Why Is All COVID-19 News Bad News?

NBER WORKING PAPER SERIES

WHY IS ALL COVID-19 NEWS BAD NEWS?

Bruce Sacerdote

Ranjan Sehgal

Molly Cook

Working Paper 28110

http://www.nber.org/papers/w28110

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

November 2020

We thank Max Grozovsky and Nashe Mutenda for superb research assistance. The views

expressed herein are those of the authors and do not necessarily reflect the views of the National

Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been

peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies

official NBER publications.

of text, not to exceed two paragraphs, may be quoted without explicit permission provided that

full credit, including © notice, is given to the source.

Why Is All COVID-19 News Bad News?

Bruce Sacerdote, Ranjan Sehgal, and Molly Cook

NBER Working Paper No. 28110

November 2020

JEL No. I12,J01

ABSTRACT

We analyze the tone of COVID-19 related English-language news articles written since January

1, 2020. Ninety one percent of stories by U.S. major media outlets are negative in tone versus

fifty four percent for non-U.S. major sources and sixty five percent for scientific journals. The

negativity of the U.S. major media is notable even in areas with positive scientific developments

including school re-openings and vaccine trials. Media negativity is unresponsive to changing

trends in new COVID-19 cases or the political leanings of the audience. U.S. major media readers

strongly prefer negative stories about COVID-19, and negative stories in general. Stories of

increasing COVID-19 cases outnumber stories of decreasing cases by a factor of 5.5 even during

periods when new cases are declining. Among U.S. major media outlets, stories discussing

President Donald Trump and hydroxychloroquine are more numerous than all stories combined

that cover companies and individual researchers working on COVID-19 vaccines.

Bruce Sacerdote

6106 Rockefeller Hall

Department of Economics

Dartmouth College

Hanover, NH 03755-3514

and NBER

[email protected]

Ranjan Sehgal

6106 Rockefeller

Dartmouth College

Hanover, NH 03755

[email protected]

Molly Cook

Brown University

69 Brown Street

Providence, RI 02912

[email protected]

Introduction

On February 18

, the Oxford Mail published a story that Professor Sarah Gilbert and her

colleagues at Oxford’s Jenner Institute were working on a vaccine for the novel coronavirus and

that rapid vaccine development could be possible given the scientists’ existing work and

experience with a possible MERS vaccine.

In contrast to Oxford Mail’s reporting, the U.S. major

media outlets of Fox News, CNN, The New York Times, and The Washington Post did not begin

coverage of Professor Gilbert’s COVID-19 related work until late April.

The U.S. based stories

emphasized caveats from health officials and experts downplaying the optimistic timeline and past

success of the Oxford researchers. The earliest available (major outlet) U.S. story is from CNN

on April 23

and begins with a quote from England's Chief Medical Officer Chris Whitty saying

that the probability of having a vaccine or treatment "anytime in the next calendar year" is

"incredibly small."

There is a similar disconnect between U.S. major media reporting on school reopenings and

scientific findings on the same topic; the reporting is overwhelmingly negative, while the scientific

literature tells a more optimistic story. Oster (2020) collects data on school reopenings and

COVID-19 infections within schools and districts.

She finds that infection rates among students

remain low (at 0.14 percent) and schools have not become the super-spreaders many feared.

Guthrie et al (2020) and Viner et al (2020) review the available evidence and reach similar

https://www.oxfordmail.co.uk/news/18243665.scientists-working-coronavirus-vaccine-oxford/

We base this statement on a LexisNexis search for the terms “Sarah Gilbert” or “Sarah Gilbert and vaccine” since

January 1, 2020.

https://statsiq.co1.qualtrics.com/public-

dashboard/v0/dashboard/5f62eaee4451ae001535c839#/dashboard/5f62eaee4451ae001535c839?pageId=Page_1

ac6a6bc-92b6-423e-9f7a-259a18648318.

https://www.theatlantic.com/ideas/archive/2020/10/schools-arent-superspreaders/616669/.

conclusions. However, ninety percent of school reopening articles from U.S. mainstream media

are negative versus only 56 percent for the English-language major media in other countries.

