Topic Analysis of Climate Change Coverage in the UK

Transcription

1 Topic Analysis of Climate Change Coverage in the UK Graham Beattie University of Pittsburgh September 1, 2017 Abstract The UK newspaper market is dominated by large national newspapers that compete for the same set of readers. I use topic modelling techniques to analyze how newspapers compete with each other through their coverage of climate change. Coverage across newspapers is highly correlated and driven by news events, but newspaper differentiate themselves by having areas of focus and increasing coverage more significantly when events related to those areas arise 1 Introduction For the majority of the public, the media is the main source of information about climate change and other environmental issues (Leiserowitz et al., 2011). The media is far from homogenous, however, outlets can present different perspectives on the news, so the information that one consumer receives may differ significantly from the information others do. Media outlets have an incentive to differentiate themselves to appeal to consumers, but this differentiation is constrained, by the fact that, each outlet has the same set of news events to report on; thus it is limited in its ability to fabricate, alter, or hide facts. If an outlet presents coverage that is incomplete or inaccurate, it may face a reputation penalty if these deficiencies are revealed. Based on this set of constraints, competing outlets should cover roughly the same set of topics, but each outlet will devote more coverage to topics that are of interest to the consumers it is trying to appeal to. In this paper, I study this relationship in the context of the national newspaper market in the UK by analyzing the complete set of articles mentioning climate change or global warming in seven national newspapers during the period , and further identifying 1

2 the commonalities and differences in their coverage. I use Latent Dirichlet Allocation, an unsupervised machine learning technique, to classify articles according to the topics they cover and then examine how these topics vary across newspapers and over time. Latent Dirichlet Allocation identifies 20 topics, which fit into 3 major categories: climate change, hard news, and soft news. Although the distribution of coverage over these categories is relatively similar across newspapers, the overall quantity of coverage varies greatly. More left-leaning newspapers, such as the Guardian, devote more coverage to all topics related to climate change than more right-leaning newspapers, such as the Telegraph, a gap which is increasing over time. Further, quality newspapers devote more coverage to these topics than tabloids. The coverage of many of these topics is highly correlated across newspapers. When one newspaper devotes more space than usual to a particular topic, it is likely that other newspapers do so as well. This indicates that coverage of these topics is driven by real world events, which create commonalities in coverage across newspapers. However, newspapers are able to differentiate their coverage of these topics. During periods when all newspapers increase their coverage of these news-driven topics, I find that newspapers that cover these topics most in other periods respond by increasing their coverage by a greater amount. For example, if a newspaper generally devotes more space to international climate agreements, and an event related to international climate agreements occurs, this newspaper will produce more additional coverage of the event than newspapers that generally devote less space to international climate agreements. The results suggest that while newspapers cover the same events related to climate change, they differentiate themselves by the amount of coverage they provide for each event. I check this categorization of news-driven events by analyzing how newspapers respond to general elections in the US and UK, providing examples of newspapers differentiating themselves in response to a major event. All newspapers increase their coverage of politics and climate change during a major election, but newspapers that cover politics and climate change the most in non-election years increase their coverage by the largest amount during election periods. As a second example of newspaper responses to external events, I analyze how newspaper coverage of climate change responds to weather. I find that newspapers cover climate change more when the weather is unseasonably hot, cold, or rainy. 2

3 2 Background This paper contributes to the understanding of competition between newspapers. The newspaper market is competitive, and consumers have different tastes, which gives newspapers an incentive to differentiate themselves to attract a particular subset of readers. However, newspapers receive homogeneous real world events to report on. One avenue of differentiation occurs when newspapers serve different geographic markets, as newspapers can provide coverage of local events that are of interest to consumers in one area but not in others. This model does not apply in the UK, however, as the market is dominated by national newspapers. British newspapers must differentiate themselves through their coverage of the same set of available news events. Previous work has studied this type of differentiation, which is driven primarily by consumer preferences. Consumers tend to prefer news which conforms to their biases because of a taste for perspectives that agree with their own (Mullainathan and Shleifer, 2005), a perception that news which conforms to their biases is more accurate (Gentzkow and Shapiro, 2006), or a belief that endorsements and opinions from sources that share their biases are more useful (Chan and Suen, 2008). Given that consumers prefer newspapers that conform to their biases, a profit-maximizing newspaper has an incentive to choose bias to attract the most consumers. Gentzkow and Shapiro (2010) show empirically that newspapers choose their political slant to conform to the political orientation of their potential consumer base. This paper focuses on the mechanism by which newspapers produce differentiated coverage, given that the events available to report on are identical for all newspapers. Mullainathan and Shleifer (2005) model one such mechanism: all newspapers receive the same information, but newspapers can selectively choose which of this information to relay to consumers. For example, one newspaper could focus on coverage of the risks of climate change, and another newspaper could provide more information about the costs of mitigation. This model creates a version of Hotelling product competition among newspapers, where consumers choose the coverage that most closely conforms to their biases, and newspapers choose their coverage to attract the largest possible share of consumers. If newspapers use this strategy, they will produce strongly differentiated coverage, and a consumer would have to read multiple newspapers to receive unbiased information. If, on the other hand, consumers have a taste for accurate and unbiased information from each newspaper they consume, competition between newspapers should create alignment in their reporting. Gentzkow and Shapiro (2006) present a model in which a newspaper has a choice between 3

