ν + △θ + △θ + δ S ≡ (1.4) BO 2δ

The probabilities that the neutral and Beijing-friendly firms’ will advertise on Apple are then simply SNA ≡ (1 − SNO) and SBA ≡ (1 − SBO)) respectively. On the supply side, our view is that newspaper advertising space is flexible and can be easily adjusted; there is no capacity constraint such that newspapers can print as many pages as needed23, and that there is no ad quota to fill so newspapers do not have to fill pages with ads to reach a certain number of pages. I model newspapers price-setting asa simple Bertrand competition where ad prices are set simultaneously by the newspapers and newspapers take into account of the best response strategy of the other paper to maximize expected revenue. We assume that the marginal cost of printing an additional ad is constant and equals to 0. Furthermore, we assume that newspapers have a correct expectation of the political climate over the period. Given these assumptions, Apple and Oriental choose ad price to maximize the following profit function respectively:

πA ≡ pA [r(1 − SBO) + (1 − r)(1 − SNO)]

πO ≡ pO [r(SBO) + (1 − r)(SNO)]

The first order condition with respect to ad price yields the newspapers’ bestre- sponse pricing function. Substituting the pricing function of one newspaper into another yields the closed-form solution for prices. The equilibrium share can be obtained by evalu-

23In contrary to TV commercials studied in DellaVigna et al (2015), ads spaces on newspapers are much more flexible since editors can simply rearrange news stories to fit however many number of pagesneeded to fit all advertisements. Note that we have implicitly assumed that readers do not get disutility fromads. 14 ating the ad share equations at the respective equilibrium ad prices.

∗ rν + △θ + △θ p = δ + O 3 ∗ rν + △θ + △θ p = δ − A 3 ∗ 1 rν + △θ + △θ S = + O 2 6δ ∗ 1 rν + △θ + △θ S = − A 2 6δ

Ad prices are sticky in the short run, and therefore ad share is the main moving variable if other parameters in the model change. Holding ad price constant, the model predicts that an increase in political events occurrences leads to a fall in Apple’s ad share. In the medium run, newspapers adjust prices and the model predicts that an expected increase in political events leads to wider price gap. This prediction can be verified in the widening ad price gap in 2013 and 2014. In the empirical section, we will formally test the following predictions:

1 Apple’s ad share falls relative to Oriental Daily in politically sensitive periods.

2 Beijing-friendly firms are less likely to advertise on Apple Daily relative to Oriental Daily in all periods.

3 All firms are less likely to advertise on Apple in politically sensitive periods.

This framework can be readily expanded to incorporate endogenous readers’ newspaper choice. Readers can become more salient of the newspapers’ political stance in turbulent times also. If the underlying population purchases more copies of Apple Daily24 in turbulent times, the economic benefit of firms advertising on Apple Daily will rise.Our reduced-form estimates will then underestimate the political cost of political salience im- posed on all firms. It is also possible that there is increased sorting of readers in turbulent times. Firms may be able to better match readers with its consumers better as a result. We discuss this possible mechanism in section 7 and find no empirical evidence of sorting. We consider this possibility in the appendix.

24Readers tend to buy newspapers when big news happened (Gentzkow et al 2007)). Given that Apple is the only strong pro-Democracy paper, the economic benefit of Apple relative to Oriental should risein turbulent times. 15

1.5 Data

1.5.1 Advertising

I use a dataset that includes all the ads that have appeared on the newspapers between 2010 and 2014. The dataset is available from a Hong Kong-based private company, Wisers. We exclude tender, announcement, evens promotion and obituary. All ads that were posted or partly sponsored by the newspaper themselves are also discarded. The information for each ad includes the ad headline, date of the ad, the newspaper that the ad appeared, the industry(s) that the ad product(s) belongs to, the company(s) of the product(s), and the section of newspaper that the ad appears. By matching the company name, I merge the ads database with a company information dataset that I constructed. In the company dataset, I classify a company as ”Local”, ”Foreign” or ”Mainland” based on the company’s location of headquarter 25. I also classify an organization as ”listed on Hong Kong Stock Exchange”, ”for-profit”, ”not-for-profit” organizations or government agencies based on other publicly available information. Finally, I look up the mother company of the advertiser if (For example, the fast food chain KFC belongs to the YUM group) they are missing from the Wisers ad dataset 26. There are 17 possible industries assignment 27, which are provided by Wisers. Ads can have multiple company and industry assignments because more than one company could appear on an ad and each ad product can be classified into multiple industries. We identify an unique ad by the headline and the date of which the ad appeared. An ad with the same headline could appear on the newspapers on multiple days, but they are considered as different ads in our dataset. Since we do not observe the actual imagery, an ad with the same headline and appeared on two newspapers on the same date but with different size would still be considered as an unique ad appearing on two newspapers. We consider an unique ad assignment as a combination of an unique ad and the company associated with the ad. For example, in a cell phone ad for which both the cell phone carrier

25I manually look up the headquarter location from the companies’ website. Companies from Taiwan are classified as ”Foreign” whereas Companies from Macau are classified as ”Mainland” 26In most cases, Wisers lists the owner of the brand and I cross check to make sure the mother company is correct 27Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel). 16 company and the manufacturer company are present, the ad will then have two unique ad assignments. In summary, there are 116887 unique ads, 13868 unique firms, and 206425 unique ad assignments that appeared on either Apple, Oriental or both. On average, there are 1.77 companies, and 1.23 industries assignment per advertisement. 33209 ads had a HKSE(Hong Kong Stock Exchange)-Listed advertiser assignment. 41055 ads have at least one or more foreign advertisers, but only 4581 ads have at least one or more mainland advertisers. 970 ads are related to the government, and 4636 have at least one or more not-for-profit organizations. Over our sample period, there are 27302, 25926, 22451, 22548 and 18660 ads in each year between 2010-14. Finally, only 4897 unique ads and 9069 ad-company assignments appear on both Apple and Oriental.

1.5.2 Politically Connected Advertisers

We classify a firm listed on the stock exchange as ”connected” to Beijing ifat least one of the firm’s owner or major shareholders or board members appears on the list of connected individuals. The approach follows the like of Faccio (2006) and Accemoglu et al (2014). We can identify the board members through a list of all board of directors from the Hong Kong Exchange and Clearing Limited (http://www.hkexnews.hk/reports/ dirsearch/dirlist/directorlist_c.htm). The list of connected individuals consists of all Hong Kong representatives of the NPC and PCC from the 9th to 12th session. The tenure for each member in each session is 5 years, and the 9th congress was in session from 1998 to 2003. Hence, our members list only include members selected after the 1997 turnover. In total, we have 73 non-repeated NPC members. There are 264 non-repeating PCC members. Since there are overlapping members between NPC and PCC, the combined list results in 331 unique individual. The list excludes representatives of other provinces who are Hong Kong residents28. For advertisers that are not listed on the stock exchange or are not-for-profit organizations, I cannot formally identify their ownership information. Instead, I gathered the

28There is no rule that restricts the PCC representatives of a province to be residents of that province. For example, Lawrence Ma( ) is the representative of Shanxi province in PCC, and Karson Choi () is the representative of Guangxi province. Both are Hong Kong residents. Most of the representatives of Hong Kong are Hong Kong residents, however. 17 profile of the list of connected individuals from the official website of the respective organizations (NPC and PCC )29. The profiles list the occupation and other outside positions of each member. This information is not complete since many individuals have multiple affiliations but the online profiles only include one affiliation. I use this supplementary information to determine whether the respective company/organization is connected or not. Government organizations and educational institutes are discarded. I then check whether the businesses and organizations associated with the profiles appear in the ads dataset 30. In summary, there are 20370 unique ads that have at least one or more connected advertiser. Table 1.5 shows the total number of all and connected ad assignments by industry. The ”Banking” industry has the most ad assignments at over 35,000, followed by the ”Travel”, ”Retail” and ”Restaurants” industries. The percentage of connected ad assignments within each industry varies. The ”Telecommunications” industry has the highest percentage of connected ad assignment at about 32%, followed by the ”Banking”, ”Insur- ance” and ”Real Estate” industries. There are many business conglomerate in Hong Kong, and they have a strong presence in almost all of the industries. For example, New World Mobility (telecommunication) and New World Development (real estate) belong to the New World Group and therefore both are classified as politically connected. Similarly, the Che- ung Kong Group (owner: Li Ka-Shing) owns Hutchison (telecommunications) and Watsons (A large retail chain in Hong Kong). These companies are all major players in their respective industries. The presence of these business conglomerate means that despite the relatively few number of connected individuals, these connection can have a large effect to newspapers’ advertising revenue. Figure 1.6 shows the percentage of connected ads assignments by firm characteristics. Foreign companies are responsible for over 50,000 ad assignments (an ad can have more than one foreign company) as opposed to less than 10,000 ad assignments from mainland companies. Less than 10% of the foreign ad assignments are considered as connected.

29We retrieve the profile of all connected individuals using a python script to scrape their profile onthe official websites of the NPC and PCC. The websites are: http://www.npc.gov.cn/npc/gadbzl/xgdbzl_11/node_8514.htm http://www.cppcc.gov.cn/CMS/icms/project1/cppcc/wylibary/wjWeiYuanList.jsp

30Only the main occupation of the members are listed, and many individuals are board members for multiple companies. 18

In contrast, the percentage of ad assignment considered as ”connected” is over 60% among HKSE-listed ad assignments, which accounts for over 40,000 ad assignments. Figure 1.7 plots the percentage of connected ad assignment in each quarter of the sample period by newspaper. In all but one quarter, the percentage of connected to total number of ad assignment is higher at Oriental. The gap appears to be bigger in the first half of our sample period. Furthermore, there appears to be no trend inthe average percentage of connected ads over time. Figure 1.8 plots the percentage of mainland ad assignment by newspaper for each quarter in our sample period. The percentage of mainland ad assignments to total ad assignment at each newspaper is higher at Oriental in all quarters. The gap also appears to diverge in 2014.

1.5.3 Readers’ Demography

Firms might select a newspapers based on the newspapers reader demography. I use the 2014 AC Nielsen’s media index to compare the readers demographics between the two newspapers. AC Nielsen surveys over 6000 individuals of the Hong Kong general population aged 12-64. The samples are weighted to the known population of Hong Kong based on government statistics, making the sample representative to the 5,280,000 residents in Hong Kong. Table 1.4 shows the basic demographics of the newspapers readers. The average reader of Oriental Daily is slightly older than that of Apple Daily while the average household income is very similar between the two newspapers. Both newspapers have slightly more than 1/3 of their readers living in government housing31. Readers of Apple Daily is more educated with the percentage of readers who have completed university or above 5% higher than Oriental readers. While differences in readers’ demography between the newspaper exist, it would be difficult for firms do readership profiling to isolate acertain population segment by just choosing one newspaper over another. It is possible that the demographic composition changed over time. However, a limitation of the readership dataset is that the date of the interviews are not recorded and therefore, we cannot estimate if and how readership is affected during politically-polarized dates. In addition, we are unable to tell whether readers’ demographics change in the sample period since our demographic information is for 2014 only.

31Government housing residents are in general poorer compared to the average population. 19

1.6 Measuring Slant Gap

I follow the methodology of Gentzkow and Shapiro (2010) to quantify slant gap between Apple and Oriental in each period. The main idea is to compare the relative frequencies of the ”diagnostic” phrases used in the news article with the hard-line pro- Beijing and pro-Democracy rhetoric. These ”diagnostic” phrases are politically-charged languages that slant to one side or the other. For example, one would expect hard-line pro- Beijing rhetoric to use words such as ”stability” more heavily and hard-line pro-Democracy rhetoric to use words such as ”freedom” more often. I have access to the archives of major newspapers in Hong Kong through Wisers, a private company based in Hong Kong. I use all headline articles between 2010 and 2014 that appeared on Apple and Oriental. As another input to the slant measure, I use the official commentary of the Wen Wui daily as the reference hard-line Pro-Beijing rhetoric. Wen Wui Daily is a state-owned newspaper. It is controlled by the Liaison Office of the Central Government, and is widely regarded to follow the hard party line. The official commentary was published daily (sometimes multiple commentary are posted in a day) and is publicly available on their website. I wrote a Python script to scrape all commentaries from the newspaper’s official website. I use a column written by 7 famous pro-Democracy politicians/lawyers as the reference of the hard-line pro-Democracy rhetoric, which is also available from Wisers. The name of the column is ”Law and Politics” and appeared on Ming Pao daily. The contributors include Margaret Ng, Audrey Eu, Ronny Tong, Alan Leong, Gladys Li, Johannes Chan, Martin Lee, Albert Ho and Benny Tai. Most of the contributor belong to a pro-Democracy party, the Civic Party, in Hong Kong. This column appears in Ming Pao every day since the end of 2003. In principle, each politician takes turn to contribute to the column but some of the authors dropped out and others joined. Benny Tai - the initial advocate of ”Occupy Central” movement, which later became the umbrella movement, has contributed to the column since 2011. I excluded all articles written in English. The processing of Chinese language using text mining technique differs from that of English only slightly. The first step involves removing all numbers, punctuation, and English words from the article. Because the Chinese language does not have white spaces 20 like English, we have to apply an extra procedure to parse the sentences into meaningful phrases(We call bi-gram, tri-gram and quad-gram as phrases made up of 2, 3 and 4 Chinese characters respectively.). This is done by implementing an existing package (rwordseg) in R. We need to supplement the package with a list of words that are specific to Hong Kong because local dialogues or newly created phrases in the news report would not exist in the dictionaries provided in the R package. I went through two procedures to create a list of new words. The first procedure is manual. I eyeballed all the phrases decomposed bythe computer script and manually check whether some phrases should be joined together. For example, many names of the officials were not recognized by the script. so I manually added them to the dictionary. Second, I wrote a program to concatenate one-word phrase into two- or three- word phrases, and rank them by the frequency they appear. We then manually check whether these high-frequency terms represent actual phrases. The high-frequency terms would then be added to the dictionary.

1.6.1 Selecting Phrases

We focus our attention on the phrases that are most likely to contain ideological valence. To select these phrases, we first let fplw and fpll denote the total number of times phrase p of length l (two, three or four words) is used by Wenwui and Law&Politics(LP),

respectively. Let f−plw and f−pll denote the total occurrences of length-l phrases that are not phrase p used by Wenwui and LP, respectively. We compute the Pearson’s χ2 statistic for each phrase (of any length):

− 2 ( fplw f−pll fpll f−plw) χ2 = (1.5) ( fpll + fplw)( fpll + f−pll)( fplw + f−plw)( f−pll + f−plw)

Pearson’s χ2 statistic is a test statistic for the hypothesis that a phrase is equally likely to be used by the pro-Beijing and pro-Democracy reference. A high Pearson’s χ2 suggests that the phrase is unlikely to be used by both the pro-Beijing and pro-Democracy reference, and would imply high partisanship. For each period, I rank the phrases by their Pearson’s χ2 and identify 80 phrases (of any combinations of bi-gram, tri-gram and quad-grams32) with the highest χ2 . Table

32Gentzkow and Shapiro (2010) use equal number of bi-grams, trigrams and quad-grams. We deviate from their treatment because this gives us higher predictive power. Specifically, some quad-grams might be 21

1.5 shows the phrases that are more heavily used by Wenwui’s commentary and table 1.6 shows the phrases that are more heavily used by Pro-Democracy commentary in each quarter. The procedure identifies many phrases that confirms the intuition that many ofthe phrases are chosen strategically for their partisan impact. For example, the economy and foreign relations are emphasized more in the Wenwui commentaries. In 2011, phrases such as ”Development”, ”Economy”, ”Financial Crisis”, and ”United States” are among the most pro-Beijing phrases. ”Ai Weiwei” (A famous Chinese dissident), ”Election”, ”Democracy” and ”Freedom of Press” are among the most pro-Democracy phrases. The emphasis changes over time. In quarter 4 of 2014, Wenwui stressed that the ”Occupy Central” movement as unlawful by using phrases such as ”Law-Violating Central Occupation” and focusing more on the ”police”. In contrary, the pro-Democracy language includes phrases such as ”Civil disobedience” to justify the occupation and emphasize the government’s wrongdoing. In the appendix I show the distribution of the number of pro-Democracy over the sample period33.