The tone of media coverage impacts both human health and attitudes towards preventative

measures including vaccination, mask wearing, and social distancing (Bursztyn et al 2020, Van

Bavel and Baicker et al 2020, Simonov et al 2020, Kearney and Levine 2015, Ash et al 2020)

The proportion of U.S. adults who exhibit depression symptoms has risen threefold since the start

of the novel coronavirus pandemic (Etman et al 2020, Fetzer et al 2020). In discussing this increase

in mental health problems, U.S. Centers for Disease Control and Prevention recommend against

heavy consumption of news stories about the pandemic

Our results suggest the CDC’s warning is prescient. We categorize by topic over 9.4 million

published news stories on COVID-19 since January 1, 2020. We then conduct several forms of

textual analysis on roughly 20,000 COVID-19 news stories to examine levels of negativity by

subtopic, source of the news, and time period. We have five major findings. First, COVID-19

stories published by the top 15 U.S media outlets (by readership/viewership) are 25 percentage

points more likely to be negative in content than more general U.S. sources or major media outlets

outside the U.S.

Second, the time pattern in observed negativity is at most weakly related to the

actual time trend in new weekly cases of COVID-19 in the U.S. Third, the most popular stories in

Bannerjee et al (2020) find that text messaging can significantly increase reporting of COVID symptoms and use of

social distancing and other health promoting measures. Nyhan et al (2014) find that it’s difficult to correct

misperceptions around vaccine safety.

https://www.cdc.gov/coronavirus/2019-ncov/daily-life-coping/managing-stress-anxiety.html

This regression-based estimate controls flexibly for article length and week of publication. The unadjusted

probability of an article being negative is 91 percent for US major media versus 54 percent for English-language

non-US major media.

The New York Times have high levels of negativity, particularly for COVID-19-related articles.

Fourth, negativity appears to be unrelated to the political leanings of the newspaper’s or network’s

audience (Niven 2001). Finally, U.S. major media stories that discuss the benefits of social

distancing or alternatively the benefits of mask wearing are less numerous than stories about

President Trump not wearing a mask. Similarly, the terms “Trump and hydroxychloroquine”

receive more coverage than do all stories about companies and researchers developing vaccines.

Overall, we find that relative to other media sources, the most influential U.S. news sources are

outliers in terms of the negative tone of their coronavirus stories and their choices of stories

covered. We are unable to explain these patterns using differential political views of their

audiences or time patterns in infection rates. This is analogous to Niven (2001) which finds a

strong negative bias in the U.S. media when covering unemployment and limited evidence of

partisanship. U.S. major outlets do demonstrate an above- average interest in promoting prosocial

behavior like mask wearing and social distancing. Consistent with the existing literature

(Gentzkow and Shapiro 2010 and Gentzkow, Glaeser and Goldin 2006), our results suggest that

U.S. major outlets publish unusually negative COVID-19 stories in response to reader demand and

interest.

Data Description

We obtain counts of COVID-19 articles and separately the text of COVID-19 articles using the

LexisNexis database. We use all English news sources and a date range of January 1, 2020 to July

31, 2020. We divide our universe of sources into the top (most widely read or watched) sources

This is consistent with the findings of Gentzkow and Shapiro (2010) who find that media respond strongly to

consumer preferences. Eshbaugh-Soha (2010) finds that negativity media coverage of the President responds to

local support for the President.

and all other sources. We further stratify by U.S. versus non-U.S. sources. The top non-TV

sources for the U.S. that are also included in LexisNexis are Newsweek, the New York Post, Los

Angeles Times, USA Today, Politico, The Hill, and the New York Times. For the top television

sources we include both written articles and television transcripts from ABC, CBS, CNN, Fox

News, MSNBC and NBC. Further details for our data downloading procedure and the search

terms used are contained in Appendix 1.

We also gather the text of articles discussing COVID-19 vaccines from five widely read scientific

and medical journals namely Science, JAMA, The New England Journal of Medicine, The Lancet,

and Nature. We gather the New York Times most popular articles from their website from

September 4-October 6

2020. We rely on the New York Times most read articles in our current

investigation, but future versions of this paper will also incorporate “most emailed” articles, outlets

beyond the New York Times, and a larger date range.

We analyze the text of 20,000 articles that fall within three subtopics regarding COVID-19:

vaccines, increases and decreases in case counts, and reopenings (of businesses, schools, parks,

restaurants, government facilities, etc). We limit ourselves to roughly 20,000 articles given the

legal requirement to “manually” download the articles from LexisNexis 100 articles at a time

We classify all articles using two different but related methods. First, we measure the fraction of

words that are negative according to established dictionaries of negative words. See Liu 2012,

Tetlock 2007, Loughran and McDonald 2011 for canonical examples of this approach.