4 making an unbiased report or a report that appeals to consumers biases. The threat of having biases, omissions, or inaccuracies revealed by competitors enforces less biased reporting. If a newspaper distorts its coverage, another newspaper can reveal this and inflict reputational damage on the newspaper that distorted the news. Under this model, there will be commonalities in the coverage of competitive newspapers. This paper analyzes these mechanisms from an empirical perspective. I evaluate how newspapers that have the same real world events available to report on differentiate their coverage. In this respect, it is related to the work of Puglisi and Snyder (2008), who show that in the US, newspapers that support one party tend to report more on scandals involving politicians of the opposite party, and Larcinese et al. (2011), who show that newspapers that support a given party tend to report positive economic news more if that party is in power and negative economic news more if the opposite party is in power. The setting of this analysis coverage of climate change provides additional motivation. The science communication literature has documented extensively (Boykoff and Boykoff, 2004, 2007; Boykoff and Roberts, 2007) that while media coverage of climate change has increased dramatically since the 1980 s, coverage often contains non-scientific analysis, opinion, and climate skepticism, which can lead to confusion and misinformation for consumers. I contribute to this literature by analysing the factors that trigger coverage of climate change and how newspapers respond to those triggers. 3 The British newspaper market 3.1 Overview The newspaper market in the United Kingdom, particularly in England, is dominated by newspapers with national circulation. This feature makes it a suitable context for this analysis. In North America, many major newspapers serve regions surrounding a major city and do not compete directly with newspapers in other regions. For example, comparing coverage in the Los Angeles Times and the Boston Globe presents challenges because if the Los Angeles Times covers a topic more, it is difficult to disentangle whether this is an editorial choice about which topics to emphasize or simply because the topic affects the west coast more than the east coast. In contrast, national newspapers in the UK compete for an identical geographical market, so differences in coverage are not a result of local stories. There are two main classes of newspaper in the UK: the quality press, or broadsheets, and the popular press, or tabloids. The quality press is aimed towards a primarily middle 4

5 class audience and contains lengthier articles, more objective coverage, and more coverage of international issues. The popular press is targeted towards a more working class audience and contains more opinions and editorials, and more coverage of sports, entertainment and scandals (Boykoff and Mansfield, 2008). Newspapers also differentiate themselves within these categories, particularly by using political orientation. Among the quality press, the Guardian is generally considered to be left-wing, the Indepedent centre-left, the Times centre-right, and the Telegraph right-wing. The Financial Times focuses on business news, and its political slant tends to be centrist. Among the popular press, the Mirror is left-wing, while the Daily Mail, Sun and Express are right-wing. These orientations can be seen in the newspapers endorsements, of the leftwing Labour Party, the right-wing Conservative Party and UK Independence Party, or the centrist Liberal Democrats, in recent general elections, shown in Table 1, as well as in the political positions of readers, shown in Table 2. Table 2 contains results from the British Election Survey, which asks respondents about their newspaper reading habits and about their political positions. The results show that the readers of the Guardian, Independent, and Mirror are more liberal, more likely to vote Labour, and more likely to prioritize the environment over economic growth. It also shows that readers of quality newspapers are more likely to support environmental protection than readers of tabloids. 3.2 Coverage of climate change Studies the communications literature have analyzed coverage of climate change in UK newspapers. Boykoff (2007) examines quality newspapers during the period , and shows that, in comparison with major US newspapers, reporting in major newspapers in the UK reflects the scientific consensus relatively accurately. On the other hand, over the same time period, tabloid newspapers provided coverage which was skeptical about climate change and emphasized the economic costs of mitigation efforts (Boykoff and Mansfield, 2008). Painter and Gavin (2015) analyze a later period, , and find that the prevalence of climate change skepticism increased over this period, particularly in more conservative newspapers. Quality and tabloid newspapers on the right-wing of the political spectrum provided less coverage of climate change overall, were more skeptical in the coverage they did provide, and often discussed climate change in opinion pieces by conservative political columnists instead of in scientifically-oriented articles. The analysis in this paper differs from the analysis in these previous studies in two ways. Previous work relies on subjective reading of a selection of articles, while I use an 5