1.6.2 Mapping Phrases to Slant

The list of 80 phrases gives us a basis to evaluate the partisanship of the news report in each period. In order to generate a slant measure, we compare the frequencies of these phrases in the reference languages with that in the news report. Specifically, for each quarter index the phrases by p ∈ {1...80} (Ignore phrase length, year for notational convenience.). ∑ ˜ ≡ 80 Let fpn denote the frequency of phrase p on newspaper n. Let fpw fpw/ p=1 fpw denote the ˜ relative frequency of phrase p in the Wenwui commentaries in each period. fpl, the relative frequency of phrase p, is defined similarly for phrases in the LawPolitics commentaries. We estimate slant in each period for each of the newspapers as follows:

i For each phrase p, we calculate the difference in relative frequencies between LawPol- ˜ ≡ ˜ − ˜ itics and Wenwui: ∆ fp flp fww.

˜ ii I regress the relative frequencies of the selected phrases in newspaper n on ∆ fp. The more informative than bi-grams in some periods or others. Using only the top 80 chi-square phrases help us capture that. 33There is consistently more pro-Democracy phrases than pro-Beijing phrases. The reason is that the length of the pro-Democracy reference is shorter than that of the pro-Beijing reference, which means that the chi-square of a particular pro-Being phrase will always be lower. 22

slope estimates, which is our slant index, gives: ∑ 80 ˜ p=1 ∆ fp fpn θn = ∑ (1.6) 80 ˜ 2 p=1 ∆ fp

This approach is a modified version of Gentzkow and Shapiro (2010)34. First, taking the LawPolitics and Wenwui commentaries as the benchmark language, the larger the difference of the relative frequencies ∆ f of a phrase, the more pro-Democracy it implies. Another way ˜ of seeing this is that from equation 1.6, ∆ fp does not contribute much to the slant index if it is small. Second, we infer the slant of the news report by asking whether a given newspaper tends to use phrases favored by more Pro-Democracy commentaries. If the phrase is used ˜ heavily and has a large ∆ fp, it would contribute to an increase in θn. The word choices of the reference languages changed over time. Certain words were ideologically relevant in some periods but not others. One might also reasonably suspect that the diagnostic phrases became more polarized in politically sensitive periods. We cannot compare the slant of the newspapers across periods directly: our measure does not tell us whether the newspapers harden or soften their political stance over time. Nevertheless, we can still compare how slant gap evolved over the time period. The slant gap is calculated by subtracting the slant of Oriental Daily from the slant of Apple Daily. In politically sensitive periods, newspapers use phrases that are more diagnostic of their political stance. The word choice would then be closer to the respective benchmark, and further away from each other’s reporting stance. In the appendix, we consider a measure of polarization of the benchmark language over time. The evolution of the slant gap between the two newspapers is presented in figure 1.9. A large difference indicates that the two newspapers is more polarized in their news report slant. In all but one period, the gap is positive, which suggests that Apple Daily is more pro-Democracy than Oriental Daily in general. The periods with large slant gap correspond to periods that intuition would suggest to be politically polarizing. For example, the slant gap increases sharply in 2012 quarter 3, which corresponds to the National and Moral Education controversy. First and second quarter of 2014 also register high slant

34Gentzkow and Shapiro (2010) use speech by all congressmen as references and therefore need to account separately their individual ideological leaning. We have a single reference for each end of the ideology spectrum and therefore, do no have to implement that step. 23 difference. Quarter 4 of 2014, during which the Umbrella Movement took place, hasthe highest slant difference in the sample period.

1.7 Empirical Evidence

The model predicts Apple’s ad share falls in politically turbulent periods. To test this, we examine the correlation between slant gap and the share of ads on Apple relative to Oriental by regressing Apple’s ad share relative to Oriental in quarter q of year t on slant gap only. Without any controls, the resulting coefficient on slant gap is statistically significant at 5% level. Adding linear time trend renders the coefficient statistically insignificant but this is expected given the short time series. As shown in figure 1.7, the percent of connected to total ad assignments is relatively flat throughout the sample period, so it cannot explain the time variation of Apple’s ad share. To visualize the result, we show Apple Daily’s and Sing Tao’s ad share relative to that of Oriental Daily over the sample period. Sing Tao is also known to adopt a pro-Beijing political stance35 but has a much a smaller readership. By comparing and contrasting Apple and Sing Tao relative to Oriental, we can shed light on the importance of newspapers’ political stance on advertising revenue. Figure 1.1 plots the ad share of Apple and Sing Tao relative to the sum of ads on all 3 newspapers (Apple, Oriental and Sing Tao). Figure 1.2 plots the ad share relative to the sum of ads on either Apple or Sing Tao and Oriental only. The dotted straight line is a fitted line over this period. Figure 1 shows that Apple’s ad share fell fromabout39% in 2010 quarter 1 to about 36% in 2014 quarter 4 while Sing Tao’s ad share stayed flat or increased slightly in the same time period. Likewise, figure 1.2 shows that Apple’s ad share fell from around 50% to 45% while Sing Tao remains at around 33% relative to Oriental. In particular, Apple’s ad share declined consistently thoroughly 2014 when the political climate was most polarized. The next model prediction suggests that ad price between apple and oriental diverges in more polarizing times. The first two columns of table 1.3 show the ad price of the full-color page in the run-of-paper36 in the sample period. In 2010, Oriental charged 23%

35Sing Tao is commonly perceived as having a more pro-Beijing stance than Oriental. 36Newspapers charge different prices for ads at different positions, size and color scheme. Wedonot observe the actual price charged for each ad. 24 higher than Apple Daily. The price gap shrink to 6% in 2012 but diverged again to 37% in 2014. Despite the increase in price gap in 2014, Apple’s relative advertising volume still shrunk considerably. This provides indirect evidence that the expected advertising demand for Apple relative to Oriental fell significantly in 2014. Decline in Apple’s readership could potentially explain Apple’s falling ad share. To investigate this channel, I turn to the estimated readership in Apple Daily’s annual investor report. Their estimates came from AC Nielsen37. Unfortunately, Apple only reports the combined readership of print and online readers but not separately. The numbers are presented in columns 3 and 4 of table 1.3. In 2010, the total readership was quite close between the two newspapers, but readership of Oriental Daily gradually shrunk from that point on. In contrary, the readership has remained relatively flat for Apple at around 1,500,000 total readers. We have information on print and online readers composition in 2014 only. We noted that in 2014, Apple’s readers are much more likely to be online readers: the fraction of total readers who are online readers in Apple and Oriental are 41% and 11% respectively. Hence, it is still possible to conclude that the print version of Apple Daily had become less attractive relative to print version of Oriental Daily if we assume that Apple’s online readership grew very fast in the period38. Next, we test the second model prediction that Beijing-friendly advertisers are less likely to advertise on Apple by running a multinomial logistic regression with three choice categories: Apple (A), Oriental (O), and Both Apple and Oriental (AO). Beijing-friendly firms include politically-connected, or mainland firms. We estimate the following model:

β β β δ ν , { , , } P(yitq = b) = 0 + 1ci + 2Xi + tq + ib b = A AO O (1.7)

where yitq is the newspaper choice of advertiser i in year t of quarter q. ci = 1 if firm i is

37Oriental Daily reports a different readership estimate from a different market research estimate intheir annual investor report. They claim that they have about 4 million readers, while population in Hong Kong is about 7 millions. The number seems dubious and we choose to use the estimate of Apple Daily instead. Another reason we use Apple’s estimate is because we have access to AC Nielsen’s Media Index 2014 report also, and have verified their 2014 number. Finally, Oriental only reported the estimated readership fortheir own, but not the other newspapers. 38Assume that both newspapers start with 0% of online readers in 2010, and the readership rises linearly to 41% and 11% for Oriental. Apple’s readership in each year would be: 1566, 1409, 1257, 948, 994. Likewise, Oriental’s readership would be 1457, 1361,1285, 1101, 1030. The gap in readership (Apple- Oriental) would be 109, ,48, -28, -153, -36 in each year. The readership time trend has little correlation with the slant difference in this period. 25

classified as Beijing-friendly, Xi is a vector of firm-specific characteristics including the origin (Local, Foreign, Mainland), the industry fixed effects, and whether the firm is a government or not-for-profit organization. Industry fixed effects control for industry-specific tendency to place ads on a certain newspaper. For example, the youth apparel industry might find a pro-democracy paper more attractive because presumably young people are more likely to adopt a pro-democracy stance. δqt are time controls that include dummy for quarters of the year, and linear time trend. The result is presented in table 1.7 and is interpreted relative to choosing Oriental alone. The top panel shows the estimation result when only Apple was chosen by the advertiser. The bottom panel shows the estimation result when both Apple and Oriental were chosen. In the first three columns I use the full sample of ad assignments, andin columns 4-6 I only include ads that have at least one HKSE-listed firms39. The HKSE- listed firms are larger in size, and many have business exposure to mainland China.In columns 1 and 4, I include industry fixed effects only. In columns 2 and 5, I include both industry fixed effects and time controls. In columns 3 and 6, I also include quarter-specific industry fixed effects. The coefficient on ”Connect” in the top panel is negative and significant atthe 1% level for the all 6 specifications. Adding time controls and industry fixed effect controls in each set of sample does not significantly alter the magnitude of the coefficient. The magnitude of the coefficient of ”Connect” in the HKSE-listed firm sample is about halfof that in the full sample, suggesting that the avoidance from advertising on Apple is weaker among the HKSE-listed firms. On the other hand, listed foreign (mainland) companies are more (less) likely to advertise on Apple than non-listed foreign (mainland) companies. To determine the effect of being connected in the probability scale, we calculate the marginal effects at the median, as shown in parenthesis in the same table. Using the first specification, our result suggests that connected companies are 6.6% and mainland Chinese companies are 18.7% less likely while foreign companies are 2.7% more likely to advertise on Apple alone. This indicates that the aversion to Apple Daily is stronger among mainland companies than connected companies, while foreign companies are more fond of the newspaper. It is possible that foreign companies are not as aware or care about the differences in political stances of

39We do not include control on government, and not-for-profit in columns 3 and 4. 26 the newspapers and simply prefer the newspaper with better cost-effectiveness. Of course, it is also possible that foreign companies prefer newspapers with a pro-Democracy stance. The coefficients are rather stable and robust to different sets of controls for all variables. This is broadly consistent with the model prediction that Beijing-friendly companies are less likely to advertise in a pro-Democracy newspaper. Turning to the bottom panel, the coefficient on ”Connect” are negative and significant at the 1% level using the full sample, but are insignificant in the HKSE-listed sample. The coefficient on ”Foreign” are positive and significant at 5% but are insignificant inthe HKSE-listed sample. The coefficient on ”Mainland” is negative and significant at 1%level in all specifications. Putting them together, the result suggests that it is unlikely forcon- nected or mainland companies to advertise on both Oriental and Apple while the opposite holds for foreign companies. The magnitude for all three variables is considerably smaller than that in the top panel. This suggests that firms’ characteristics have smaller explanatory power on firms’ decision to advertise in Oriental Daily conditional on them advertising in Apple Daily. The model only yields prediction on firms’ choice of newspapers, but not on ad characteristics. In practice, a firm can choose different ad specifications on different newspapers. For example, it can place a large color ad on newspaper A and a small black/white ad on newspaper B. Hence a firm can discriminate newspaper B through the ad specifications it chooses. Our data does not have information on on the ad characteristics so our result only captures the extensive margin of firm’s choice. This concern of intensive margin is more relevant to firms that choose to advertise on both newspapers. If firms only advertise on Apple, even if we observe the ad characteristics, we would not observe their choice ad characteristics and make relevant comparison. Given that only small amount of ads that appear on both newspapers, we believe that the extensive margin is the more important margin in our context. Many politically-connected, as well as the mainland advertisers are large corpo- ration that have business stretching across industries and their ads are mostly politically neutral. For Apple Daily to turn away the advertisements from companies that are politically connected, it would mean to turn away ”regular ads” such as ads from grocery stores, cell phone carrier...etc. This translates to giving up a large portion of advertising revenue, 27 which is difficult for Apple to justify as a listed company on the stock exchange. Inthe appendix table, we also show that Hong Kong government ads do not seem to have a strong preference for Oriental over Apple Daily, as opposed to political ads which seem to exhibit perfect sorting between the two newspapers40. Next, we test the third model prediction that all firms are less likely to advertise on Apple Daily in politically sensitive periods. We consider the following multinomiral logit model:

β β β △θ β · △θ β δ ν , { , , } P(yitq = b) = 0 + 1ci + 2 tq + 3ci tq + 4Xi + tq + ib b = A AO O (1.8)

where △θtq denotes the slant gap in quarter q of year t, ci · △θtq denotes the interaction between connect/mainland dummy and slant gap. β2 is expected to be negative: the larger the slant gap, the less likely a firm would advertise on Apple. β3 measures the possible differential impact of newspapers’ slant difference on Beijing-friendly41 firms . The top panel of table 1.8 shows the estimation result of equation 1.8 for which only Apple was chosen, and the bottom panel for which both Apple and Oriental were chosen. In the top panel, the coefficient on θtq is negative and significant at 1% level for all 3 specifications using the full sample, but the magnitude falls as the specification becomes more stringent. Using the HKSE-listed sample, the significance of coefficient falls to5% significance as shown in column (4) and (5). However, the magnitude of the effect isstronger compared to that in the full sample. The interaction terms are generally insignificant except for listed-mainland firms. Using the specification in column (1), the coefficient saysthat one unit increase in slant gap decreases firms’ probability of advertising in Apple Daily by 4.2%. To put the number in perspective, the average slant gap in 2014 is 0.45435, which means that firms are 1.9% less likely to advertise in Apple because of the divergence ofslant in 2014. This provides evidence that listed mainland firms react more strongly in turbulent times but not the connected firms.

Turning to the bottom panel of table 1.8, the coefficient on △θtq is insignificant in

40We define an ad as pro-Beijing or pro-Democracy by the ad title. For example, the ad title ”” (Strongly oppose Central Occupation) is pro-Beijing and the ad title ”” is considered to be pro-Democracy. 41Our model does not suggest a larger effect on Beijing-friendly firms during turbulent times 28 the full sample but is negative and 1% significant in the HKSE-listed sample, as shown in column (4)-(6). This echoes with the earlier result that the magnitude of slant gap’s effect is stronger among listed companies. The result also indicates that slant gap has strong power in explaining HKSE-listed companies’ choice between advertising on both and advertising on Oriental only. This makes sense because HKSE-companies are more profitable than neutral companies in general and can afford to advertise on both newspapers. Overall, the results are broadly consistent with the mechanism that firms avoided advertising on Apple in politically volatile periods.