The

LexisNexis does not permit automated downloading of the text of stories. We manually downloaded articles in

batches of 100 articles.

Riffe Lacy Fico and Watson (2019) is an in depth presentation of these methods. Grimmer and Stewart (2013)

review the value of text analysis for summarizing political documents and transcripts.

results reported here use the Hu-Liu (2004) dictionary of positive and negative words.

compute the fraction of total words that are negative according to the dictionary and standardize

this variable to be mean 0 variance 1.

Second, we create a predicted probability that an article has a negative tone. We identify

characteristics of negative and positive media reports in a set of 200 articles classified as strongly

positive or negative by human readers. We use the two- and three-word phrases appearing in the

training articles combined with machine leaning techniques to find the phrases that best predict

whether the human reader will classify an article as strongly negative. We implement a Naïve

Bayes classification scheme (Zhang 2004 , Pazzani 1996, Antweiler and Frank 2004).

Naïve

Bayes assumes that each phrase in the article contributes independently to the probability that the

article is negative and maximizes the number of correct predictions given the phrases.

We use the resulting model to predict whether each of the 20,000 articles in our sample are

negative. For example, the inclusion of the phrases “clinical trial” and “Jenner Institute” are strong

predictors of an article being positive while “White House” and “death toll” are strong predictors

of a negative article.

Table 1 reports summary statistics at the article level for our main sample which excludes New

York Times most popular articles and a comparison sample of non-COVID articles. We analyze

roughly 23,500 articles from January 1, 2020 to July 31, 2020. On average the articles have 1652

We have conducted the same analysis using the Harvard General Inquirer dictionary of positive and negative

words and obtain qualitatively similar results. http://www.wjh.harvard.edu/~inquirer/

To keep the yardstick consistent, we standardize once within our broad sample which includes New York Times

most popular articles and a sample of non-Covid articles. Our main analysis sample excludes these two categories.

We standardize before dropping duplicates of articles which were published multiple times.

To extract phrases and implement the Naïve Bayes classification scheme we use WordStat software created by

Provalis research.

words. This count and our subsequent statistics are measured after we apply a truncation procedure

to limit the text to be within 10 lines of the words “COVID” or “coronavirus”. We applied this

truncation to deal with very long television transcripts that switched to non-COVID topics in the

middle of the transcript. However, results are quite similar with or without truncation.

The share of negative words (using the Hu-Liu dictionary) is 4 percent. As mentioned above we

standardize this variable to aid in interpreting the coefficients. By construction, our articles are

divided roughly equally between articles on increases/decreases in cases, reopenings, and

vaccines. The division among US major media, US General media, International Major Media,

and International General media is also roughly equal.

Results

Figure 1 plots the time trend in media negativity for major media outlets in the U.S. (green line)

and outside the U.S. (blue line) using the scale on the left. The most striking fact is that 91 percent

of the U.S. stories are classified as negative whereas 54 percent of the non-U.S. stories are

classified as negative. Figure 1 uses our estimated probability that an article is negative. We

obtain similar results using the Hu-Liu dictionary and the fraction of words in the article that are

negative.

The red line plots the weekly average of daily new cases of COVID-19 in the U.S. using the scale

on the right. The x-axis is the week of the year within 2020. New cases per day rise sharply from

March through mid-April. Cases decline until about June 15th, then rise rapidly until late July,

We don’t have exactly 25% of articles in each major category because our initially drawn sample included many

articles that were repeats which we then eliminated to arrive at this final sample. Our reopenings analysis is for all

reopenings articles. We also examine school reopenings specifically and for these articles “school” must appear in

the title of an article that also contains “reopen” or “re-open”.

when cases begin to decline again. Average media negativity over time is not correlated with new

case counts, as regression results confirm (not reported).

In Table 2 we regress our estimated probability that an article is negative on indicator variables for

whether the source is from the U.S. major media, U.S. general sources, or international general

sources. The omitted category is international (non-U.S.) major media sources. In the regressions

we control flexibly for the length of the article and the week the article was published. We run a

linear probability model, though results from probit and logit models are similar to those reported

here. The non-U.S. major media sources have a baseline rate of negativity of 54 percent. In

column (1) we show that relative to this omitted category, articles in the U.S. major media are 25

percentage points more likely to be negative. In contrast, U.S. general and non-U.S. general

sources have about the same level of negativity as non-U.S. major media.