6 Table 1: UK general election endorsements by newspaper Year Newspaper Guardian Labour Labour Labour Labour/Lib Labour Dem Independent Labour Labour/Lib Dem Lib Dem Labour/Lib Dem Conservative/ Lib Dem Times Conservative Labour Labour Conservative Conservative/ Lib Dem Telegraph Conservative Conservative Conservative Conservative Conservative Financial Times Labour Labour Labour Conservative Conservative/ Lib Dem Mirror Labour Labour Labour Labour Labour Sun Labour Labour Labour Conservative Conservative Daily Mail Conservative Conservative Conservative Conservative Conservative Express Conservative Labour Conservative Conservative UKIP Sources: general-election-newspaper-support, 417fa1a2-ab60-11d9-893c-00000e2511c8.html Newspaper Table 2: Circulation and positions of readers, 2015 Circ. Av. Orientation Voted Voted Voted (Feb. Conser- Labour Liberal 2016) (0=Lib., vative Democrat 10=Cons.) Voted UKIP Av. Priority (0=Econ., 10=Env.) Guardian 158, % 50% 6% 2% 6.5 Independent 53, % 39% 10% 5% 6.0 Times 391, % 15% 9% 6% 5.8 Telegraph 459, % 10% 8% 9% 5.4 Mirror 721, % 49% 2% 7% 5.7 Sun 1,669, % 19% 1% 11% 5.2 Daily Mail 1,462, % 15% 5% 15% 5.3 Express 404, % 12% 7% 13% 5.7 Sources: Audit Bureau of Circulation and British Election Study,

7 objective measure applied to the complete set of articles mentioning climate change or global warming, allowing for a more detailed categorization of coverage. I also focus on topic analysis, rather than sentiment analysis, which provides data about the contexts in which a newspaper discusses climate change, but not about what perspective the newspaper takes. 4 Quantifying coverage of climate change The dataset consists of all articles in major British national newspapers published during the period which mention global warming or climate change. The set of newspapers include four quality newspapers, the Guardian, Independent, Times, and Telegraph and three tabloid newspapers the Sun, Mirror, and Daily Mail. Along with the Financial Times and the Express these are the most widely circulating newspapers in the UK. The Financial Times is excluded from the analysis because its focus on business news may bias topic analysis, which identifies which topics newspapers choose to report on. The Express is excluded because of significant gaps in the available archived coverage. I combine the Sunday editions of these newspapers with their weekday counterparts, by including the Observer (sister paper to the Guardian), Independent on Sunday, Sunday Times, Sunday Telegraph, Sun on Sunday, Sunday Mirror, and Mail on Sunday. 4.1 Latent Dirichlet Allocation The complete dataset contains approximately 72,000 articles, which are categorized using the Latent Dirichlet Allocation (LDA) topic modelling method, an application of unsupervised machine learning using a bag-of-words model of text. Bag-of-words models describe text as a distribution of words or phrases. When constructing a text, the topics that an author wants to discuss and the perspective she wants to take determine the distribution that she draws from. LDA identifies words that frequently appear together in one article as being part of the same topic. These topics are used to classify articles; an article which contains many words which are indicative of a particular topic is assigned a high score for that topic. Thus, LDA classifies articles which have many words in common as being about the same topic, and also generates a list of indicative that the researcher can use to name the topic. Prior to implementation of LDA, each article is represented as a multinomial distribution of words, so that only the frequency, and not the order, of words within an article is relevant. To ensure that results are not driven by grammar, these distributions are modified by re- 7

8 moving stop-words and stemming the remaining words. Stop-words are extremely common words, such as pronouns, articles, and auxiliary verbs, which appear in many articles and are not indicative of the topics the article discusses. Stemming removes grammatical indicators such as plurals and verb tenses, so that different grammatical forms of the same word, such as emission and emissions, are treated as identical. The LDA model (Blei et al., 2003) treats a text as a latent distribution of topics, each of which is in turn a distribution of words. Specifically, the LDA process with k topics assumes that each article i is generated using the following process: 1. Draw a k-dimensional distribution of topics θ i from a Dirichlet distribution parameterized by the vector α. 2. For each word position j in article i of length N i, draw a topic z ij from a multinomial distribution parameterized by θ i. 3. For each word position j in article i of length N i, choose a word w ij from the multinomial distribution characterizing topic z ij, parameterized by a row of the matrix β (which contains one row vector for each topic). The probability of a set of M articles occurring is: M N i p(θ i α) p(z ij θ i )p(w ij z ij, β) dθ i. z ij i=1 j=1 Finding optimal values for α and β using maximum likelihood estimation is computationally infeasible, but the model can be estimated using Gibbs sampling (Griffiths and Steyvers, 2004). This generates a distribution of topics θ i for each article. The researcher chooses the number of topics k, but does not choose the topics themselves. The topics chosen by LDA consist of words which tend to appear together within articles. It is up to the researcher to describe the generated topics in order to give meaning to the output. 4.2 Results of topic modelling Table 3 shows the results of Latent Dirichlet Allocation using 20 topics. The list of most frequently occurring words for a given topic z are the words that are assigned the highest value of p(w z). If an article was written exclusively about topic z, these are the words that 8