1.7.1 Persistence of Slant Gap’s Effect

Our time series unit is set to quarter. This means that the observed advertising decision on a given date could happen before a major political event occur, contradicting what our mechanism suggested. In addition, slanted newspaper reporting in volatile periods can have a lasting impact on firms’ decision. To addresses these two issues, we investigate whether lagged slant gap could affect firms’ newspaper choice in the next period. Firstwe note that the autocorrelation of the slant gap time series is relatively low at 0.13 with a 95% confidence interval between -0.31 and 0.57. This suggests that there is no direct mapping between contemporaneous and lagged slant gap. We consider two regression specifications. In the first specification, we replace slant gap, and all the interaction terms containing slant gap with the lagged slant gap by one quarter in equation 1.8. The rationale is that since advertisers could not foresee the occurrences of political events, they could respond only after they observed newspapers slant in the last quarter. This is especially relevant to ads placed in the beginning of the quarter because firms’ newspaper choice are presumably driven by reporting in thelast quarter. In the second specification, we include both the contemporary and lagged slant gap in the multinomial regression to control for responses :

P(yitq = b) = β0 + β1ci + β2△θtq + β3△θtq−1 + β4ci · △θtq + β5ci · △θtq−1 + β4Xi+ (1.9) δ ν , { , , } tq + ib b = A AO O (1.10)

The results are presented in table 1.9. We used the full sample for both specifications. 29

Focusing on the top panel of which the sample include ads appearing on Apple only, column (1) to (3) present results for which only lagged slant gap and the respective interaction terms are included. In all three columns, lagged slant gap has a negative and significant effect.

Comparing these coefficients with the coefficients on △θtq in table 1.8, the magnitude of the effect appears to be stronger when △θtq−1 is the independent variable. This provides suggestive evidence that advertisers responded more strongly to slant gap in the previous period by shunning Apple. Column (4) to (6) in table 1.9 show the results when both contemporaneous and lagged slant gap are included In all three specifications, both contemporaneous and lagged slant gap are negative and highly significant. Comparing the results with the first 3 columns in table 1.8, it appears that including lagged slant gap only mildly depresses the effect of contemporaneous slant gap on newspaper choice in the top panel. The result suggests that newspaper’s slant had lingering impact on firms’ choice of location. We did not extend our analysis to consider longer lags because of limited data, but occurrence of larger and more intense political events is likely to have a longer effect than smaller ones.

1.7.2 Real Estate Industry

Jimmy Lai’s comment in the interview suggested that the businesses in the real estate industry have avoided advertising in Apple Daily. If this is true, the real estate industry should exhibit higher tendency to avoid Apple relative to other industries even after controlling for political connection and slant gap. To investigate, we first examine the industries fixed effects and their interaction with firm’s political connectivity and slantgap on likelihood of advertising on Apple from regression 1.8. Figure 1.10 plots the confidence interval of the industry fixed effects for adsap- pearing on Apple only and figure 1.10 for ads appearing on both newspapers, using the real estate industry as the base category. For ads appearing on Apple only, all but the Food & Beverage industry have a positive fixed effect. This suggests that most industries are more willing to advertise on Apple relative to the real estate industry. In particular, the automobile industry is 1.1 log odds (.754 in probability) more likely to advertise on Apple alone than the real estate industry. The effect is statistically significant at 95% level for all industries except the Travel and Retail industry. The positive industry fixed effects dis- 30 appear for most industries when we consider ads appearing on both newspapers, as shown in figure 1.11. This suggests that the real estate industry’s preference for Oriental alone is not as strong comparing with both Apple and Oriental. We also examine whether industries exhibited differential response to slant gap. Figure 1.12 plots the interaction term of each industry and slant gap, again using the real estate industry as the base category. The larger confidence intervals show that the effect is less precisely measured compared to the industry fixed effects. The evidence suggests that industries did not react differentially to slant gap shocks. The respective interaction terms are insignificant from zero in many industries as shown in figure 1.13. The industry fixed effect estimates do not allow us to tear out the effect ofpolitical consideration from economic motives. Preferences for advertising on Oriental alone by the real estate industry can be potentially driven by perceived better match between potential customers and Oriental readers. In other words, there might be unobservable newspapers readers characteristics that explain firms’ decision. To partially address this issue, we focus on the connected companies and investigate whether connected companies in the real estate industry has stronger aversion from Apple Daily. Figure 1.14 and 1.15 plot the interaction between firms’ political connectivity with the industries fixed effect for ads that appeared on Apple only and ads that appearedon both, relative to Oriental only. The controls include all regressors in equation 1.8. The interaction term captures the differential likelihood of connected companies to advertise on Apple in different industries. For both figures, the base category is the interaction term between the real estate industry and firm’s political connectivity. The estimates are positive and significant at 95% confidence interval for all industries interaction infigure 1.14. This suggests that connected companies in the real estate industry are much less likely to advertise on Apple alone comparing to other industries even after controlling for industry fixed effects. In particular, the magnitude is largest for the education industry. Fromthe figure, connected companies in the education industry is 2.56 log odds (.93 in probability) more likely than connected companies in the real estate industry to advertise on Apple alone. Figure 1.15 paints a similar picture with all industries have a positive estimate, but the effect is less precisely estimated as evident by the wider confidence intervals. 31

1.7.3 Alternative Measure of Political Climate

Our premise is that the slant gap reflects concurrent political climate. While we have shown in figure 1.9 that periods with a large slant gap is associated with periods with high-profile political events, to further illustrate that firms are wary of the political environment, we consider another proxy for political salience: number of protesters. Protest is a clear political expression: the more the protesters, the stronger the pressure it exerts on the government. Number of protesters thus conveys a sense of political volatility 42, but it can lag slant gap as individuals could become more aggrieved and decided to protest after being exposed to slanted news reports. A number of protests had erupted in our sample period. While the specific demand of each protest was different, all but one of the protests could be classified asanti- government. We count the total number of protesters (estimated by the public opinion program of Hong Kong University 43 ) for all protests in each quarter. Table 1.10 lists all of the street protests and the number of protesters in our sample period. The Pearson correlation between number of protesters and slant gap is 0.419 with a two-tailed p-value of 0.066. This suggests a moderately positive correlation between the two measures of political climate. We estimate the multinomial logistic regression 1.8 again but replace slant gap with the number of protesters on street. Table 1.11 presents the result on the Protesters variable only. As before, columns (1)-(3) use the full sample whereas columns (4)-(6) use ad assignments from HKSE-listed companies only. In both samples, the coefficient on the number of protesters in the top panel is negative and statistically significant at the 1% level in the specification with least number of controls but the power declines rapidly withan increasing set of control. The same pattern can be observed in the full sample in the bottom panel. In the bottom panel, the listed sample is statistically significant at 1% level across all 3 specifications. The result indicates that number of protesters has strong statistical power in explaining HKSE-listed companies’ choice between advertising on both newspapers and advertising on Oriental only.

42Acemoglu et al (2014) used a similar approach in estimating the effect of protests in Arab Springs on Egyptian stocks 43The website: https://www.hkupop.hku.hk/english/features/rallies/summary.html. The website provides an estimated range of the number of protesters for each protest, and we use the middle of the range. 32

Since the observation on the number of protesters and slant gap is limited to 20 quarters, our result inevitably suffers from a lack of power. Nevertheless, both measure - slant gap and number of protesters - produce results consistent with our hypothesis that Apple is systematically discriminated when politics becomes a focal point in daily life.

1.8 Mechanisms

Although the empirical evidence presented above is consistent with a theory in which 1). firms’ political salience increase in politically-volatile periods and 2). firms close to the government suffer a cost from advertising in Apple, other explanations may be suggested. First, one might suspect that direct political pressure from the mainland government could intensify during volatile times as motivated by anecdote. In other words, the mainland Chinese government could directly persuade companies to boycott Apple Daily behind the scene. While we cannot dispel the hypothesis completely, given the large advertisers pool, it is quite unlikely that the government can apply direct pressure on each and every individual firm. To start with, one would expect direct pressure to be appliedto politically-connected companies only. However, our result indicates that neutral firms also respond to slant gap, suggesting that firms refrain from Apple in self-motivated manner. The second plausible mechanism is that firms have an economic interest in advertising to readers with a certain political stance only. To a certain extent, wealthier individuals tend to be more politically conservative (Powdthavee and Oswald, 2014) so readers’ political ideology could be a strong predictor of their purchasing power. To test this hypothesis, we compare advertising responses of different industries when slant gap is large. The idea is that there will be increased sorting of readers in volatile times, and the match between industries and firms would improve in volatile period. If firms are driven by economic motives, we would expect to see industries that target young customers to shift to Apple, and industries that target older customers to shift to Oriental. Figure 1.12 and figure 1.13, which plot the interaction terms between industry and slant gap, shed light on the validity of this hypothesis. Tor the majority of other industries, the effect is not statistically significant, which suggest that firms are not very motivated by improved newspaper-reader match. The result provides confidence to our interpretation that firms 33 react to political shocks rather than economic shocks.

1.9 Apple Daily’s Revenue Loss

We are interested in separately estimating the revenue impact on Apple Daily in 2014 due to 1). a heightened political awareness and 2). Beijing-friendly firms’ preference for Oriental Daily over Apple Daily. To calculate the effect of 1), we use regressions results in table 1.8 to help forecast what ad volume would have been in 2014 if the degree of political awareness remained fixed at its values in 2010. To calculate the effect of2),we use the regression results in table 1.7 to predict the number of ads placed by connected and mainland companies on Apple if they do not exhibit political preference. We focus on 2014 alone because both qualitative and quantitative evidences suggest that it is the most politically volatile period. In order to calculate the impact on revenue, we need to know the size of ad that firms will buy, which is unavailable in the Wisers dataset. To circumvent this issue,I collected a smaller sample of 2045 physical ads with ad size from a local public library in Hong Kong and manually recorded the size of the ads (Excluding tender, announcement, event promotion) from the following randomly selected 10 days in 2014: 3/6, 4/4, 5/3, 6/4, 7/1, 8/4, 9/6, 10/10, 11/3, 12/7. The dates are chosen arbitrarily to cover several politically important dates such as 6/4 (Annual anniversary of Tiananmen Student Massacre ) and 7/1( Annual protest to voice various demands to the HKSAR government). The average ad size using all ads is 0.588 page, and connected and mainland ads is 0.783 page. Revenue Impact of Political Salience - The average slant gap is 0.20497 in 2010, and 0.45435 in 2014. Slant gap is larger in all quarters in 2014 except quarter 3 compared with 2010. Using the estimation result of −0.042 from column (1) in the top panel of table

1.8 as our β2 estimate and holding the industry effects at their means, this yields a 1.05 percent drop in probability of advertising on Apple for neutral firms44. This is equivalent to an ad loss quantity of 196 ads for Apple Daily45. Finally, we multiply the average ads size and the ad price in 2014 to the ad loss quantity to arrive at a ad revenue loss for Apple

44(0.45435 - 0.20497) * 4.2% = 1.05% 45The combined (Apple + Oriental) ad quantity was 18660 in 2014. 18660 ∗ 0.0105 = 196 34 of $HKD 26.3 million($USD 3.4 million) 4647 Revenue Impact of Aversion by Beijing-friendly firms - Connected companies are 6.6% and mainland companies 18.8% less likely to advertise on Apple in 2014, which translates to a total ad loss of 273 ads in Apple Daily.48, and leads to an ad revenue loss of $HKD 48.7 million ($USD 6.1 million ) 49. Putting the number in perspective, the advertising revenue at Apple in 2014 is HKD 343.7 million. This means that ad revenue loss due to political salience amounts to 7.7% of total advertising revenue, and Beijing-friendly firms’ political preference contributes to another 14.2%. In sum, Apple Daily loses 21.9% of its advertising revenue due to political reasons. In an economy where it is increasingly difficult for print media to remain profitable, the effect of politics on media is sizable. For smaller newspapers, this financial pressurecould very well affect their position on the political spectrum. An obvious limitation of the above calculation is that we cannot account for the endogenous ad price adjustment. Newspapers can adjust the ad price in expectation of heightened political awareness or knowing the fraction of Beijing-friendly firms in the pool of potential advertisers. Lowing the ad price could potentially mitigate the revenue impact. However, given the short time series, we are unable to say much concerning the degree to which the evolving prices were due to increased political awareness or political preferences of Beijing-friendly firms. Another limitation of this calculation is that we do not consider possible existence of pro-Democracy companies. The existence of such companies would generate a politically- induced economic benefit for Apple Daily. While our regression analysis suggests that Foreign firms are more likely to advertise on Apple, it is unclear whether foreign firmsare pro-Democracy or simply react to the economic incentives due to a more attractive price-to- reader ratio at Apple. Finally, Apple’s readership might increase in volatile periods, leading to higher profit from paper sales. But since we do not observe readership composition at

46The revenue impact due to neutral advertisers is : 196 ∗ .588page ∗ 1886.8cm2 ∗ 120.84$/cm2=$HKD 26,276,649 47We have ignored the ads that advertise on both newspapers because the coefficient on slant gap is not significant. 48Connected firms accounts for 15% of total ad, and mainland firm 2.5%. There was a total of18660ads in 2014. The expected number of ad loss due to firms’ political connection: 18660*0.15*0.066 = 185 ads, and mainland firms: 18660*0.025*0.188 = 88 ads. Total ad loss: 185 +88=273 49−273 ∗ .783page ∗ 1886.8cm2 ∗ 120.84$/cm2= 48.7 million 35 different time period, we cannot account for the effect of this channel onrevenue.

1.10 Conclusions

This paper has provided empirical evidence on the effect of newspapers’ political stance on firms’ newspaper choice of advertising. Using daily advertising data ofthe two major newspapers in Hong Kong, I have shown that increase in political salience in politically-sensitive period leads to stronger aversion from the pro-Democracy Apple Daily among advertisers. Furthermore, Beijing-friendly advertisers, which include politically connected and mainland Chinese firms, exhibit stronger aversion from pro-Democracy newspaper relative to the neutral firms even in relatively stable periods. Using the regression results, I estimated that Apple Daily suffered from an ad revenue loss equivalent to 21.9% of its total advertising revenue in 2014 due to political reasons. It is important to note that our reduced-form findings do not account for the endogenous ad price. Accounting for price change will likely amplify the effect of political reasons on Apple’s advertising revenue because Apple Daily could adjust ad price to attract advertisers. While Hong Kong has a unique political institution, the relevance of the findings is not restricted to regions or countries of specific political system. Rather, the implications pertain to both democratic or nondemocratic countries in which large businesses share cozy relationship with the government (e.g. South Korea50) or owned by the government (e.g. China). This paper has shown that businesses can inherit government’s preference and behave as extension of the state. The mechanism highlighted in this paper - politically- induced advertising pressure - can generate an unfavorable impact on the media’s willingness to adopt a liberal political stance. This finding is especially important in an era when online news, of which the main source of revenue is advertising but not readers subscription, become more a dominant information channel. There are several papers that analyze media bias in response to advertisements in specific contexts51 but this paper did not address whether increasing advertising pressure could lead to an intensification of self-censorship. This is an important question leftfor

50Schoenherr documented that politically connected firms in South Korea allocated contracts in favor of firms from the same connected network, resulting in a total annual cost of about 0.21-0.32% ofGDP. 51Reuter and Zitzewitz (2006) studied mutual fund recommendation and Dewenter and Heimeshoff (2014) studied automobile reviews in response to advertising from the respective products. 36 future research. Chapter 1, in full, is currently being prepared for submission for publication of the material. Lam, Onyi. ”Advertisers Capture: Evidence from Hong Kong”. The dissertation author was the sole author of this material. 37

Bibliography

[1] D. Acemoglu, T. Hassan, and A. Tahoun. The power of the street: Evidence from egypt’s arab spring. C.E.P.R. Discussion Papers, 2014.

[2] R. K. Aggarwal, F. Meschke, and Y. T. Wang. Corporate political donations: Invest- ment or agency? Business and Politics, 14(1), 2012.

[3] T. Besley and A. Prat. Handcuffs for the grabbing hand? media capture and government accountability. American Economic Review, 96(3):720–736, 2006.

[4] P. Bordalo, N. Gennaioli, and A. Shleifer. Salience and consumer choice. Journal of Political Economy, 121(5), 2013.

[5] P. Bordalo, N. Gennaioli, and A. Shleifer. Competition for attention. Review of Economic Studies, 2015.

[6] J. Y. Cheng. The emergence of radical politics in hong kong: Causes and impact. China Review, 14(1), 2014.