In column (2) we switch the dependent variable to the share of negative words in the article. We

standardize the outcome to be mean 0 standard deviation 1. The U.S. major media publish stories

that are .23 standard deviations more negative relative to non-U.S. major media. U.S. general

media outlets are significantly less negative than all other categories of sources. In columns (3)-

(5) we examine media negativity by subtopic within COVID-19. Relative to both types of

international media, U.S. major media vaccine articles are particularly negative. Vaccine stories

in the U.S. major media are 45 percentage points more likely to be negative relative to stories in

the non-U.S. general media.

In Figure 2 we present the mean share of negative words (standardized) by source and topic

(COVID-19 versus not). For Figure 2 only, we add a large sample of non-COVID articles and the

New York Times most popular articles. Starting with the bars at the bottom of the chart, we see

that in a sample of non-COVID-19 stories (pre-January 2020), the U.S. major media are only

modestly more negative than the rest of the sample. In covering COVID-19 (the second bar from

the bottom), U.S. major media negativity is .31 standard deviation above the average while the

non-U.S. major media are .17 standard deviation below average. Notably, scientific media articles

on COVID-19 vaccines are a full standard deviation below average in negativity. In contrast, the

New York Times’ most popular articles are .6 standard deviations above the sample mean in

negativity for non-COVID-19 stories and 1.5 standard deviations above the mean when covering

COVID-19 topics. Readers of the U.S. major media (as represented by the New York Times) are

attracted to negative stories in general and negative stories about COVID-19 in particular.

The next two figures look specifically at the share of words that are negative within vaccine articles

(Figure 3) and within school reopening articles (Figure 4). We standardize across the entire sample

(all topics) and hence are comparing the negativity in the vaccine articles to the overall sample

mean. For vaccine articles, all media categories are meaningfully below the overall sample mean

for negativity, except for the U.S. major media which produces articles on COVID-19 vaccines

that are .35 standard deviations higher on negativity.

These data were gathered prior to Pfizer’s positive stage three trial result announced on November

. Our results show that on a relative basis, U.S. major media gave much less positive coverage

to the developments that lead up to Pfizer’s breakthrough. We hypothesize (but have not yet

tested) that U.S. major media coverage of vaccines remained more negative than other categories

of media during and after the Pfizer announcement.

For school re-opening articles (Figure 4), the U.S. major media is .18 standard deviations more

negative than the overall sample mean. All other media categories are less negative than the

sample mean. The U.S. general media vaccine articles are .4 standard deviations less negative.

A natural question is whether media negativity varies greatly by the specific news source and

whether that variation is related to the political beliefs of the readership. Our results are perhaps

surprising. COVID-19 stories from all the major U.S. outlets have high levels of negativity and

the variation that does exist is not correlated with readers’ political leanings. See Figure 5. We

plot the share of negative words (standardized) by U.S. media source versus the probability that

conservative-leaning people say that this is a “trusted media source.” The latter comes from a

2019 Pew survey of 12,000 people about their consumption of election news.

The estimated probability that a COVID-19 article is negative varies from 70 percent to 100

percent among major U.S. outlets. These probabilities are not correlated with the likelihood that

conservative consumers of news trust the source. COVID-19 stories from Fox News are about as

negative as those from CNN. We obtain similar results using our estimated probability that a story

is negative.

We now take a broader look at which COVID-19 topics the media choose to emphasize. Table 3

provides an overview of the number of COVID-19-related articles during our sample period

(January-July 2020) and counts of articles by topic, where one article can cover multiple topics.

Overall, we found 2.6 million articles from U.S.-based sources and 6.4 million from non-U.S.

sources. The rows represent different search terms we included while the columns represent four

broad categories of sources, namely U.S. versus non-U.S. interacted with major media outlet

versus general media. We are most interested in the relative coverage of different topics. For

example, among the U.S. major media (column 2) 15,000 stories mention increases in caseloads

https://www.pewresearch.org/fact-tank/2020/01/24/qa-how-pew-research-center-evaluated-americans-trust-

in-30-news-sources/

while only 2,500 mention decreases, or a 6 to 1 ratio. Even when caseloads were falling nationally

(April 24

to June 27

), this ratio remains relatively high at 5.3 to 1.

In row 3, we show results for mentions of COVID-19 vaccines and any names of the top ten

institutions or companies working on a COVID-19 vaccine. The U.S. major outlets ran 1,371 such

stories. During the same period they ran 8,756 stories involving Trump and mask wearing and

1,636 stories about Trump and hydroxychloroquine.