9 would occur most frequently in that article. The topic label is a subjective choice based on the most frequently occurring words. I further group the topics into three broad categories: climate change, hard news, and soft news. Table 3: Results of LDA estimation: Topics used in articles mentioning climate change or global warming Topic label Most frequently occurring words Category Wildlife species year bird plant fish sea tree Climate change Climate science scientist climate year ice temperature research rise Climate change Emissions car carbon emission green transport air year Climate change Inter. agree. country emission world carbon global climate china Climate change Energy energy power nuclear carbon wind renewable coal Climate change Consequences water flood year weather temperature record Climate change Energy costs energy tax government year bill cost pay Climate change Sustainability work business company make person environment project Climate change US politics obama president state america bush house war Hard news Law police claim report group public person government Hard news Business oil company year bank market price investment Hard news Global change person world make live time political change Hard news Education child person live school year woman family Hard news UK politics party minister labour cameron government brown leader Hard news Aus. politics time block publish australia government minister abbott Hard news Agriculture food world grow farmer year crop person Hard news Travel city day build year world island travel Soft news The arts film book live show year write play review Soft news Society service royal london community sir prince john Soft news Miscellaneous make time day person year back live N/A The topics chosen by Latent Dirichlet Allocation include some which are directly related to the environment or climate change, such as wildlife, climate science, emissions, international agreements, energy, consequences, energy costs, and sustainability. Other topics, classified as hard news topics, are less directly related to climate change, but are common subjects of news coverage which are often discussed in conjunction with it. Examples of this type of topic include US politics, law, business, global change, education, UK politics, Australian politics, and agriculture. A third category of topic includes more soft news topics such as travel, the arts, and society. The final topic is difficult to attach a meaningful topic label to, so it is described as miscellaneous 9

10 and is left uncategorized. While the topics categorized as hard news or soft news are not directly related to climate change, articles must contain a mention of climate change or global warming to appear in the dataset. LDA does not assign a single topic to each article; it assigns to each article a probability for each of the 20 topics. For example, if for a given article, climate science is assigned a weight of 0.75, UK politics is assigned a weight of 0.25, and all other topics are assigned a probability of 0, then the article contains words indicative of both climate science and UK politics, but the proportion of climate science words is higher. This could indicate that most of the article covers climate science and some covers UK politics, or that the article covers the intersection of the two topics, focussing more on climate science issues. For the purpose of analysis, it is useful to classify articles by primary topic, which is the topic to which LDA assigns the highest probability for each article. The mean probability attached to a primary topic is It is also useful to normalize by the number of articles in a given newspaper. For each topic-newspaper-month triple, the primary topic share variable is the share of articles in that newspaper and month with that topic as the primary topic. The three major categories have very different interpretations. Topics in the climate change category contain words directly related to the environment and climate. A newspaper that dedicates a significant amount of coverage to this type of topic is likely to be providing its readers with a significant amount of information about climate issues. On the other hand, words indicative of hard news topics are not directly related to environmental issues. Although they must mention global warming or climate change to be included, articles written using mainly these topics are not likely to discuss climate-related issues in detail. If a newspaper produces many articles of this type, it might indicate that hard news events they are reporting on involve climate change for example, if climate change becomes an election issue. It might also indicate that the newspaper focusses on the climate-related aspects of hard news events for example, how an election outcome could affect climate change. Similarly, topics in the soft news category do not attach a high probability to environment or climate related words, but these articles must contain global warming or climate change in order to be included in the data set. Articles that place a high probability on these topics could be about a soft news event which involves climate change, such as a celebrity discussing it. It is important to note that LDA uses topic analysis, rather than sentiment analysis. It does not identify the perspective that a newspaper takes. For example, if an article endorses the Conservative Party in a general election, topic analysis will distinguish it from 10

11 articles which do not talk about UK politics, but will not distinguish it from an article endorsing the Labour Party. Similarly, topic analysis will identify articles which discuss climate science, but will not distinguish articles which are more environmental from articles which are more skeptical. Therefore this analysis does not address climate change skepticism directly. However, the quantity of coverage of climate change is still an informative measure of the environmental nature of coverage. If a newspaper reports extensively on climate issues, readers will be more informed and concerned about climate change, and if a newspaper downplays climate change by covering it less, readers will be less informed and may be less concerned. From a demand perspective, more coverage of climate change may provide a signal that it is a priority for readers, so that the newspaper has an incentive to provide more coverage. 5 Differentiation among newspapers 5.1 Type and quantity of coverage Latent Dirichlet Allocation uses the newspaper article as the main unit of analysis, allowing for the comparison of coverage of climate change between newspapers and over time. Table 4 shows a cross-newspaper comparison at the category level. Newspaper Table 4: Topic comparison across newspapers Climate change topics Hard news topics Soft news topics Average share per article As primary topic per 1000 Average share As primary topic per 1000 Average share As primary topic per 1000 Guardian Independent Times Telegraph Mirror Daily Mail Sun The first, third and fifth columns of Table 4 show the average share of each article mentioning climate change that is devoted to climate change, hard news, and soft news topics respectively. These shares are similar across newspapers, although a slightly greater share is 11