[7] A. Cheung and P. Wong. Who advised the hong kong government? the politics of absorption before and after 1997. Asian Survey, 44(6), 2004.

[8] F. Cingano and P. Pinotti. Politicians at work. the private returns and social costs of political connections. Journal of the European Economic Association, 11(2):433–465, 2013.

[9] R. Coulomb and M. Sangnier. The impact of political majorities on firm value: Do electoral promises or friendship connections matter? Journal of Public Economics, 115:158–170, 2014.

[10] S. DellaVigna. Psychology and economics: Evidence from the field. Journal of Eco- nomic Literature, 47(2):315–72, 2009.

[11] S. DellaVigna, R. Durante, B. Knight, and E. La Ferrara. Market-based lobbying: Evidence from advertising spending in italy. American Economic Journal: Applied Economics, 2014.

[12] R. Dewenter and U. Heimeshoff. Media bias and advertising: Evidence from a german car magazine. Review of Economics, 65, 2014.

[13] R. Di Tella and I. Franceschelli. Government advertising and media coverage of corruption scandals. American Economic Journal: Applied Economics, 3(4):119–51, 2011.

[14] T. Eisensee and D. Strömberg. News droughts, news floods, and u. s. disaster relief. Quarterly Journal of Economics, 122(2):693–728, 2007.

[15] M. Ellman and F. Germano. What do the papers sell? a model of advertising and media bias. The Economic Journal, 119, 2009.

[16] R. Enikolopov, M. Petrova, and E. Zhuravskaya. Media and political persuasion: Evi- dence from russia. American Economic Review, 101(7):3253–85, 2011. 38

[17] M. Faccio. Politically connected firms. American Economic Review, 96(1):369–386, 2006.

[18] R. Fernandez and D. Rodrik. Resistance to reform: Status quo bias in the presence of individual-specific uncertainty. American Economic Review, 81(5):1146–55, 1991.

[19] S. Gehlbach and K. Sonin. Government control of the media. Journal of Public Economics, 118, 2014.

[20] M. Gentzkow. Valuing new goods in a model with complementarities: Online newspapers. American Economic Review, 97(3), 2007.

[21] M. Gentzkow, N. Petek, J. Shapiro, and M. Sinkinson. Do newspapers serve the state? incumbent party influence on the us press, 1869-1928. Journal of the European Economic Association, 2015.

[22] M. Gentzkow and J. Shapiro. Media bias and reputation. Journal of Political Economy, 114(2), 2006.

[23] M. Gentzkow and J. Shapiro. What drives media slant? evidence from u.s. daily newspapers. Econometrica, 78(1), 2010.

[24] M. Gentzkow, J. Shapiro, and M. Taddy. Measuring polarization in high-dimensional data: Method and application to congressional speech. Working Paper, 2016.

[25] A. Gerber, D. Karlan, and D. Bergan. Does the media matter? a field experiment measuring the effect of newspapers on voting behavior and political opinions. American Economic Journal: Applied Economics, 1(2):35–52, 2009.

[26] T. Groseclose and J. Milyo. A measure of media bias. Quarterly Journal of Economics, 120(4):1191–1237, 2005.

[27] J. Jensen, E. Kaplan, S. Naidu, and L. Wilse-Samson. Political polarization and the dynamics of political language: Evidence from 130 years of partisan speech. Brookings Papers on Economic Activity, 2012.

[28] A. I. Khwaja and A. Mian. Do lenders favor politically connected firms? rent provision in an emerging financial market. Quarterly Journal of Economics, 120(4):1371–1411, 2005.

[29] B. Knight and C.-F. Chiang. Media bias and influence: Evidence from newspaper endorsements. Review of Economic Studies, 78(3):795–820, 2011.

[30] B. Koszegi and A. Szeidl. A model of focusing in economic choice. Quarterly Journal of Economics, 128(1):53–104, 2013.

[31] S. Luechingera and C. Moser. The value of the revolving door: Political appointees and the stock market. Journal of Public Economics, 119:393–10749, 2014.

[32] N. Ma. Political Development in Hong Kong: State, Political Society, and Civil Society. Hong Kong University Press, 2007. 39

[33] J. McMillan and P. Zoido. How to subvert democracy: Montesinos in peru. Journal of Economic Perspectives, 18(4), 2004.

[34] I. Najih and M. Yanai. Media coverage of fukushima nuclear power station accident 2011 (a case study of nhk and bbc world tv stations). Procedia Environmental Sciences, 17, 2013.

[35] A. Oswald and N. Powdthavee. Does money make people right-wing and inegalitarian? a longitudinal study of lottery winners. IZA Discussion Paper No. 7934, 2014.

[36] M. Petrova. Newspapers and parties: How advertising revenue created an independent press. American Political Science Review, 104(4):790–808, 2011.

[37] A. Prat and D. Stromberg. The political economy of mass media. CEPR Discussion Papers, (8246), 2011.

[38] B. Qin, D. Stromberg, and Y. Wu. The determinants of media bias in china. Working Paper, 2014.

[39] J. Reuter. Does advertising bias product reviews? an analysis of wine ratings. Journal of Wine Economics, 4(2):125–151, 2009.

[40] K. Reuter and E. Zitzewitz. Do ads influence editors? advertising and bias in the financial media. Quarterly Journal of Economics, 121(1):197–227, 2006.

[41] D. Schkade and D. Kahneman. Does living in california make people happy? a focusing illusion in judgments of life satisfaction. Psychological Science, 9(5), 1998.

[42] D. Schoenherr. Political connections and allocative distortions. Unpublished.

[43] M. Sing. Politics and Government in Hong Kong: Crisis Under Chinese Sovereignty. Routledge, 2008.

[44] J. Snyder and D. Strömberg. Press coverage and political accountability. Journal of Political Economy, 118(2), 2010.

[45] M. Spence. Job market signaling. The Quarterly Journal of Economics, 87(3):355–374, 1973.

[46] D. Strömberg. Radio’s impact on public spending. Quarterly Journal of Economics, 119(1):189–221, 2004.

[47] D. Wank. The institutional process of market clientelism: Guanxi and private business in a south china gity. The China Quarterly, 9(147), 1996. 40

1.11 Figures and Tables

Figure 1.1: Ad Share on print version of Apple Daily and Sing Tao relative to Sum of Apple, Oriental and Sing Tao.

Note: Tender ads, and ads that are event promotion, announcement are excluded. Sing Tao issued an ad-filled real-estate magazine that was free with the purchase of the newspaper between 2010to early 2012. This makes the number of ads from the real estate industry very high in that period. Therefore we excluded ads by the real estate industries because to make the comparison across time consistent. 41

Figure 1.2: Ad Share on print version of Apple Daily and Sing Tao relative to Oriental Only.

Figure 1.3: 2016 Press Freedom Index issued by the Freedom House 43

Table 1.1: Readership (’000) in 2014 (Including online readership)

Apple Economic Journal Economic Times Ming Oriental Sing Tao Sun Apple 1684 54 65 168 316 103 119 Economic Journal 54 93 37 47 19 13 12 Economic Times 65 37 142 32 55 37 28 Ming 168 47 32 356 76 83 83 Oriental 316 19 55 76 1158 48 60 Sing Tao 103 13 37 48 83 157 36 Sun 119 12 28 36 83 60 224 Note: 2014 Data from AC Nielsen Media Index Report. 44

Figure 1.4: Hang Seng Index and relative Stock Price of the corporations that own Apple Daily (Next Digital, Ticker: 0282) and Oriental Daily (Oriental Press, Ticker: 0018)

Note: The data comes from Google Finance. The percentage represents the stock price relative to that on 1/1/2010. The spike in Apple Daily’s stock price in 2012 corresponded to the sale of its Taiwanese subsidiary.

Table 1.2: Apple Daily’s Revenue Composition

year subscription advertising fraction of advertising revenue to total revenue 2010 388,600,000 706,600,000 0.645 2011 349,900,000 679,700,000 0.662 2012 308,800,000 596,900,000 0.659 2013 305,400,000 500,600,000 0.621 2014 264,000,000 343,700,000 0.566 Note: Data comes from Next Digital’s annual investor relations report 45

Figure 1.5: Total number of ads assignments and ads assignments placed by connected organizations in each industry.

Figure 1.6: Ads Assignment by company characteristics 46

Figure 1.7: Percentage of Connected Ad Assignment

Figure 1.8: Percentage of Mainland Ad Assignment 47

Table 1.3: Ad Price and Readership

Apple Oriental Apple Oriental Ad Price Readership (Print + Online) 2010 85.3 104.95 1,566 1,457 2011 109.02 124.88 1,535 1,392 2012 114.48 121.95 1,503 1,344 2013 114.48 150.68 1,411 1,207 2014 120.84 165.92 1,684 1,158 Note: Ad price measures the price per cm2 of color ads in the run-of-paper. The unit of readership is 1000. Price data from 2010-2012 comes from Wisers. Prices from 2010-2012 come from a private mainland company (http://www.cmtad. com.cn/). Unit of price is Hong Kong dollars per cm2. Reader- ship data from Next Media annual financial report.

Table 1.4: Readers Demography. The number indicates the fraction of readers reading the respective newspaper, except average Age, household income and totals. Source: AC Nielsen 2014.

Apple Daily Oriental Daily Total Average Age 43 47 40 Average Household Income HK$ Per Month 33878 32835 32115 Female 46.37 48.4 53.69 Government Housing 35.58 34.08 29.21 Working 71.98 65.44 63.34 Student 3.73 4.74 11.71 Retired 7.76 11.62 7.3 Unemployed 4.03 7.84 6.49 Primary Completed 11.49 16.07 10.52 F4-F5 24.4 25.17 25.06 University Or Above 22.58 17.23 27.88 48

Figure 1.9: Difference in slant between the two newspapers. 49

Table 1.5: Phrases with highest χ2 and used by Wenwui Daily in each year and quarter

Bigram Trigrm Quadgram 2010 Q1 (Hong Kong) (Renminbi) (Civic Party & League of Social Democrats) (Economy) (Investor) (Financial Crisis) (Market) (Real Estate) (Mainstream Opinion) Q2 (Hong Kong) (Opposition Group) (Finance Crisis) (Economy) (Renminbi) (Civic Party & League of Social Democrats) (Real Estate Market) (Investor) (Mainstream Opinion) Q3 (Hong Kong) (Renminbi) (Columbarium field) (China) (Diao-Yu Islands) (Financial Crisis) (Economy) (Octopus Card) (Switch Off Idling Vehicles) Q4 (Hong Kong) (Renminbi) (Directly Subsidized Schools) (Economy) (Investor) (Monetary Policy) (United States) (Real Estate) (Inflationary Pressure) 2011 Q1 (China) (Libya) (Minimum Wage) (Economy) (Nuclear Crisis) (Inflationary Pressure) (United States) (League of Social Democrats) (Economist) Q2 (China) (Renminbi) (Inflationary Pressure) (Economy) (Plasticizer) (Small and Medium-Sized Enterprises) (Hong Kong) (Research Institute) (Food Safety) Q3 (Economy) (Investor) (Debt Crisis) (China) (Renminbi) (European Debt Crisis) (Development) (Research Institute) (Price of Pork) Q4 (Economy) (Renminbi) (European Debt Crisis) (China) (Opposition Group) (Monetary Policy) (Market) (Reserve Fund) (Debt Crisis) 2012 Q1 (China) (Renminbi) (Love the Country Love Hong Kong) (Market) (Syria) (Monetary Policy) (Market) (Competitive Power) (Research Institute) Q2 (China) (Philippines) (European Debt Crisis) (Economy) (Legislative Council) (Economists) (United States) (Renminbi) (Monetary Policy) Q3 (Japan) (Diao-Yu Islands) (Chinese Government) (China) (Nationalization) (Economist) (Economy) (Research Institute) (Restructuring) Q4 (China) (Diao-Yu Islands) (Reform and Open) (Economy) (Xi Jinpin) (Chinese Navy) (Japan) (18th National Congress of the Communist Party of China) (Minimum Wage) 2013 Q1 (China) (Diao-Yu Islands) (Policy Address) (Japan) (Xi Jinpin) (Fire Control Radar) (Economy) (Renminbi) (Chinese Military) Q2 (China) (Diao-Yu Islands) (Improve Livelihood) (Japan) (Opposition Group) (National People’s Congress) (Economy) (Hong Kong Confederation of Trade Unions) (Pan-politicalization) Q3 (China) (Opposition Group) (Love the Nation Love Hong Kong) (Economy) (Clifford Hart) (Chinese Navy) (Japan) (Landfill) (Rights in the Ocean) Q4 (China) (Opposition Group) (Third Plenary Session) (Japan) (Basic Law) (Universal Suffrage of Chief Executive) (Hong Kong) (Diao-Yu Islands) (Unflinching) 2014 Q1 (Hong Kong) (Opposition Group) (Policy Address) (Economy) (Basic Law) (Fulfill Universal Suffrage) (United States) (Run Run Shaw) (Militarism) Q2 (United States) (Opposition Group) (Occupy Central Referendum) (China) (Diao-Yu Islands) (Northeast New Territories) (Japan) (Finance Committee) (Militarism) Q3 (Hong Kong) (Opposition Group) (Fulfill Universal Suffrage) (Accord to Law) (Basic Law) (National People’s Congress’s Decision) (Council Members) (Benny Tai) (Mainstream Opinion) Q4 (Occupy Central) (Opposition Group) (Occupation) (Economy) (Shanghai-Hong Kong Stock Connect) (Occupy Central Schemer) (Police) (Basic Law) (Law-Violating Central Occupation) 50

Table 1.6: Phrase with highest χ2 and used by pro-Democracy politicians in each year and quarter

Bigram Trigrm Quadgram 2010 Q1 (Democracy) (Post-’80) (Functional Constituency) Constitutional Forms) (Benny Tai) (De facto Referendum) (People) (Liu Xiaobo) (Charter 08) Q2 (Democracy) (Hong Kong people) (Functional Constituency) (Voting) (Democratic Group) (De facto Referendum) (Election) (Benny Tai) (Democratic Movement) Q3 (Democracy) (Benny Tai) (Judiciary System) (Chief Executive) (Basic Law) (Functional Constituency) (Politics) (Article 23) (Chief Executive) Q4 (Democracy) (Liu Xiaobo) (National Education) (Rule of Law) (Zhao Lianhai) (Socialism) (Counts) (Legislative Concuil) (Universal Values) 2011 Q1 (Government) (Legislative Concuil) (Judicial Review) (Democracy) (Benny Tai) (Headstrong) (Border Entry) (Hong Kong people) (High Degree of Autonomy) Q2 (Courts) (Ai Weiwei) (Judiciary Independence) (Apply to) (Legislative Council) (National Education) (Legislative) (Benny Tai) (Court of Final Appeal) Q3 (Elections) (Hong Kong people) (Freedom of Press) (Law) (Candidates) (Core Values) (Candidate List) (Basic Law) (Martial Arts Fiction) Q4 (Chief Executive) (Law Reform Commission) (Judicial Review) (CY Leung Chun Ying) (Private Universities) (Law) (Civic Party) (6.5 billions) 2012 Q1 (Chief Executive) (Hong Kong people) (HKSAR government) (Council Members) (CY Leung Chun Ying) (HKSAR government) (Monitor) (Liaison Office of the PRC in HKSAR) (Core Values) Q2 (Law) (CY Leung Chun Ying) (Rules of Procedure) (Rule of Law) (Liaison Office of the PRC in HKSAR) (Secretary of Justice) (Democracy) (Li Wangyang) (Court of Final Appeal) Q3 (Democracy) (Creditor) (Civil Education ) (Education) (Benny Tai) (Negative Criticism) (Sequence) (CY Leung Chun Ying) (Judiciary Traditions) Q4 (Chief Executive) (CY Leung Chun Ying) (Judiciary Independence) (Courts) (Law Circles) ( Court of Final Appeal) (Judges) (Hong Kong people) (Civic Society) 2013 Q1 (Law) (CY Leung Chun Ying) (Peaceful Occupation of Central) (Lawyers) (Benny Tai) (Civil Disobedience) (Action) (Hong Kong people) (Peaceful Occupation of Central) Q2 (Negotiation) (Hong Kong people) (Occupy Central) (Democracy) (Timothy Tong) (Red Cross) (Father) (CY Leung Chun Ying) (Commissioner of ICAC) Q3 (Nomination) (Injustice) (Civil Disobedience) (Justice) (Committee) (Peaceful Occupation of Central) (Voter) (Democratic group) (Civic Society) Q4 (Election) (Mandela) (Court of Final Appeal) (Son) (CY Leung Chun Ying) (New Hong Kong people) (Screening) (Hong Kong people) (Freedom of Speech) 2014 Q1 (Media) (Hong Kong people) (Freedom of Press) (Ming Pao) (Kevin Lau) (Freedom of Speech) (Democracy) (Li Wei-ling) (Public Interest) Q2 (Chinese Communist) (Hong Kong people) (Civil Nomination) (Democracy) (Democratic Group) (Peaceful Occupation of Central) (Citizen) (Nominating Committee) (Election rules) Q3 (Democracy) (Hong Kong people) (Civil Disobedience) (Screening) (Take it on board first) (Peaceful Occupation of Central) (Disobedience) (Law Circles) (Judiciary Independence) Q4 (Citizen) (Hong Kong people) (Democratic Movement) (Umbrella) (CY Leung Chun Ying) (Civil Disobedience) (Democracy) (Candidate) (Joint Declaration) 51