A natural question is whether the media is promoting prosocial behaviors (Simonov et al 2020 and

Burstyn et al 2020). While we cannot answer whether the U.S. media are “doing enough” to

promote transmission-reducing behavior in absolute terms, we can compare how emphasis of the

benefits of mask wearing or social distancing varies across media categories. Five percent of

COVID-19 articles in U.S. major outlets mention the benefits of mask wearing compared to .6

percent for non-U.S. outlets and 2 percent for general U.S. sources. U.S. major media outlets are

also much more likely to discuss the benefits of social distancing (4 percent of stories) than their

non-U.S. counterparts (1 percent of stories). This suggests the U.S. media are outperforming the

non-U.S. media in promoting prosocial behavior, though perhaps because such messages are more

needed in the U.S.

Overall, we find that COVID-19 stories from U.S. major media outlets are much more negative

than similar stories from other U.S. outlets and from non-U.S. sources. The negativity does not

respond to changes in new cases. Potentially positive developments such as vaccine stories receive

less attention from U.S. outlets than do negative stories about Trump and hydroxychloroquine.

Overall, we are unable to explain the variation in negativity with political affiliation of an outlet’s

See Della Vigna and La Ferrara (2015) for a summary which discusses more generally the impact of media

consumption on human behavior.

audience, or U.S case count changes, but we do find that U.S. readers demand negative stories (as

evidenced by article popularity). We conclude that the CDC’s implicit “warning label” against

consuming too much U.S. COVID-19 media may be warranted.

References

Antweiler, Werner, and Murray Z. Frank. "Is all that talk just noise? The information content of

internet stock message boards." The Journal of finance 59, no. 3 (2004): 1259-1294.

Ash, Elliott, Sergio Galletta, Dominik Hangartner, Yotam Margalit, and Matteo Pinna. "The

Effect of Fox News on Health Behavior During COVID-19." Available at SSRN

3636762 (2020).

Banerjee, Abhijit, Marcella Alsan, Emily Breza, Arun G. Chandrasekhar, Abhijit Chowdhury,

Esther Duflo, Paul Goldsmith-Pinkham, and Benjamin A. Olken. Messages on covid-19

prevention in india increased symptoms reporting and adherence to preventive behaviors

among 25 million recipients with similar effects on non-recipient members of their

communities. No. w27496. National Bureau of Economic Research, 2020.

Bursztyn, Leonardo, Aakaash Rao, Christopher Roth, and David Yanagizawa-Drott.

"Misinformation during a pandemic." University of Chicago, Becker Friedman Institute

for Economics Working Paper 2020-44 (2020).

Eshbaugh-Soha, Matthew. "The tone of local presidential news coverage." Political

Communication 27, no. 2 (2010): 121-140.

Ettman, Catherine K., Salma M. Abdalla, Gregory H. Cohen, Laura Sampson, Patrick M. Vivier,

and Sandro Galea. "Prevalence of depression symptoms in US adults before and during

the COVID-19 pandemic." JAMA network open 3, no. 9 (2020): e2019686-e2019686.

Fetzer, Thiemo, Lukas Hensel, Johannes Hermle, and Christopher Roth. "Coronavirus

perceptions and economic anxiety." Review of Economics and Statistics (2020): 1-36.

DellaVigna, Stefano, and Eliana La Ferrara. "Economic and social impacts of the media."

In Handbook of media economics, vol. 1, pp. 723-768. North-Holland, 2015.

Gentzkow, Matthew, and Jesse M. Shapiro. "What drives media slant? Evidence from US daily

newspapers." Econometrica 78, no. 1 (2010): 35-71.

Gentzkow, Matthew, Edward L. Glaeser, and Claudia Goldin. "The rise of the fourth estate. How

newspapers became informative and why it mattered." In Corruption and reform:

Lessons from America's economic history, pp. 187-230. University of Chicago Press,

2006.

Gentzkow, Matthew, Bryan Kelly, and Matt Taddy. "Text as data." Journal of Economic

Literature 57, no. 3 (2019): 535-74.

Groseclose, Tim, and Jeffrey Milyo. "A measure of media bias." The Quarterly Journal of

Economics 120, no. 4 (2005): 1191-1237

Gurun, Umit G., and Alexander W. Butler. "Don't believe the hype: Local media slant, local

advertising, and firm value." The Journal of Finance 67, no. 2 (2012): 561-598.