12 devoted to hard news in quality newspapers, particularly the more left-leaning Guardian and Independent, and a corresponding higher share is devoted to soft news in tabloid newspapers. Since tabloids devote more space to soft news topics in general, it is unsurprising that this holds true on the subset of articles considered in the analysis. Overall, conditional on writing an article mentioning climate change or global warming, the distribution of topics used in the article is similar in each newspaper. The second, fourth, and sixth columns of Table 4 show a much greater differentiation between newspapers. These columns show the number of articles with climate change, hard news, and soft news as the primary topic per 1000 articles that each newspaper prints. The share of all articles mentioning climate change and falling into any of the three categories of topics is much greater in quality newspapers than in tabloids. Within quality newspapers, the more liberal Guardian discusses climate change much more frequently than the more conservative Times and Telegraph, with the Independent in between. The one newspaper which is an exception to these rules is the Daily Mail, which devotes a greater share of its coverage to climate change than most other tabloids. Previous work (Boykoff and Mansfield, 2008; Painter and Gavin, 2015) shows that the coverage the Daily Mail does provide is more skeptical than other tabloids, so this anomaly does not necessarily imply that the Daily Mail provides readers with more information about the scientific consensus about climate change. Taken together, these results suggest that British newspapers differentiate themselves using the quantity of coverage of climate change, rather the category-level composition of their coverage. The share of articles mentioning climate change ranges from less than 0.2% of all articles in the Sun to almost 2% of all articles in the Guardian. Figures 1 and 2 illustrate how coverage evolves over the course of the sample period in each newspaper. Each plot shows the number of articles per 1000 which mention climate change or global warming. Among the quality newspapers, the difference between newspapers has been increasing over time. The Times and Telegraph have tracked each other closely, while the Independent and particularly the Guardian have separated themselves from the two more right-leaning newspapers and devote a higher proportion of their coverage to climate issues. Figure 2 shows that the anomalous Daily Mail coverage primarily occurred prior to After that point, the coverage of the Daily Mail converged towards the coverage of the other tabloids, and, although it is higher than the Sun and Mirror, in recent years it is comparable to the Times and Telegraph. 12

13 Figure 1: Coverage of climate change in quality newspapers 5.2 Correlation of coverage across newspapers The detailed topic-level classification allows for a more in-depth analysis of the differentiation between newspapers. If different newspapers provide the same distribution of topics at any point in time, one would expect that the share of coverage of a given topic in one newspaper could be predicted by the share of coverage of that topic in other newspapers. Table 5 shows tests of this hypothesis. For each topic listed in Table 5, the coefficient is the value of β 1 from the following regression: PrimShare nt = β 0 + β 1 PrimShare n,t + φ n + ψ t + ɛ nt. where PrimShare nt is the share of all articles not only those mentioning climate change in newspaper n and month t with a given topic as the primary topic, PrimShare n,t is the share of articles in all other newspapers with that topic as the primary topic, and φ n and ψ t are newspaper and month fixed effects. 13

14 Figure 2: Coverage of climate change in tabloid newspapers The results in Table 5 vary significantly across topics. Some topics have significant positive coefficients. Coverage of these topics is highly correlated across newspapers; when one newspaper devotes an unusually high share of coverage to them, other newspapers tend to as well. Topics of this type include climate science, emissions, international agreements, energy, consequences, energy costs, US politics, law, education, UK politics, agriculture, and the arts. Although these results do not address the causal relationship, the most reasonable hypothesis, which will be explored in greater depth in Section 6, is that this correlation is news-driven real world events are being reported on by multiple newspapers. For example, an election in which environment and climate issues are a topic of debate would drive coverage of politics and climate change, and the release of a report by the Intergovernmental Panel on Climate Change would drive coverage of climate science and the consequences of climate change. Other topics are uncorrelated or even negatively correlated across newspapers. Topics of this type include wildlife, sustainability, business, global change, Australian politics, travel, and society. When one newspaper covers one of these topics, other newspapers are no more likely to. This suggests that coverage of these topics is not driven by real world events, and 14