Table 1.7: Multinomial Logit Regression Results

Dependent variable: 1 if only Apple is chosen Connect −0.2799∗∗∗ -0.2457∗∗∗ -0.2463∗∗∗ -0.1108∗∗∗ -0.1165∗∗∗ -0.1173∗∗∗ 0.014 0.014 0.014 0.022 0.022 0.022 (-0.0664) (-0.0570) (-0.0573) (-0.0273) (-0.0264) (-0.0268) Foreign 0.1133∗∗∗ 0.0948∗∗∗ 0.0948∗∗∗ 0.3814∗∗∗ 0.3861∗∗∗ 0.3858∗∗∗ 0.011 0.011 0.011 0.036 0.036 0.036 (0.0272) (0.0222) (0.0223) (0.0868) (0.0830) (0.0836 ) Mainland −0.7883∗∗∗ -0.7814∗∗∗ -0.7806∗∗∗ -1.1421∗∗∗ -1.1417∗∗∗ -1.1413∗∗∗ 0.031 0.031 0.031 0.061 0.061 0.061 (-0.1877) (-0.1823) (-0.1826) (-0.2589) (-0.2448) (-0.2468)

Dependent variable: 1 if both Apple and Oriental is chosen Connect −0.2165∗∗∗ -0.2079∗∗∗ -0.2076∗∗∗ 0.0870 0.0857 0.0879 0.034 0.035 0.035 0.055 0.055 0.055 (-0.0023) (-0.0028) (-0.0027) (0.0056) (0.0040) (0.0039 ) Foreign 0.0669∗∗∗ 0.0612∗∗ 0.0611∗∗ 0.1339 0.1295 0.1313 0.026 0.026 0.026 0.109 0.109 0.109 (0.0003) (0.0006) (0.0006) (-0.0006) (1.29e-05) (1.908e-05) Mainland -0.5640∗∗∗ -0.5620∗∗∗ -0.5622∗∗∗ -0.4582∗∗∗ -0.4512∗∗∗ -0.4522∗∗∗ 0.078 0.078 0.078 0.132 0.133 0.133 (-0.005) (-0.0062) (-0.0061) (-0.0006) (-0.0022) (-0.0020) Industry FE X X X X X X Quarter FE X X X X X X Linear Time Trend X X X X X X IndustryQuarter FE X X X X Quadratic Time Trend X X Observations 206425 206425 206425 45036 45036 45036

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. The first row of each variable shows the coefficient, the second the standard deviation, and the third the marginal effect. Column (4)-(6) uses ad sample from companies listed onHKSE. The second row of the coefficients indicate the p-value. Industry classification consists of the following categories:Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel). 52

Table 1.8: Results on Slant Gap and Interactions

Dependent variable: 1 if only Apple is chosen ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗ △θtq -0.1722 -0.1721 -0.1621 -0.2392 -0.1966 -0.1652 0.037 0.037 0.039 0.101 0.101 0.106 (-0.0420) (-0.0408) (-0.0381) (-0.0443) (-0.0353) (-0.0281) ∗ Connect·△θtq 0.0433 0.0644 0.0636 0.2277 0.1676 0.1645 0.080 0.080 0.080 0.124 0.125 0.125 (0.0021) (0.0097) (0.0096) (0.0354 ) (0.0247) (0.0246) ∗∗∗ ∗∗∗ ∗∗∗ Mainland ·△θtq -0.2465 -0.2004 -0.1993 -1.1364 -1.1744 -1.1734 0.185 0.185 0.185 0.407 0.407 0.409 (-0.0520) (-0.0423) (-0.0421) (-0.2863) (-0.2698) (-0.2705) Foreign ·△θtq 0.0854 0.0686 0.0677 0.2537 0.2623 0.2600 0.06 0.061 0.061 0.192 0.194 0.194 (0.0217) (0.0169) (0.0167) (0.0506) (0.0510) (0.0510)

Dependent variable: 1 if both Apple and Oriental is chosen ∗∗∗ ∗∗∗ ∗∗∗ △θtq -0.0660 -0.0659 -0.0953 -0.8321 -0.8885 -0.9929 0.092 0.092 0.097 0.275 0.279 0.290 (0.0008) (0.0002) (-0.0007) (-0.0262) (-0.0211) (-0.0230) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Connect·△θtq 0.5240 0.5290 0.5310 1.2986 1.3955 1.4064 0.195 0.196 0.196 0.319 0.323 0.323 (0.0160) (0.0132) ( 0.0131) (0.0429) (0.0343) (0.0331) ∗∗ ∗∗ ∗∗ Mainland ·△θtq -0.5759 -0.5555 -0.5582 1.6677 1.7057 1.7062 0.491 0.492 0.492 0.803 0.807 0.808 (-0.0143) ( -0.0124) (-0.0123) (0.0743) (0.0535) (0.0513) Foreign ·△θtq -0.0236 -0.0325 -0.0312 0.6201 0.7178 0.7216 0.148 0.148 0.148 0.608 0.606 0.605 (-0.0022) (-0.0016) (-0.0016) (0.0185) (0.0162) (0.0156) Firms Characteristics X X X X X X Industry FE X X X X X X Quarter FE X X X X X X Linear Time Trend X X X X X X IndustryQuarter FE X X X X Quadratic Time Trend X X Observations 206425 206425 206425 45036 45036 45036

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. The first row of each variable shows the coefficient, the second the standard deviation, and the third the marginal effect. Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. Column (4)-(6) uses ad sample from companies listed on HKSE. Firms characteristics include political connectivity and country of origin (foreign, mainland). Industry classification consists of the following categories:Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel). 53

Table 1.9: Results on Slant Gap Lags

Dependent variable: 1 if only Apple is chosen ∗∗∗ ∗∗∗ ∗∗∗ △θtq -0.1814 -0.1767 -0.1586 0.038 0.038 0.040 (-0.0431) (-0.0422) (-0.0376) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ △θtq−1 -0.1875 -0.2237 -0.2001 -0.1609 -0.1976 -0.1864 0.041 0.042 0.042 0.042 0.042 0.043 (-0.0432) (-0.0525) (-0.0467) (-0.0369) (-0.0463) (-0.0435)

Dependent variable: 1 if both Apple and Oriental is chosen

△θtq -0.0752 -0.0728 -0.0906 0.094 0.095 0.098 (0.0001) (8.487e-05) (-0.0006) ∗ △θtq−1 -0.1654 -0.1690 -0.1810 -0.1548 -0.1579 -0.1700 0.102 0.103 0.105 0.104 0.104 0.105 (-0.0029) (-0.0019) (-0.0024) (-0.0029) (-0.0019) (-0.0023) Firms Characteristics X X X X X X Industry FE X X X X X X Quarter FE X X X X X X Linear Time Trend X X X X X X IndustryQuarter FE X X X X Quadratic Time Trend X X Observations 206425 206425 206425 206425 206425 206425

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. The second row of the coefficients prints the standard deviation and the third prints the marginal effects at the median.Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. Column (4)-(6) uses ad sample from companies listed on HKSE. Firms characteristics include political connectivity and country of origin (foreign, mainland). Industry classification consists of the following categories:Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel).

Table 1.10: List of Major Protests in Hong Kong between 2010 and 2014

Date Protest Number of Participants (Estimated) 2010/1/1 New Year Protest 10000-12000 2010/7/1 Annual July 1st Protest 22000-26000 2011/7/1 Annual July 1st Protest 59000-67000 2012/4/1 Protest against Mainland Influence on the Chief Executive Election Few thousands 2012/06/10 Death of Li Wangyang 25000 2012/7/1 Annual July 1st Protest 90000-100000 2012/7/29 National and Moral Education 90000 2012/12/30 Support Leung Protest 50000 2013/1/1 New Year Protest 30000-33000 2013/7/1 Annual July 1st Protest 88,000-98,000 2013/10/20 HKTV Free Licence Controversies 120,000 2014/1/1 New Year Protest 13,000-16,000 2014/3/2 Kevin Lau Knife Attack 13,000 2014/6/27 White Paper Controversies 1800 2014/7/1 Annual July 1st Protest 150,000 - 166,000 2014/8/17 Anti- Occupy Central 79,000-88,000 2014/9/25 Student Protest for Universal Suffrage 4000 2014/09/28-2014/12/11 Umbrella Movement > 200,000 Note: The estimate of the number of participants comes from the public opinion programme conducted by the University of Hong Kong. 54

Table 1.11: Results on Protesters Turnout and Interactions

Dependent variable: 1 if only Apple is chosen ∗∗∗ ∗∗∗ ∗∗ ∗ Protesterstq -0.0008 -1.873e-05 5.569e-05 -0.0018 -0.0007 -0.0005 7.59e-05 9.81e-05 0.000 0.000 0.000 0.000

Dependent variable: 1 if both Apple and Oriental is chosen ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Protesterstq -0.0009 0.0004 0.0004 -0.0037 -0.0282 -0.0024 0.000 0.000 0.000 0.001 0.006 0.001 Firms Characteristics X X X X X X Firms Characteristics Interaction X X X X X X Industry and Quarter FE X X X X X X Linear Time Trend X X X X IndustryQuarter FE X X Quadratic Time Trend X X Observations 206425 206425 206425 45036 45036 45036

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. Protesters turnout measured in 000s. The second row of the coefficients indicate the standard deviation.Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. Column (4)-(6) uses ad sample from companies listed on HKSE. Industry classification consists of the following categories:Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel). 55

Figure 1.10: Industry Fixed Effects Relative to the Real Estate Industry

Note: Dependent variable: 1 if only Apple is chosen. The fixed effects and the interaction terms are estimated with these controls: firms characteristics, linear time trend, quarter fixed effects, slant gap and slant gap interaction terms.

Figure 1.11: Industry Fixed Effects Relative to the Real Estate Industry

Note: Dependent variable: 1 if both Apple and Oriental are chosen. The fixed effects and the interaction terms are estimated with these controls: firms characteristics, linear time trend, quarter fixed effects, slant gap and slant gap interaction terms. 56

Figure 1.12: Interaction terms between Industry and Slant Gap relative to the Real Estate Industry

Notes: Dependent variable: 1 if only Apple is chosen. The fixed effects and the interaction terms are estimated with these controls: firms characteristics, linear time trend, quarter fixed effects, slant gap and slant gap interaction terms.

Figure 1.13: Interaction terms between Industry and Slant Gap relative to the Real Estate Industry

Notes: Dependent variable: 1 if both Apple and Oriental are chosen. The fixed effects and the interaction terms are estimated with these controls: firms characteristics, linear time trend, quarter fixed effects, slant gap and slant gap interaction terms. 57

Figure 1.14: Interaction terms between industry fixed effects and firms’ connectivity for ads appearing on Apple only

Figure 1.15: Interaction terms between industry fixed effects and firms’ connectivity for ads appearing on both newspapers

Table 1.12: Average Ad Size by Industry

Industry Average Size Baby Products 0.7187500 Clothing & Accessories 0.6057348 Communications & Internet 1.0000000 Consumer Durable 0.8144531 Cosmetics & Skincare 0.5637097 Credit Card 0.9583333 Education 0.3004167 Financial Services 0.5474954 Food & Beverages 1.0087535 Government 0.3560606 Holidays & Travel 0.6102201 Household Products 0.8381757 Lifestyle 0.5310256 Motoring 0.6287538 Others 0.5096026 Pharmaceutical & Personal 0.5401542 Real Estate 0.6344787 Restaurants 0.6173077 Note: Unit of Size is number of page. The data comes from a smaller sample of old newspapers collected through visits to local public library in Hong Kong 60

1.12 Appendix

1.12.1 Sample of Apple Daily and Oriental Daily Headline during Occupy Central

Figure 1.16: Headline of Apple Daily on December 12, 2014. Translated as ”Do not forget the original intention. We will be back”

Figure 1.17: Headline of Oriental Daily on December 12, 2014. Translated as ”Financier Jimmy Lai Pan-Democratic politicians accept bribes. Occupy Central Schemers All Caught” 61

1.12.2 Endogenous Readers’ Choice of Newspaper

We extend the model in section 3 to incorporate endogenous readers’ choice of newspaper while holding newspaper’s per-period slant exogenous in this section. The goal is to illustrate the effect of newspaper slant at each period on firms’ ad location decision through a change in readership composition. This has the effect of changing firms’ economic benefit from advertising in each newspaper. I highlight the effect of an increase inslantgap on readers sorting. Consider an economy of which the population is composed of young and elderly. The fraction of young people is denoted as g. Young people are more pro-Democracy whereas elderly more pro-Beijing. We assume that young readers derive higher utility from Apple when Apple slants more to pro-Democracy in turbulent times. Formally, denote b and b − δ, δ > 0, as the utility for which young and old readers obtain from reading Apply Daily in stable periods respectively. Let a be the utility that elderly readers obtain from Oriental Daily, and a − δ be that of young readers. Furthermore, young (elderly) readers obtain an extra benefit equivalent to the slant gap △θ from Apple Daily (Oriental Daily) . Finally, there is an idiosyncratic utility εkfor reader to read newspaper , assumed to j k j εk distribute type-I extreme value. Individuals receive only an idiosyncratic benefit of 0 from not reading any newspaper. Given all these, the demand is:

exp(b + ∆θ) F = (1.11) ya 1 + exp(b + ∆θ) + exp(a − δ) exp(b − δ) F = (1.12) ea 1 + exp(b − δ) + exp(a + ∆θ) where Fya is the probability that young readers choose Apple. Turning to the firms, we assume that firms either sell products that only elderlyor young people would buy. The economic benefit of advertising on Apple for firms targeting elderly then depends on Apple’s ability to reach to the elderly: ee = (1− g) f (Fea)+νi, where

f is an increasing function of its argument and νi follows a uniform distribution [−δ, δ].