Guthrie, Brandon L., Jessie Seiler, Lorenzo Tolentino, Wenwen Jiang, Molly Fischer, Rodal

Issema, Sherrilynne Fuller, Dylan Green, Diana M. Tordoff, Julianne Meisner, Ashley

Tseng, Diana Louden, Jennifer M. Ross, Alison L. Drake “Summary of Evidence Related

to Schools During the COVID-19 Pandemic” Report from COVID-19 Literature Report

Team, Washington State Department of Public Health. October 2020.

Grimmer, Justin, and Brandon M. Stewart. "Text as data: The promise and pitfalls of automatic

content analysis methods for political texts." Political analysis 21, no. 3 (2013): 267-297.

Hu Minqing and Bing Liu. "Mining and Summarizing Customer Reviews." Proceedings of the

ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

(KDD-2004), Aug 22-25, 2004, Seattle, Washington, USA

Kearney, Melissa S., and Phillip B. Levine. "Media influences on social outcomes: The impact of

MTV's 16 and pregnant on teen childbearing." American Economic Review 105, no. 12

(2015): 3597-3632.

Larcinese, Valentino, Riccardo Puglisi, and James M. Snyder Jr. "Partisan bias in economic

news: Evidence on the agenda-setting behavior of US newspapers." Journal of public

Economics 95, no. 9-10 (2011): 1178-1189.

Liu, Bing. "Sentiment analysis and opinion mining." Synthesis lectures on human language

technologies 5.1 (2012): 1-167.

Loughran, Tim, and Bill McDonald. "When is a liability not a liability? Textual analysis,

dictionaries, and 10‐Ks." The Journal of Finance 66, no. 1 (2011): 35-65.

Martin, Gregory J., and Ali Yurukoglu. "Bias in cable news: Persuasion and

polarization." American Economic Review 107, no. 9 (2017): 2565-99.

Niven, David. "Bias in the news: Partisanship and negativity in media coverage of presidents

George Bush and Bill Clinton." Harvard International Journal of Press/Politics 6.3

(2001): 31-46.

Nyhan, Brendan, Jason Reifler, Sean Richey, and Gary L. Freed. "Effective messages in vaccine

promotion: a randomized trial." Pediatrics 133, no. 4 (2014): e835-e842.

Oster, Emily, “Schools Aren’t Super-Spreaders”, The Atlantic, October 9, 2020.

Pazzani, M. J. 1996. Search for dependencies in Bayesian classifiers. In Fisher, D., and Lenz, H.

J., eds., Learning from Data: Artificial Intelligence and Statistics V. Springer Verlag.

Puglisi, Riccardo. "Being the New York Times: the political behaviour of a newspaper." The BE

journal of economic analysis & policy 11, no. 1 (2011).

Riffe, Daniel, Stephen Lacy, Frederick Fico, and Brendan Watson. Analyzing media messages:

Using quantitative content analysis in research. Routledge, 2019.

Simonov, Andrey, Szymon K. Sacher, Jean-Pierre H. Dubé, and Shirsho Biswas. The persuasive

effect of fox news: non-compliance with social distancing during the covid-19 pandemic.

No. w27237. National Bureau of Economic Research, 2020.

Tetlock, Paul C. "Giving content to investor sentiment: The role of media in the stock

market." The Journal of finance 62, no. 3 (2007): 1139-1168.

Van Bavel, Jay J., Katherine Baicker, Paulo S. Boggio, Valerio Capraro, Aleksandra Cichocka,

Mina Cikara, Molly J. Crockett et al. "Using social and behavioural science to support

COVID-19 pandemic response." Nature Human Behaviour (2020): 1-12.

Viner, Russell M., et al. "School closure and management practices during coronavirus outbreaks

including COVID-19: a rapid systematic review." The Lancet Child & Adolescent

Health (2020).

Zhang H. (2004). The optimality of Naive Bayes. Proc. FLAIR.

Figure 1: Media Negativity and New COVID-19 Cases Over Time

Notes: Negativity is estimated using supervised machine learning on article phrases coupled with a training data set.

Articles are manually downloaded from LexisNexis for the period January 1

, 2020 to July 31

, 2020. The red line

shows the weekly average of daily confirmed new COVID-19 cases and is accessed from the New York Times

website.