15 Table 5: Correlation between primary topics Topic Coefficient of primary topic share on share in other newspapers Topic Coefficient of primary topic share on share in other newspapers Wildlife Business 0.201* (0.151) (0.113) Climate science 0.687*** Global change (0.110) (0.124) Emissions 0.750*** Education 0.402*** (0.148) (0.109) Inter. agree *** UK politics 0.890*** (0.160) (0.106) Energy 0.612*** Aus. politics (0.075) (0.147) Consequences 0.868*** Agriculture 0.468*** (0.161) (0.094) Energy costs 0.887*** Travel (0.087) (0.130) Sustainability -1.09*** The arts 0.543*** (0.261) (0.124) US politics 0.617*** Society (0.099) (0.107) Law 0.832*** Miscellaneous 0.370** (0.158) (0.168) Each coefficient is from a separate regression with own newspaper primary topic share as dependent variable and other newspapers primary topic share as independent variable, along with newspaper and year fixed effects Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 that newspapers can use their coverage of these topics to differentiate themselves from their competitors Differentiation in news-driven topics Newspapers are able to differentiate themselves with their coverage of the set of topics which are not news-driven. They can use coverage, or absence of coverage, of these topics to appeal 1 Australian politics is an unexpected member of the group of non-news driven topics. This anomaly is primarily the result of the Guardian covering Australian politics and climate change far more than other UK newspapers. 15

16 to consumers or advertisers or to support the agendas of journalists or management. If a topic is news-driven, and all newspapers choose their coverage to respond to real-world events, they may still be able to differentiate themselves using the magnitude of their response to these events. For example, if there is an election in which environment and climate issues are prominent, all newspapers will increase their coverage of politics and climate change, but more environmental newspapers, such as the Guardian, may increase their coverage by more. This can be tested using the following specification for each of the news-driven topics: PrimShare nt = β 0 + β 1 PrimShare n,t + β 2 PrimShare n,t PrimShare n, t + φ n + ψ t + ɛ nt. where PrimShare nt is the proportion of articles in newspaper n in month t which have the topic as the primary topic and φ n and ψ t are newspaper and year fixed effects. PrimShare n,t is the normalized share of coverage in other newspapers and can be seen as an indicator that real world events are driving coverage of the topic. This variable is normalized for each newspaper, to avoid bias generated by the result that if newspaper n contains relatively more coverage of a particular topic, then omitting that newspaper lowers PrimShare n,t by relatively more. If β 2 is positive, then newspapers which tend to have more coverage of a particular topic, and thus have a higher value for PrimShare n, t, increase their coverage more in response to news events. Table 6 shows the results of this specification for each of the topics shown to be newsdriven in Table 5. With the exception of three topics, climate science, agriculture, and the arts, the coefficient for the interaction effect between coverage of the topic in other newspapers and coverage of the topic in other time periods in the same newspaper is significantly positive. When coverage of a news-driven topic increases in all newspapers, it increases more in newspapers which tend to cover that topic more. This provides further insight into coverage of news-driven topics. Differences in coverage in these topics is not simply a level effect, where some newspapers always provide a fixed amount more coverage of a given topic. Instead, when an event happens which drives coverage of these topics, newspapers which specialize in coverage of these topics respond more to the event. 5.4 Discussion Taken together, the results presented in this section provide a picture of how national newspapers in the UK report on climate change. On a broad scale, the composition of coverage 16

17 Table 6: Coverage of news-driven topics Climate science Emissions Inter. agree. Energy Conseq. Energy costs PrimShare i, n,t 0.011*** *** 0.004* (0.003) (0.006) (0.008) (0.002) (0.005) (0.002) PrimShare i, n,t *** 0.712*** 0.169*** 0.495*** 0.373*** PrimShare i,n, t (0.071) (0.126) (0.102) (0.047) (0.170) (0.057) Newspaper FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Observations R-squared US Law Education UK Agriculture The arts politics politics PrimShare i, n,t 0.009*** 0.007*** *** 0.007*** (0.002) (0.003) (0.002) (0.003) (0.002) (0.003) PrimShare i, n,t 0.230*** 0.422*** 0.385*** 0.484*** PrimShare i,n, t (0.064) (0.099) (0.126) (0.065) (0.057) (0.053) Newspaper FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Observations R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 of climate change between topics related directly to climate change, hard news relating to climate change, and soft news relating to climate change is fairly similar across newspapers. However, quality newspapers cover each of these categories more than tabloid newspapers, and left-leaning newspapers cover each more than right-leaning newspapers, a difference which is increasing over time. These categories can be further divided into topics, each of which represent one specific subject relating to climate change. Some of these topics are news-driven newspapers coverage of these topics track each other closely over time. Others are less news-driven coverage in each newspaper evolves independently of other newspapers. Newspapers are still able to differentiate themselves using coverage of news-driven topics; newspapers which cover these topics more in general respond more to the factors which drive coverage of these topics. These results shed light on the models of competition and differentiation in the newspaper market. Newspapers, either because of competition or journalistic integrity, do not ignore topics that other newspapers cover. In this respect, the reporting that different newspapers 17