Similarly, the economic benefit for firms targeting young people would be ey = g f (Fya) + νi. ∗ As in section 3, elderly firms choose Apple Daily if their vi is above the threshold ve, and ∗ young firms vy. 62

The only difference from section 3 is that newspapers have to take into account of the in-and-out of young and elderly readers, and how that affect firms economic benefit ∂ ∂F | Fea | > | ya | △θ from advertising on their newspapers. Suppose ∂△θ ∂△θ such that an increase in has a larger effect in turning elderly reader away than drawing young readers. Denote the fraction of young firms as h in the market. If h > 0.5, then the aggregate economic benefit at Apple for elderly firm decreases more than that of young firms. Note that h can be a function of g: if the fraction of young people (g) in the economy is large, that the fraction of young firms can be large also. This means that there is a net decrease in expected demand for Apple Daily. When setting its ad price, Apple takes into account the lost of economic appeal and sets a lower price than it would have should there is no change in readership composition. The equilibrium ad share of Apple also falls. The main idea is that economic benefit at Apple can increase for some firms but decrease for others due to change in readership composition depending on the target audience of the advertisers. We expect greater degree of sorting of firms between the two newspapers in turbulent times. In presence of political pressure, we expect that drop in companies targeting Apple’s ad share decreases more for elderly firms than for young firms.

Endogenous Newspaper Slant

In this section, I allow the newspaper to choose how much slant to report when political events occur and show that the slant choice can be rationalized in a profit-maximizing framework. I focus on Apple Daily’s decision problem. The main trade-off that Apple Daily faces is the advertising revenue gained versus the readership lost for reporting in a more neutral stance. Readers are assumed to be aware of the occurrences of the political events and have a general sense of the intensity. They buy newspaper to get more in-depth stories and expect the slant implied in the stories to be consistent with the political ideology of the newspaper. If the news slant does not match the expectation, newspaper loses their good-will and future readership revenue. The assumption is different from Gentzkow and Shapiro (2006) in that Apple Daily wants to build a reputation as an anti-government voice, a provider of pro-Democracy information52.

52Gentzkow and Shapiro (2006) assume that media wants to be seen as unbiased and provide accurate 63

The key intuition is that when political events occur frequently, newspaper are also assessed frequently by the readers. Newspapers care about its reputation, and therefore report in accordance to readers’ expectation. Advertisers on the other hand, are made more aware of newspaper slant when reporting diverges, and this creates a negative advertising revenue impact on Apple Daily. To formalize this intuition, denote z = 1 when a political event occurs and z = 0 otherwise. We write the following as the readership revenue function which depends on the gap between newspaper slant and the political image θ¯:    r(θ − θ¯) if z = 1 r(θ) =   ¯r if z = 0 where θ¯ is assumed to be a constant. We assume that if θ − θ¯ < 0, then r < ¯r. Furthermore, when political events occur, Apple might draw new readers with its anti-government slant ∂r(θ) > in addition to be able to win good-will from its loyal readers so ∂θ 0 (The more pro-Democracy, the larger the θ). The implicit assumption here is that adopting a more pro-Democracy stance does not turn away loyal readers because they have already formed an expectation on Apple’s political stance. Advertising revenue depends on the readership and the instantaneous slant as laid out in the main text. Since Beijing-friendly firms avoid advertising on Apple regardless of slant, we can write the advertising revenue function as a(r(θ − θ¯), θ) when there is political event, and a constant a¯ when there is no political event. We assume that a¯ ≥ a(θ), which suggests that Apple’s advertising revenue is always higher in non-volatile times. So Apple’s profit-maximizing function in politically volatile times is given by: a(r(θ − θ¯), θ) + r(θ − θ¯). In equilibrium, the marginal cost must equal to the marginal benefit. [ ] ∂r(θ − θ¯) ∂a ∂a 1 + + = 0 ∂θ ∂r ∂θ

∂ θ−θ¯ r( ) ∂a As long as ∂θ is significantly larger than ∂θ , readership subscription reacts more than advertising revenue to slant, Apple Daily would prefer to report in more slant.

information 64

1.12.3 Finer Time Unit for Slant Gap

We chose the time unit of slant gap as quarterly instead of a finer unit such as monthly or weekly to balance the trade-off between having more words to choose from (which gives higher confidence of picking the most polarized words) and more data points of polarization (which gives higher statistical power). In using a finer time unit of analysis, we inevitably introduce more noise to the slant gap measure but we might be able to capture more accurate advertisers’ response to slant gap. Figure 1.18 shows the slant gap using monthly text data. The period-to-period fluctuation is high. Furthermore, the second half of 2014 does not register high slant gap, which is contrary to the consensual perception that newspapers became more polarized during umbrella movement. We are therefore hesitant in using finer time units than quarters in our slant gap measurement.

Figure 1.18: Slant Gap Using Monthly Data 65

1.12.4 Polarization of the Benchmark Language

To capture the degree of polarization of benchmark language overtime, we consider a simple measure inspired by Jensen et al (2012). For each period t, we take the top 100 phrases of length l ranked by χ2. We then compute the difference in the frequencies that the phrase p is used in Law and Politics and Wen wui Daily:

∑ ∑100 ϕ − t = ( fpll fplw) (1.13) l p

In absolute term, the measure is large when a certain phrase appears more frequently in one benchmark but not another. The measure ignores the length and number of articles that each benchmark has. This design is meant to address two issues: 1. We observe that the length of the articles tend to be longer in periods with many political events. 2. In addition, Wenwui started to post more than 1 editorial later in the sample period. This measure also disregard relative ranking of phrase p. This simple measure does not suffer from the finite-sample bias highlighted by Gentzkow et al (Working Paper) because with only two references, we do not calculate the correlation between the newspaper and the phrase frequency53. Figure 1.19 shows the evolution of the polarization measure over the sample period. There is a general downward trend, indicating that the absolute value of the measure became larger over time. This suggests that polarization of the benchmark language has worsened in our period, which reflected the general political environment. The decrease in especially in 2014. This matches the general sense of the growing polarization of the political climate. The Pearson correlation between the slant gap and the polarization measure is −0.3 suggesting a moderately strong correlation. This means that as the slant gap becomes large, frequency difference becomes more negative (the absolute value becomes larger). Since the benchmark language became more polarized over time, the slant gap used in the main text actually underestimate the degree of polarization over time. To assess the explanatory power of the polarization measure, we ran equation 1.8 again but replace slant gap and its interaction term with ϕt. The results are presented in

53Jensen et al (2012) calculate the correlation between the party of the speaker and the normalized frequency, but in any finite sample the correlation will be nonzero with positive probability, so the measure may imply some amount of polarization even when speech is unrelated to party. 66

Figure 1.19: A measure of polarization of word choices used by the benchmark

table 1.14. The coefficient on ϕt is positive and significant at 1% across the specifications. This suggests that the more negative the polarization measure, the less likely a firm will advertise on Apple Daily. The result suggests that the polarization measure, slant gap and the political events all have explanatory power on advertisers’ choice of newspaper. In our model, however, political events alone, does not drive advertisers away from Apple Daily; slant gap has to respond to political events, and the mechanism rests on advertisers responding to the news content. The rationale is that if newspapers do not slant more in respond to occurrences of political events, the newspaper will lose its reputation as a pro-Beijing or pro-Democracy paper which is costly. Once the political image of the newspaper is lost, advertisers no longer face the same decision problem and will not avoid a pro-Democracy newspaper. 67

Table 1.13: Results on Polarization and Interactions

Dependent variable: 1 if only Apple is chosen ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ϕtq 5.205e-05 5.626e-05 5.912e-05 0.0001 0.0001 0.0001 1.1e-05 1.11e-05 1.12e-05 2.88e-05 2.91e-05 2.92e-05 ∗ ∗∗∗ ∗∗ ∗∗ Connect·ϕtq 3.881e-05 3.375e-05 3.558e-05 -9.458e-05 -7.995e-05 -7.81e-05 2.19e-05 2.2e-05 2.2e-05 3.39e-05 3.42e-05 3.43e-05 ∗ ∗ ∗ Mainland ·ϕtq -1.045e-05 -1.776e-05 -1.777e-05 0.0002 0.0002 0.0002 4.8e-05 4.81e-05 4.82e-05 0.000 0.000 0.000 Foreign ·ϕtq 1.76e-05 2.217e-05 2.291e-05 4.909e-05 5.295e-05 5.466e-05 1.67e-05 1.68e-05 1.68e-05 5.36e-05 5.45e-05 5.46e-05

Dependent variable: 1 if both Apple and Oriental is chosen ∗ ∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ϕtq -4.527e-05 -5.106e-05 -5.433e-05 0.0003 0.0003 0.0003 2.67e-05 2.69e-05 2.7e-05 7.65e-05 7.82e-05 7.84e-05 ∗ ∗ ∗∗∗ ∗∗∗ ∗∗∗ Connect·ϕtq -9.8e-05 -9.717e-05 -9.659e-05 -0.0004 -0.0004 -0.0004 5.31e-05 5.32e-05 5.29e-05 8.55e-05 8.67e-05 8.59e-05 Mainland ·ϕtq -0.0001 -9.823e-05 -9.733e-05 -0.0002 -0.0002 -0.0002 0.000 0.000 0.000 0.000 0.000 0.000 ∗∗∗ ∗∗∗ ∗∗∗ Foreign ·ϕtq 0.0001 0.0001 0.0001 6.448e-05 2.921e-05 2.479e-05 3.99e-05 4.01e-05 3.99e-05 0.000 0.000 0.000 Firms Characteristics X X X X X X Industry FE X X X X X X Quarter FE X X X X X X Linear Time Trend X X X X X X IndustryQuarter FE X X X X Quadratic Time Trend X X Observations 206425 206425 206425 45036 45036 45036

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. The second row of the coefficients indicate the standard deviation.Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.001. Column (4)-(6) uses ad sample from companies listed on HKSE. Firms characteristics include political connectivity and country of origin (foreign, mainland). Industry classification consists of the following categories:Automobile, Baby products, Banking, Beauty, Computers, Education, Electronic Appliances, Fashion and Accessories, Food & Beverages, Insurance, Pharmaceutical, Real Estate /property, Restaurants, Retail, Sports, Telecommunication, Travel (airline, hotel). 68

1.12.5 Time Series of the Number of Pro-Democracy Phrases

Table 1.14: Number of Pro-Democracy Phrase over the sample period

Year Quarter Number 2010 Q1 50 Q2 55 Q3 54 Q4 56 2011 Q1 55 Q2 61 Q3 53 Q4 61 2012 Q1 55 Q2 63 Q3 58 Q4 57 2013 Q1 58 Q2 60 Q3 58 Q4 53 2014 Q1 60 Q2 54 Q3 58 Q4 58

Note: The number of pro-Democracy phrases is constantly more than 40, and fluctuates between 50 to 63. That means the ratio of pro-Democracy phrases is constantly over half (80 phrases used every period.) 69

1.12.6 Political Ads and Government Ads

Table 1.15: Number of political and government ad assignments in Apple and Oriental

Apple Oriental Pro-Beijing Political Ads 0 314 Pro-Democracy Political Ads 5 0 Government Ads 453 644 Chapter 2

Measuring Subjectivity in History Textbooks

2.1 Introduction

History textbooks provide a lens through which students view the nation’s past. Free of political considerations, history textbooks should provide an unbiased, objective and accurate account of the nation’s history. However, governments, especially that of authoritarian regimes, have strong incentives to use history textbooks ”as ideological tools to promote a certain belief system and legitimize an established political and social order.” (Apple and Christian-Smith (1991, 10)). As such, history textbooks can contain ideological content and are used to promote nationalistic sentiment. History textbooks are also arguably more formative than the mass media in shaping one’s support to the regime and easier to manipulate. Unlike mass media consumption, history textbooks are mandatory for students and students are incentivised to learn the material in order to do well in exams. Furthermore, history textbooks are studied at an age when values are most influenced1. These characteristics of history textbooks suggest that understanding how history is being portrayed, and quantify the degree of textbook subjectivity can provide an assessment of government’s intent in affecting their citizens’

1. Studies have suggested that political views are influenced most in early years. The most relevant theory in this respect, the impressionable years hypothesis, states that core attitudes, beliefs, and values crystallize during a period of great mental plasticity in early adulthood (the so-called impressionable years) and remain largely unaltered thereafter.

70 71 political view. This paper measures the degree of subjectivity in history textbooks by using tools developed in the field of sentiment analysis. By doing that it illustrates one linguistic channel through which persuasive communication manifests. Subjectivity in natural language processing refers to aspects of language used to express opinions and evaluations (Banfield 1982; Wiebe 1994). Subjective remarks can come in a variety of forms, including opinions, rants, allegations, accusations, suspicions, and speculation. Though subjectivity and bias can have similar meanings in some contexts, in this paper, we consider subjectivity as writer’s explicit emphasis on his/her views and feelings on the event and bias as factual but incomplete account of the event. In other words, an objective text can be completely biased. The main difference is that Subjective text may contain little ”hard” information such as facts and details of the state. While traditional economic models on persuasion emphasize on objectively useful information (Milgrom and Roberts 1986b; Dewatripont and Tirole 1999), the use of uninformative content in persuasion has been widely pointed out in psychology (Bless et al 1990, 1992; Batra & Stayman 1990) and has recently received attention in behavioral economics (Mullainathan et al (2008)). Studies in Psychology have also shown that the viewing of emotional phrases leads to psychological responses (Herbert et al, 2006; Larsen et al, 2003). We study high school history textbooks used in mainland China, Hong Kong and Taiwan. China observers have noted that Beijing uses ”history education as an instrument for the glorification of the party, for the consolidation of the PRC’s national identity, and for the justification of the political system of the CCP’s one party rule.” (Wang, 2008) The modern history portion is directly related to the legitimacy of the Chinese Communist Party, and the government has a strong incentive to manipulate the educational content to legitimize their ruling. Comparing textbooks between mainland, Hong Kong and Taiwan allows us to study and contrast government’s incentive in influencing the citizens’ political views. To put the degree of subjectivity in perspective, we also include in our analysis the translated version of ”Search of Modern China” (SMC) - a book on Chinese history written by the British-American scholar Jonathan Spence. SMC is commonly assigned as a reference in universities in the United States. In addition to the modern history textbooks, we also

1Yolanda Xue provides excellent research assistance. 72 analyzed two textbooks prior to major textbook reforms in the mainland and Taiwan to study the change in subjectivity over time. The interest in history textbook is not new. In particular, history textbooks in East Asia has been a source of international tension and domestic controversies. Previous studies have noted the divergence of historical account of world war II in textbooks used in East Asia (Sneider, 2013). Mainland Chinese textbooks are found to promote a nationalist view of the past while Japanese textbooks are relatively devoid of overt attempts to promote patriotism. Controversy over history textbooks also took place in Taiwan where critics of a proposed textbook reform said the changes ”China-centric” and deny the island its own perspective. While there exists several studies that measure slant in mass media (Gentzkow and Shapiro, 2010; Groseclose and Milyo, 2005 ), none has provided a quantitative measure on the subjectivity in education content. Previous studies on history textbooks (Ye(2016)2; Sneider, 2013 ) use qualitative methods and involve a certain degree of subjective interpretation. The quantitative approach used in this paper can mitigate the potential human bias in interpretation but has the limitation of only able to analyze text of the same language. By focusing on text subjectivity, we also bypass the need to find a language benchmark that is required for measuring bias. For example, in order to measure bias in newspapers in the U.S., we need to know how Democrats and Republicans talk and that serves as the benchmark. This benchmark is not necessarily easy to find in certain contexts such as ours. Although this paper focuses on the cross-region difference in subjectivity, the approach is also fruitful in analyzing changes of subjectivity over time. We use a number of metrics to measure degree of subjectivity. First, we count how often adjectives, adverbs and four-character idioms (Chengyu) are used relative to the character length. Four-character idioms are unique in East Asian languages and often carry strong philosophical connotation and convey a sense of lecturing. The motivation comes from the observation that provocative and inflammatory language are often aided by useof adjectives. Adjectives and adverbs are usually difficult to verify and quantify (what doesa ”brutal” war mean?). This means that it is difficult to dispute with the information sender. The sentiment analysis literature has also found that the presence of adjectives is strongly

2The study examines if there are opposing historical accounts in Chinese and Taiwanese textbooks. 73 predictive on whether an article belongs to an opinion piece (Hatzivassiloglou and Wiebe, Wiebe, Bruce, & O’Hara 1999; Bruce & Wiebe 2000). Second, we manually compile a list of important entities of political significance and count the number of times these entities are mentioned. Mao Zedong - the central figure of the Chinese Communist party from its founding to his death - is critical tothe CCP’s authority and presumably mentioned more in the mainland textbooks. Similarly, Chiang kai-shek - the Director-General of the Kuomintang (KMT) until his death in 1975 - is expected to be mentioned more in the Taiwanese textbook. To facilitate comparison across the textbooks, we calculate the ratio of pro-CCP to pro-KMT count. Third, we count the number of positive and negative words. The use of positive and negative words directly expresses the sender’s opinion on a certain historical episode or person. Several studies in Psychology have also found that pleasant and unpleasant adjectives are better memorized than neutral adjectives(Herbert et al, 2007; Kissler et al, 2009), suggesting that the use of emotional words could capture readers’ attention. The ruling regime has an incentive to emphasize the policy success and underplay the mistakes so the positive-to-negative ratio on specific historical episodes and time period is expected to vary on certain politically-sensitive episodes. The first three methods simply count number of times a phrase appears. To capture the ways important political entities are being described in the text in more granularity, we use word embedding to measure distance between these political entities and other adjectives. Word embeddings use numbers to represent words, and this representation allows us to measure the distance between important political entities such as ’Chinese Communist Party’ and other emotional phrases in different textbooks. The method originally introduced by Bengio et al in 2003, and the main idea is that words that tend to appear in similar contexts are likely to be related. To learn these word embedding, we use Word2vec (Mikolov et al (2012)), a state-of-art method in the field of natural language processing. to train the history textbooks. Adjective ratio and the ratio of pro-CCP to pro-KMT entities are both higher in the mainland textbooks than the textbooks in the other two regions. Mainland textbooks also have a much higher ratio of positive-to-negative words in post-1949 history. In addition, the pro-CCP political entities tend to have stronger association with positive words in 74 mainland textbooks. Overall, the result strongly suggests that the history textbooks in mainland China exhibit stronger degree of subjectivity across the three regions using our measures. We also found the adjective ratio is higher in the modern textbooks than the old textbooks in the mainland, and that the positive-to-negative words in post-1949 history registered a considerable increase over time also. In contrary, textbooks in Taiwan became less subjective over time. This paper contributes to the empirical persuasion literature conveniently summarized in DellaVigna and Gentzkow (2010), much of which has focused on the persuasive effects of mass media communications (Stromberg, 2004; DellaVigna and Kaplan, 2007; Bursztyn and Cantoni, 2012; Yanagizawa-Drott, 2014; DellaVigna et al., 2014) aside fro an important exception of Cantoni et al (2016) which documents evidence of the persuasive effect of politics textbooks in China. In addition, much of the focus of the empirical persuasion literature is paid to bias, which emphasizes factual but incomplete ”hard” information. Outside of bias, however, persuasion can take in the form of emotional appeal, of which the use of subjective language is arguably more effective. This paper also contributes tothe growing literature in using data-intensive text analysis techniques to elicit political content (Gentzkow and Shapiro 2010, Tumasjan et al 2010, Beauchamp, 2016). By pointing out a role of conveying subjective sentiment, it also expands the literature on persuasion techniques much of which focus on informative communication( Kamenica and Gentzkow, 2011), with a few exception that emphasizes on behavioral limitation (Mullainathan et al, 2010). Recent work has analyzed authoritarian regimes’ incentive to homogenize the population by shaping the views of their citizens (Alesina and Reich, 2013), to which we contribute empirical evidence that is manifested in educational material.

2.2 History Textbooks in Mainland China, Hong Kong and Taiwan

Our textbook sample consists of 3 history textbooks from the mainland, 2 from Hong Kong and 3 from Taiwan. They are all used in local high schools as of 2016. Students are generally between the age of 15 and 17. We sought to select textbooks that are most widely used in local high schools. To digitalize the textbooks into text files, the under- graduate research assistant uses a combination of voice typing and manual entry. For all 75 textbooks, we only use the history after the Opium War (A.D. 1842). In addition, we only use the main text portion and ignore all text in captions, supplementary information and appendix. History textbooks in mainland China are regulated by the central government. The three approved versions are: Renmin, Renjiao and Xueli. Localities can choose one of the three approved versions for their high schools 3. There are 3 books in each version and each book has different emphasis. The first book focuses on the historical eventsand present them in a generally chronological order. The second book focuses on the economic development in the corresponding time period, whereas the third book focuses on culture and technological progresses. The textbooks cover material that are not directly related to China in certain parts. For example, the different political systems used in the world as well as western literature and science advancement at the concurrent time period are mentioned. We opt not to include these discussion in our analysis. . To facilitate comparison across regions, we use portions of the textbooks that are directly relevant to China only. The history curriculum is separated into mandatory and elective portion. We use the mandatory potion only. In both Hong Kong and Taiwan, the education bureau issues guidance for private publishers, and the schools have the discretion to decide which version to use. In Taiwan, there are 7 popular versions, and our sample consists of three of the most widely used: Kangzi, Nane, and Lungtun. Kangzi is most popular textbook and used by 45% of the schools, followed by Nane (17%) and LungTun (16%) 4. The history curriculum is organized into four sections with a respective focus on Taiwan History, Chinese History, Ancient World History and Modern World History5 and the events are presented in chronological order. We focus on the second volume on Chinese History only. Our Hong Kong textbooks sample includes two versions from the following publishers: Manhattan Press, and the Modern Educational Research Society Limited. The two versions are among the most commonly used in Hong Kong 6. Manhatten Press separates

3http://godfreyxu.github.io/2013/01/30/high-school-history-textbook-version-of-provinces-and-cities. html 4The figure comes from a study that sampled 313 schools in 2009 (Mao, 2013) 5https://www.sanmin.com.tw/learning/public/data/course 6There is no formal study that examines the fraction of schools using which textbook versions, but Manhattan Press, Modern Educational, Hong Kong Educational and Ling Kee are considered to be the four most widely used versions: http://www.com.cuhk.edu.hk/ubeat_past/051170/64.htm 76 the curriculum into six books while Modern Education has 4 books, but the content and emphasis is very comparable between the two. To measure over-time change in subjectivity, we also analyze a copy of old history textbook used in Taiwan between 1983 and 1999 before the textbook market opened up for private publishers. The old textbook was standardized so all high schools in Taiwan adopted this version. We do not have the mainland history textbook for the corresponding years, but we have a version used in 2003, just before a major reform took place in the mainland in 2004. The old textbooks allow us to measure the change in degree of subjectivity prior and after reform in the respective regions. In addition to the textbooks used in the three regions, we also obtain a translated copy of the book ”The Search for Modern China” (SMC) written by the American Scholar Jonathan Spence7. The book has been assigned as a reference in many Chinese history classes in American universities. Since this book is relatively free of a political agenda, it serves a good comparison for the textbooks in the three regions. SMC covers many areas of Chinese history from 1685 to 1989 and in much more depth and breadth than our sample textbooks.

2.3 Descriptive Statistics

As part of preprocessing, we removed all punctuation and stop words from the text and segment the sentences into words and phrases. Unlike English, there is no white space in Chinese so we need to segment the text into phrases. To do that, we use a Python library Jieba, which segments Chinese sentences into individual phrases by finding the most probable combination based on the word frequency. Table 2.1 provides the descriptive statistics in each version. The first column indicates the total character count in each version (by combining individual books for each version). The total character count varies considerably across regions though not as much within. Textbooks in Hong Kong are on average the longest, followed by that of mainland and Taiwan. SMC is considerably longer than our textbook sample. Column (2) - (4) indicate the number of bigram, trigram and quadgrams that Jieba can detect. For all

7The book is translated by a Taiwanese 77 versions, the number of trigrams and quadgrams identified in the dictionary are considerably less than the bigrams. In all Mainland textbooks, the number of gradgrams exceeds the number of trigrams while the reverse is true for Hong Kong and Taiwan textbooks. Many quadgrams are 4-character idioms, and they are usually strong in subjective interpretation. This provides the first indication that mainland textbooks contain more subjective elements 8. Column (5) and (6) shows the character count on pre- and post- 1949 content. Character count in Post-1949 history is less than that of pre-1949 in all textbooks. On average, the fraction of post-40 to pre-49 character count is lowest in the Taiwan textbooks and highest for mainland textbooks.

2.3.1 Major Historical Episodes

We focus on several major historical episodes before and after 1949 to examine whether the treatment varies across textbooks on individual event9. We do not include SMC in this analysis because its writing styles makes it difficult to separate out the specific portion that is exclusively relevant to the major historical episodes. The top panel in table 2.2 shows the number of words and the ratio to the character length in the four major historical episodes prior to 1949: Foreign Invasion (1842 - ), XinHai Revolution (1911), Second Sino-Japanese War (1937-45), Chinese Civil War (1945 - 1949). Foreign Invasion refers to all of the wars involving a foreign power since the Opium War but excluding the second Sino-Japanese War. Xinhai Revolution includes description on Sun yut-chan, and all immediate events leading up to the Xinhai Revolution. The second Sino-Japanese War refers to the military conflict between China and Japan from 1937 to 1945. Finally, the Chinese Civil war includes all events surround the conflict between the Chinese Communist Party and Kuomintang after the surrender of Japan in 1945, but not any previous conflicts. We avoid capturing stand-alone sentences that may be relevant to the event, and capture a block of text whenever possible. Foreign Invasion has the largest word count among the pre-1949 individual events for 5 of the 8 versions in all regions. In Mainland Renjiao’s version, Civil war has the

8In the sentiment analysis literature, Pang et al. (2002) also find evidence that higher-order n-grams are useful features in predicting opinionated piece. They report that unigrams outperform bigrams when determining determine whether a movie review is positive or negative 9Since mainland versions have different emphasis in each of the book, we exclude text in the bookson economics and culture and technology and only use the text in the general history. 78 largest word count, whereas in the Taiwan Kangxi and Lungtun version Xinhai Revolution has the largest word count. The word count varies significantly across versions within the same region. For example, Mainland’s Xueli only has 618 characters on the Sino-Japan war whereas Renmin has 3293 characters. In terms of length ratio, the three Taiwan versions have the highest ratio for the Civil War. On Foreign Invasion, the two Hong Kong versions have the highest ratio across all 3 regions The bottom panel of table 2.2 indicates the character count and ratio on three post-1949 historical episodes: the Great Leap Forward (1958 - 61), Cultural Revolution (1966-76), the Reform and Open Policy (1976 - current, encompassing the Tiananmen Square Protests). In terms of character count, the reform and open policy is the largest in all versions. Within each event, the Great Leap Forward and the Cultural Revolution have the highest ratio in the two Hong Kong versions. The Great Leap Forward has the lowest ratio in mainland’s Renjiao, whereas the Cultural Revolution has the lowest ratio in mainland Xueli. For the Reform and Open policy, Taiwan Lungtun has the highest ratio and the lowest goes to Taiwan’s Nane, suggesting considerable variation in event emphasis within Taiwanese versions. The Tiananmen Square Protests of 1989 is noticeably absent in all mainland textbooks but are mentioned in all versions in both Hong Kong and Taiwan.

2.4 Subjectivity Analysis

We construct a number of metrics to measure subjectivity. First, we count the number of times an adjective is used and calculate the ratio of adjectives to the total word count. Work on subjectivity detection revealed a high correlation between the presence of adjectives and sentence subjectivity (Hatzivassiloglou and Wiebe, 2000). Second, we count the number of times an important political entity such as Mao Zedong and Chang Kai-shek is mentioned. Knowing the affectual aspect of the opinions like ”happiness” or ”mood” is also important in understanding government’s objective. Third, we calculate the positive to negative word ratio using a sentiment dictionary. Forth, we use word embedding to calculate distance between relevant politically-relevant entities and other adjectives. 79

2.4.1 Adjectives and Adverbs

To determine whether a phrase is an adjective, we write a script in Python to automatically look up all the phrases in the Chinese dictionary pre-installed on a Mac notebook10. A phrase is considered as an adjective if the word definition contains the character (English: adj.) and a four-character idiom if the character is present, and an adverb if is present. I compile by hand a list of stop words that do not have any discriminate power11. Table 2.3 lists the top 5 bigram and quadgrtam adjectives and their respective frequencies in each textbook version. (Act independently and of one’s own initiative) is the most common quadgram for all 3 mainland versions. All popular quadgrams in mainland versions have a positive connotation. For the two Hong Kong versions, (Domestic trouble and foreign invasion) is among the top 5 quadgrams. (Revolution/Revolutionary) and (Economy/Economic) are the two most popular bigrams in versions of all regions. Perhaps surprisingly, (Democracy/Democratic) is the third most commonly used bigrams in all mainland versions12. Figure 2.4 and 2.5 plot the adjective and adverb ratio for each textbook version. The three mainland versions have the highest adjective ratio among all textbooks. On average, the adjective ratio in mainland textbooks is more than 1% higher than Hong Kong and Taiwan textbooks. As a reference, SMC has a similar adjective ratio as the Hong Kong and Taiwan textbooks. The adverb ratio is slightly higher for Taiwan textbooks relative to Hong Kong and mainland versions, but is similar with that of SMC. The ratio is similar between Hong Kong and mainland versions. The difference in adverb ratio is smaller than that of the adjective ratio, with the difference between the largest and smallest about 0.7%. Figure 2.6 and figure 2.7 plot the adjective ratio for the 4 pre-1949, and 3 post-1949 historical episodes in each version. In the civil war portion, the adjective ratio of the three mainland versions stands out compare to versions in other regions. Across all versions, the adjective ratio is highest for Xinhai Revolution compared with the other pre-1949 episodes.

10There is a function in the Jieba module that returns the part of speech tagging but the result is not very extensive 11Many single character are classified as adjective even though the most common usage of the character does not serve as an adjective. For example, the character can mean ‘correct‘ (adjective) in some context but in most situations, the character is used to to mean ”is”. 12The bigram adjectives are commonly regarded as nouns as well. This can make our result difficult to interpret. We also use jieba’s part-of-spech tagging method to determine adjectives. 80

Among post-1949 historical episodes, mainland versions have a higher adjective ratio in the Reform and Open portion. Adjective ratio is also higher in the Great Leap Forward portion in mainland versions though the variation is also higher among different versions and across regions13. We next examine the adjective ratio in pre and post-1949 content. Figure 2.8 plots the results. The mainland versions in figure 2.8 is calculated using only first book because the first book concerns exclusively with the general historical flow14 ofevents . While all textbook versions have an increase in adjective ratio from pre- to post-194, the average increase in adjective ratio in mainland textbooks from pre-1949 to post-1949 is more pronounced when compared with Hong Kong and Taiwan versions. In comparison, the adjective ratio of SMC also has a mild increase from pre- to post-49 period.