18 provide is correlated, and consumers receive coverage of broadly similar topics regardless of the newspaper that they consume. On the other hand, newspapers do differ in the amount of the coverage that they allocate to a given news topic, and these allocations are consistent over time within a newspaper. In this respect, reporting varies across newspapers, and consumers who read different newspapers will receive differing amounts of coverage on each topic, and thus may receive very different information about which issues are highest priority. This result complements the models of Mullainathan and Shleifer (2005) and Gentzkow and Shapiro (2006). Newspapers do have preferences over topics, which may be demanddriven. However, they are not able to ignore a topic if it is receiving considerable coverage in other newspapers. Doing so might lower their reputation as a reliable source of information. Therefore, they maintain their preferences for topics by covering their preferred topics extensively when they are relevant, while still providing some coverage of all relevant topics. 6 Timing of coverage of climate change The previous section analyzed the commonalities and differences in coverage of climate change between newspapers. When producing coverage, all newspapers face the same set of inputs, namely real-world events. This is particularly true for national newspapers, which serve the same geographical market, so do not have different local events to report on. For this reason, the high level of correlation in coverage is very intuitive. To further explore the mechanics underlying this relationship, I analyze newspapers responses to two very different types of news: elections, which are major and relatively rare events; and weather, which is high-frequency and often minor news. 6.1 Elections This categorization of some topics as news-driven assumes that the correlation in coverage identified in the previous section is driven by news events, such as national elections, an assumption which can be tested. The sample period contains two UK general elections, in May 2010 and May 2015, and two US presidential elections, in November 2008 and November During the 2010 UK election all three major parties listed environmental and climate change priorities in their platforms and discussed them during campaigns. Polls indicated that voters were split about which of the three major parties had the best environmental policies (Rootes and Carter, 2010), so coverage of environmental issues was not a strongly 18

19 partisan topic. Similarly, in 2015, all three parties expressed concern over environmental issues, signing a joint pledge to address climate change prior to the election. It also became an occasional campaign issue, as the traditionally more environmentally-oriented parties, Labour and Liberal Democrat, criticized the incumbent Conservatives for their handling of environment and climate issues. 2 Political positions on environmental and climate issues are more divided in the US. In the 2008 and 2012 elections, parties and candidates differentiated themselves using their support for offshore oil drilling, subsidies for alternative energy sources, and participation in international climate treaties (Bomberg and Super, 2009). In general, the Republican Party is in favour of less regulation or taxation of environmental goods, while the Democratic Party is more in favour of environmental protection. This difference extends to climate science, as many prominent Republicans express skepticism about the scientific consensus about climate change. Each of the elections during the sample period should increase the prevalence of coverage of politics (US or UK) and climate change. Elections bring politics to the forefront of public perception and media coverage, so it is intuitive that newspapers would increase their coverage of all political topics, including the relationship between politics and climate change. This is illustrated in the first and third columns of Table 7. These columns show regressions of the share of articles in which UK or US politics is the primary topic (per 1000 articles written in a year) on election dummies representing months before, during and after an election, as well as newspaper and year fixed effects. Relative to the coverage the rest of the year, coverage of UK politics and climate change is significantly higher in the month of an election and the month immediately before. Coverage of US politics and climate change spikes during the month of the election, and again two months later when the new administration takes office. The second and fourth columns of Table 7 address a more substantive question. In these columns, the election dummies are weighted by the primary topic share of the politics topics in non-election years. These specifications also include newspaper fixed effects, so they test whether the difference between newspapers in coverage of politics and climate change is a simply a difference in levels, or if newspapers which tend to cover these topics more respond more to elections. The results show that the latter hypothesis is true, which is consistent with the findings in the previous section. During elections, politics and climate change is a 2 Source: 19

20 Table 7: Coverage of politics and climate change during elections UK election primary topic share UK election primary topic share US election primary topic share US election primary topic share Election in month t 0.127*** *** (0.035) (0.024) (0.026) (0.022) Election in month t *** * (0.029) (0.021) (0.007) (0.010) Election in month t (0.008) (0.008) (0.007) (0.008) Election in month t (0.012) (0.012) (0.007) (0.006) Election in month t (0.008) (0.009) (0.009) (0.009) Election in month t *** (0.012) (0.009) (0.014) (0.015) Election in month t (0.009) (0.010) (0.008) (0.010) Election in month t 2.66*** 2.48** non-election prim. topic share (0.672) (0.985) Election in month t *** non-election prim. topic share (0.654) Election in month t * non-election prim. topic share (0.161) (0.150) Election in month t *** non-election prim. topic share (0.314) (0.189) Election in month t non-election prim. topic share (0.186) (0.259) Election in month t ** 1.28*** non-election prim. topic share (0.245) (0.375) Election in month t non-election prim. topic share (0.228) (0.411) Newspaper FE Yes Yes Yes Yes Year FE Yes Yes Yes Yes Observations R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 20

21 relevant news topic, so all newspapers increase their coverage. However, newspapers do still differentiate themselves, as newspapers such as the Guardian and the Independent which cover politics and climate change more also increase their coverage by the largest amount during elections. This is consistent with the results in Table 6, and validates the assumption that news events drive the correlations between newspaper coverage and the differential responses. 6.2 Weather Elections are rare, major events. This type of event attracts a large amount of media coverage, and this coverage spills over to related topics, such as politics and climate change. Analyzing elections does not show whether coverage of climate change is sensitive to minor day-to-day events or, alternatively, if it is only disrupted systematically across newspapers by major events. To analyze this, I use monthly weather data from Heathrow Airport provided by the UK Met Office. These data include average daily high, average daily low, and total rainfall. I also use the midpoint between the average daily high and the average daily low as a measure of the average temperature. The variable of interest for generating newspaper coverage is unusual weather, as a very mild winter is more newsworthy than an average summer, even if the average summer is warmer. To measure this, I demean each monthly temperature and rainfall observation using the 50-year average values for that month of the year. Table 8 shows the results of regressions of measures of media coverage on these variables. The dependent variable in the first two column of Table 8 is the share of articles which have one of the topics in the climate change category as opposed to hard news or soft news as the primary topic. This share is calculated by dividing the number of articles in a given newspaper and month by the total number of thousands of articles in that newspaper during that year. The first column shows that weather that is hotter and wetter than normal leads to more coverage of climate change. The second column includes a dummy for a month being hotter than the 50 year average, and an interaction term between this dummy and the difference the temperature and the 50 year average temperature for that month of the year. The results show a significant negative coefficient for the demeaned average temperature and a significantly larger in magnitude positive coefficient for the interaction between demeaned average temperature and the hotter than average dummy. This suggests that unusual temperatures, either hotter or colder than average, increase the amount of coverage of climate change. 21

22 Table 8: Coverage of climate change and weather Climate change primary topic share Climate change primary topic share Climate science primary topic share Emissions primary topic share Conseq. primary topic share Temp.- av ** *** *** temp. ( C) (0.009) (0.011) (0.003) (0.002) (0.002) Hotter than * average (0.011) (0.005) (0.005) (0.007) Hotter*(Temp *** *** 0.023*** av. temp.) ( C) (0.022) (0.004) (0.003) (0.005) Rainfall - av *** 0.011*** 0.001** 0.002*** 0.002*** rainfall (cm) (0.003) (0.003) (0.001) (0.001) (0.001) Newpaper FE Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Observations R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 The final 3 columns of Table 8 break down the climate change category into some of its component topics. The results show that coverage of climate science responds to rainfall but not to temperature, and that coverage of emissions responds to rainfall and hotter, but not colder, weather. The share of coverage that is devoted to the consequences of climate change is higher in unusually hot, cold, or wet weather. This last result is intuitive given that weather and temperature are among the words that Latent Direchlet Allocation assigns the highest probability to for the consequences topic. However, since an article has to include one of the phrases global warming or climate change this does not simply mean that unusual weather increases mentions of the weather. Instead, it shows that unusual weather increases the share of articles which discuss climate change in the context of the weather. This set of results sheds light on the direction of public engagement with environmental and climate issues. As climate change creates unusual weather patterns, coverage of climate change increases. Applying a demand-driven model of media coverage suggests that this weather drives public interest in climate change. Further, applying a persuasive model of media coverage suggests that if unusual weather increases coverage, this will increase public awareness and interest in climate issues. Assuming that this interest is also manifested in 22

23 concern and willingness to adopt climate-friendly policies, the results suggest that if the effects of climate change become more prevalent, media coverage is one mechanism that can lead to increased public support for mitigation policies. 7 Conclusion Newspapers that serve the same geographical market, such as national newspapers in the UK, have the same set of news events to report on, yet they turn such news into differentiated reporting. This paper sheds some light on the mechanisms that they use to do this, with a particular application to climate change. I look at the complete set of articles mentioning global warming or climate change in seven national newspapers in the UK over the period and apply Latent Dirichlet Allocation to classify each article by the topics it covers. I find that the distribution of coverage across broad categories climate change, hard news, and soft news is similar across newspapers, but the total volume of coverage that newspapers devote to climate change varies. Left-leaning newspapers provide more coverage than right-leaning newspapers, and quality newspapers provide more coverage than tabloids. The difference between coverage in left-leaning newspapers and right-leaning newspapers increased dramatically over the period Many of the topics identified by Latent Dirichlet Allocation are news-driven. Coverage of these topics is strongly correlated across newspapers, as newspapers tend to discuss these topics simultaneously, in response to news events. Nevertheless, newspapers are able to differentiate their responses. If a newspaper usually covers a given topic more, they respond more to an event that drives coverage of that topic. Competing newspapers cover a similar set of topics, but choose which topics to emphasize and which to downplay in order to differentiate themselves from their competition. References D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. The Journal of Machine Learning Research, 3: , E. Bomberg and B. Super. The 2008 us presidential election: Obama and the environment. Environmental Politics, 18(3): ,