2.4.2 Political Entities

The number of times a political entity is mentioned is a direct indication of the political entity’s importance in history, but it can also suggest the political entity’s significance specifically to the regime. For example, ”Mao Zedong”, as well as ”Chinese Commu- nist Party” (CCP) are expected to be more frequently cited in mainland textbooks because of their significance in mainland politics. In contrary, ”Kuomintang” (KMT) is expected to mention more in Taiwan textbooks15. Here we count the number of times some polarizing figures are mentioned in each version. I then calculate the ratio of pro-CCP to pro-KMT entities mention by considering the following list of politically-relevant individuals and entities: Mao Zedong, the Chinese Communist Party, Karl Marx, Chang Kai-Shek16 , Sun Yat-Sen, the Kuomintang. The first three are considered to be pro-CCP and the last three pro-KMT. The ratio is presented in figure 2.3. Mainland versions have the three highest pro-CCP to pro-KMT ratio among all versions. In particular, the Renmin version mention a pro-CCP entity in about 3.5 to 1

13Due to the low total character count in Great Leap in mainland versions, the ratio is not the best measure of subjectivity. 14In figure 2.9, we plot the adjective ratio using all 3 versions. The adjective ratio in pre-49 period is considerably higher for the mainland versions but there is still a sharp increase from pre- to post-1949. 15By counting the number of times a think tank is cited in the newspaper, Groseclose and Jeffrey Milyo (2005) have considered a similar idea to capture liberal bias in the media. 16The Taiwanese versions address Chang Kai-Shek using his adopted name ”Zhongzheng”, and in many cases in the old Taiwanese textbook, the honorific ”Chairman Chang” is used. 81 ratio, while SMC has a 2 to 1 ratio. Hong Kong and Taiwan textbooks have a ratio between 1 to 2 with Taiwan version on average the lowest among the three regions. Table 2.7 presents the number of mentions for each of the 6 individuals and entities. The contemporary mainland versions mention Chang less than 15 times but Mao is mentioned more than 45 times. The opposite is true in Taiwan versions. Marx is mentioned less than 5 times in all Taiwan versions but at least 19 times in the mainland versions. Hong Kong versions mention Marx 9 times on average. The number of mentions of CCP is comparable between mainland and Taiwan versions, with the average being 113 times in mainland versions and 103 times in Taiwan versions. Sun also appears to receive relatively equal emphasis between mainland and Taiwan versions. Both region mention Sun about 30 times on average. In comparison, Hong Kong versions mention Sun 65 times. Interestingly, KMT is mentioned more times in mainland versions that in Taiwan versions. On average, mainland versions mention KMT 96 times while Taiwan versions 47 times. Taken together, the count suggest that Marx, Mao and Chang Kai-Shek receive more drastic difference in treatment between mainland and the Taiwan, whereas CCP, KMT and Sun are relatively comparable.

2.4.3 Text Polarity

Determining the valence or polarity of a piece of text, whether it is positive, negative, or neutral is an important part of subjectivity analysis. It is also commonly referred as ”sentiment analysis”. Studies in Marketing have found that text valence in product reviews can change initial attitudes (Lee et al., 2012), suggesting the potential importance of emotional words in shaping political attitude. To determine the polarity of the sentiment implied in the text, we assign a phase as either positive or negative using a binary sentiment Chinese dictionary available online17. The dictionary consists of 4570 positive and 4374 negative phrases. Phrases not found in the dictionaries are ignored. Table 2.8 reports the number of positive, negative phrase, and the positive-to-negative ratio for each version. Mainland textbooks have the highest higher positive to negative words ratio, followed by Hong Kong and Taiwan. In particular,

17The dictionaries are available at: https://github.com/Fansion/CCUED/blob/master/README.md. The dictionary has the drawback of not having many idioms. 82

Renmin’s ratio is about 2 times higher than the ratio of SMC, which has the lowest ratio of all. On average, the positive-to-negative ratio is slightly higher in Hong Kong than in Taiwan. Figure 2.12 and 2.13 plot the results of the positive to negative word ratio in the major historical episodes. The ratio is similar across all versions for Foreign Invasion. There is more variation between versions for the other three events. The variation within mainland versions is particularly large for Xinhai Revolution: Renjiao has the lowest ratio whereas Renmin has the highest. Within Hong Kong versions, the ratio vary much less across all 4 episodes compare to the other two regions. Shifting to post-1949 episodes, the ratio is considerably higher in all three mainland versions for the Reform and Open section: it is about 2 times higher than Hong Kong and about 3 times than Taiwan. The ratio is similar across versions and regions in Cultural Revolution and Great Leap Forward. Figure 2.10 plots the pre- and post-1949 positive to negative word ratio for each version18. Mainland versions stand out as having the largest increases from pre- to post- 1949. The average ratio for Hong Kong versions has a slight increase, while Taiwan has a small decrease on average. SMC has a roughly similar ratio pre- and post-1949. The result strongly supports the notion that the mainland textbooks portray a more positive image of the regime compared with the textbooks of the other two regions.

2.4.4 Word Embedding

Cutting the text by time period and historical episode is a crude way to assess how adjectives are associated with the regime. To understand how a political entity is being portrayed in more granularity, we can study the word embedding of the political entity. In linguistics, word embeddings aim at quantifying semantic similarities between words based on their distributional properties. The basic idea is that words with similar distributions of other words have similar meanings. Mathematically, a word embedding W : words → RN is a parameterized function mapping words in some language to high-dimensional vectors. We can capture the sentiment surrounding the political entities of interest by looking at the distance with their surrounding adjectives in the high-dimensional embedding space trained

18Figure 2.11 plots the same ratio using all 3 texbooks. The increase from pre- to post-49 is consistent for mainland versions. 83 by the history textbooks from each region. We use Word2vec (Mikolov et al, 2013), which is a group of related models that are used to produce word embeddings, to consider the associations between entities of interest (such as ’Taiwan’, ’Chinese Communist Party’) and words with emotional connotation. Word2vec is a particularly computationally-efficient predictive model for learning word embeddings from raw text. One type of model within Word2vec is called Continuous Bag of Words (CBOW) (Mikolov et al, 2013). It can be used to predict co-occurrence relation- ships using the conditional probability of observing the target word given the input context words. Context words is represented by multiple words for a given target word. For example, Word2vec treats ”The”, ”cat”, ”over”, ”the”, ”puddle” as context and and predicts the target word ”jumped”. The training goal of the CBOW model is to arrive at vector representations of words that best predict the target word. Formally the objective function is given by: ∑ 1 T Jθ = log p(w |w ) (2.1) T t C t=1

where θ represents all the variables we optimize. wt denotes the target word, and wC denotes the context words. t denotes the training step. Denote V as the vocabulary in the text. The conditional probability of the target word can be represented by a softmax function, which uses a neural network structure to learn the parameters:

T exp(u vw ) ( | ) = ∑ t c (2.2) p wt wC T ∈ ( ) wi V exp uwvwc uw and and vware two representations of the word w. uw comes from rows of the input to hidden weight matrix, and vw comes from columns of hidden to output matrix. The inner product T computes the log-probability of word , which we normalize by the sum of ut vwc wc the log-probabilities of all words.The goal of the learning is to then learn the weights in the input to hidden, and hidden to output matrix19. (A more detailed explanation of the CBOW model and neural network learning can be found in Rong (2014)). To calculate the distance between words, we calculate the cosine distance between the word vectors in the high-dimensional embedding space. The cosine distance between

19This can be achieved by the gradient descent method using a random initialization. 84 vector A and B is defined as: A · B (2.3) ∥A∥ ∥B∥ The resulting similarity ranges from -1 meaning exactly opposite, to 1 meaning exactly the same, with 0 indicating orthogonality. We can then examine the distance between important political entities to other To maximize statistical power in explaining cross-region difference, we combine individual textbook versions into a corpus for each region and train our model based on region-specific text. We trained each model with embedding dimension of 500 andwith context size of 6 using a python library gensim, which has built in a CBOW model. We search the top 20 closest adjectives around those words in each embedding space, and calculate the ratio of positive to negative surrounding adjectives. Table 2.9 shows that Mainland versions have higher positive-to-negative ratio on Mao, Sun and Marx. Taiwan versions have higher positive-to-negative ratio only on Chiang and Hong Kong versions have averagely mild positive-to-negative ratio on all 6 entities. More examples of words and distance can be found in Table 2.4, 2.5 and 2.6. We can visualize each embedding space by using the t-Distributed Stochastic Neighbor Embedding (t-SNE) (Maaten et al, 2008). This is a dimensions reduction technique that is particularly well suited for the visualization of high-dimensional datasets. Fig 2.19, Fig 2.20 and Fig 2.21 showed the the relationship between Mao, CCP, Marx, Chang, KMT, and Sun in each space. In Fig 2.19, we can see that Marx to CCP and Mao is like Sun to Chiang and KMT, but they are far from each other. Fig 2.20 shows three clusters, CCP and Marx as one, Mao as another itself and KMT, Chiang and Sun as the third. Fig 2.21 shows that Mao, CPP, KMT, Chiang and Sun are clustered and Marx alone is far away.

2.5 Prior and After Textbooks Reform

History textbooks in both mainland China and Taiwan have undergone significant revision in the last 2 decades. Before 1996, Taiwan’s national textbooks were designed and published by the National Institution for Compilation and Translation under the Ministry of Education (MOE) and classroom use of nationally standardized textbooks was required. 85

Since 1996, however, the MOE has opened up the textbook market to competition among private publishing companies. The opening of the textbook market suggests that textbooks that do not follow strict government stance can be published. Textbooks will try to give a more complete evaluation of history and we therefore expect that the subjectivity of language decreases. On the other hand, a nationwide education reform was undertaken by the Chinese central government in 2001, and the first entering cohorts to study under the new curriculum were students entering high school in 2004. The reform has an explicit goal to shape students’ political views and students are taught to ”love socialism” and be more patriotic. 20. Subjectivity is therefore expected to increase for mainland textbooks. We first examine some summary statistics. The bottom 2 rows oftable 2.1 print the book length and the portion allocated to pre-1949 and post-1949 history for the old versions. The old versions of both mainland and Taiwan are longer than the modern versions. The post-1949 portion of the old Taiwanese (OT) version is significantly smaller compared to the modern version with only one paragraph allocated to post-1949 mainland China. The old mainland (OM) version also allocates a larger fraction to pre-1949 history compared to the modern versions: the ratio of word count of pre- to post-1949 is over 2 in the old version and less than 1.5 in the modern versions on average. As shown in figure 2.4, the aggregate adjective ratio increases slightly from 0.52 to 0.55 for the mainland textbook. The adjective ratio of OT is about the same with the average of the modern versions. Figure 2.14 plots the adjective ratio of the old version with the average of the modern versions for each historical episode. Since the text size of individual event is rather small and changes considerably before and after event, we have to warrant caution in drawing conclusions on the change in subjectivity for each event. Nevertheless, the adjective ratio appears to be relatively similar in the four pre-1949 episodes. Among the post-1949 episodes, there is a noticable increase for the Great Leap Forward but decreases in Cultural Revolution in mainland textbooks. For the Taiwan versions, the adjective ratio

20From Appendix C of Cantoni et al(2016), in the Ministry of Education’s “Framework for Basic Education Reform” (2001), the ministry delineates the motivation and the objectives for the future curricular reforms. The previous basic educational curriculum, it is said, cannot satisfy the needs of development in this new age. Thus, a new curriculum should meet the following objectives (in the order of appearance in the original document): it should reflect the times, and make students patriotic, communitarian, [and] love socialism. Students should inherit and carry forward the great traditions of the Chinese nation and its revolution; and be equipped with an awareness of the legal system under a socialist democracy. The new curriculum should promote compliance with national laws and with societal ethics, and gradually form in students a correct worldview, a correct view of life, and a correct value system. 86 appear to stay relatively similar or decrease slightly for the four pre-1949 episodes. Figure 2.16 (2.17) plots the adjective ratio (positive to negative phrase ratio) in pre- and post-1949 for the old version and the average of the modern versions by region. In terms of adjective ratio, the increase in OT from pre- to post-1949 is much larger than that in modern Taiwan, and the increase is comparable with OM and the average of modern mainland versions. In terms of the positive to negative phrase ratio, the average of modern mainland versions has a much higher ratio OM, suggesting that there is more emphasis on the positive aspects since 1949. For Taiwan, Since the text size on post-1949 in OT is very small, the low positive to negative phrase ratio can be misleading. CCP is mentioned 152 times in OM and 74 times in OT. Interestingly, the KMT is mentioned more in mainland than Taiwan: 116 times in OM and 32 times in OT. Chang is mentioned about equal number of times with 41 at OM and 42 at OT, as well as Sun, mentioned 37 and 39 times in OM and OT respectively. After the reform, Chang is mentioned in a 2 to 1 ratio in modern Taiwan and mainland textbooks. In contrast, Mao is mentioned 62 times in OM and only 2 times in OT, and Marx is mentioned 14 times OM and 1 time in OT. After the reform, Mao is mentioned in higher frequency but Marx has remained scarcely mentioned in modern Taiwan versions. This suggests that Mao and Marx are more polarizing than Chang and Sun before the reform. After the reform, Chang became more polarizing but Mao less so.

2.6 Writing Style

Can the result be attributed to difference in writing styles across regions? That is, do mainland Chinese tend to use more adjectives in their writing? While the difference in speech and informal writing is large across the regions, formal writing is usually not significantly different. However, to evaluate this possibility, we measure the adjective ratio of the news content in popular newspapers of the 3 regions. We select 3 representative newspapers that have reasonably large circulation and focus on the section on economy only since the economy section should be relatively free of subjective interpretation. We avoid choosing a newspaper that operates under direct government influences such as the state-owned newspaper. The newspaper sample are Shanghai Morning Post () circulated 87 mainly in Shanghai, Oriental Daily () in Hong Kong and Liberty Times () in Taiwan. The news content is available from Wisers, a Hong Kong private company. We use all news reports in year 201621. The result is summarized in table 2.10. The newspaper adjective ratio is higher than that in the history textbooks in all 3 regions, suggesting that the writing style of newspapers tend to be very different from that of history textbooks. In addition, Taiwan newspaper has the highest adjective ratio, followed by that of Mainland and Hong Kong. The result suggests that the use of adjective does not tend to be higher in the mainland.

2.7 Conclusion

This paper provides the first quantitative analysis on history textbooks by focusing on text subjectivity. We have shown that mainland Chinese textbooks exhibit a stronger degree of subjectivity by the following measures: the adjective ratio, the ratio of pro- CCP to pro-KMT entities mentions, the ratio of positive to negative phrase, and distance between political entities and adjectives. The result broadly confirms the notion that the authoritarian government has a stronger incentive to present subjective content. In addition, by studying the textbook before and after reform in mainland China and Taiwan, we show that looser government control of education content led to a more objective evaluation of history. This paper focuses on a neglected aspect of persuasive communication by exam- ining subjectivity rather than bias. The methods shown in this paper can also be used to analyze the subjectivity of other politically-relevant content such as speeches of politician and judges. There are important caveats, however. First, our approach is comprised of several individual measure, so it could be difficult to access the overall subjectivity when different measures give different results. Second, our method only considers adjective, and does not capture other subjective elements such as subjective nouns. Previous study in Natural Language Processing has found that certain extraction pattern can be leveraged to identify subjective words (for example, the pattern “expressed object” often extracts subjective nouns, such as “concern”, “hope”, and “support”), and some bootstrapping al-

21We exclude all non-news articles including columns that tend to represent one’s opinion. 88 gorithms can automatically generate these extraction patterns(Riloff et al., 2003). We have not considered this possible extension in this paper. How opinions can be persuaded by language subjectivity is unanswered in this paper. Persuasion can take different channels. One channel is through careful considerations by the information recipients. Attitude changes are then determined by the issue-related arguments of the message claims. This channel is analyzed extensively in the Economics literature (Crawford and Sobel 1982, Dewatripont and Tirole 2005 , Dziuda 2011, Kartik 2009 among others). Another channel is referred to as coarse thinking (Mullainathan et al. 2008) and explains how persuasion can take place in absence of information transfer. While previous study in Marketing (Chevalier and Mayzlin 2006) have shown that text with strong valence can change product purchase decision, the next step is to explicitly examine how subjectivity influences opinion through different channels of persuasions. Chapter 2, in full, is currently being prepared for submission for publication of the material. Lam, Onyi; Lin, Eddie. The dissertation author was the primary investigator and author of this material. 89