Figure 2:

Media Negativity by Source for COVID-19 and Non-COVID-19

Articles

Notes: Negativity is estimated as the fraction of negative words in the article and is standardized. Dark blue bars are

for COVID related articles and light blue bars are for non-COVID related articles. The raw share of negative words

is .043 with a standard deviation of .021. Negative words are defined by the Hu-Liu (1997) dictionary. Articles and

transcripts are manually downloaded from LexisNexis for the period January 1

, 2020 to July 31

, 2020 and

websites for Science, JAMA, The New England Journal of Medicine, The Lancet, and Nature. The New York Times

website is used for the list and text of the most popular articles.

Figure 3:

Media Negativity by Source for COVID-19 Vaccine Articles

Figure 4:

Media Negativity by Source for School Reopening Articles

igure 5: Media Negativity and Audience Political Leanings

Notes: Negativity is estimated as the fraction of negative words in the article and is standardized. The raw share of

negative words is .043 with a standard deviation of .021. Negative words are defined by the Hui-Lu (1997)

dictionary. Articles and transcripts are manually downloaded from LexisNexis for the period January 1

, 2020 to

July 31

, 2020. “Trusted source” is measured in a 2019 Pew Survey of U.S. adults.

Table 1: Summary Statistics

(1)

(2)

(3)

(4)

(5)

mean

min

max

Word Count of Article

23,486

1,652

2,654

57,166

Estimated P(Article is Negative)

23,486

0.663

0.384

Share of Words That Are Negative

23,486

0.0420

0.0200

0.190

Share Words Negative Standardized

23,486

-0.118

0.999

-2.212

7.269

Is a Scientific Article on Vaccines

23,486

0.00860

0.0923

Is an Increase/Decrease in Cases Article

23,486

0.303

0.460

Is a Reopenings Article

23,486

0.350

0.477

Is a Vaccine Article

23,486

0.347

0.476

US Major Media

23,486

0.292

0.454

US General Media

23,486

0.236

0.425

International Major Media

23,486

0.270

0.444

International General Media

23,486

0.194

0.395

Fraction of Conservatives Who Trust This Source (US Major Media)

8,131

0.252

0.137

0.116

0.742

Notes: We present summary statistics for our main variables. Each article is one observation. Probability of the article being negative is estimated using

supervised machine learning on article phrases coupled with a training data set. Share of negative words is estimated as the fraction of negative words in the

article and is standardized. The raw share of negative words is .043 with a standard deviation of .021. Negative words are defined by the Hui-Lu (1997)

dictionary. Articles are manually downloaded from LexisNexis for the period January 1

, 2020 to July 31

, 2020.

able 2: Negativity by Media Category and Topic

(1)

(2)

(3)

(4)

(5)

Probability

Article is

Negative (All

Articles)

Share of

Negative

Words

Standardized

(All Articles)

Probability

Article is

Negative

(Vaccine

Articles)

Probability

Article is

Negative

(Case Count

Articles)

Probability

Article is

Negative

(Reopening

Articles)

US Major Media

0.253***

0.234***

0.452***

0.188***

0.203***

(0.00666)

(0.0192)

(0.0121)

(0.00805)

(0.00776)

US General Media

-0.00196

-0.422***

-0.0281**

0.0521***

0.142***

(0.00834)

(0.0206)

(0.0126)

(0.0111)

(0.00862)

International General Media

-0.00727

-0.0750***

0.0212**

0.0771***

0.0896***

(0.00648)

(0.0184)

(0.00935)

(0.00824)

(0.00789)

Observations

20,909

7,246

6,295

7,367

R-squared

0.388

0.223

0.546

0.422

0.423

Mean Negativity for Intl Major Media

.541

-.160

.242

.686

.645

(the Omitted Category)

Notes: Probability of the article being negative is estimated using supervised machine learning on article phrases coupled with a training data set. Share of

negative words is estimated as the fraction of negative words in the article and is standardized. The raw share of negative words is .043 with a standard deviation

of .021. Negative words are defined by the Hui-Lu (1997) dictionary. Articles are manually downloaded from LexisNexis for the period January 1

, 2020 to

July 31

, 2020. All columns use OLS with robust standard errors. *** p<0.01, ** p<0.05, * p<0.1

Table 3: Total COVID-19-Related Media Articles by Topic: January 31

2020 to July 31

, 2020

Topic

U.S. Total

U.S.

mainstream

U.S. non-

mainstream

Non-U.S.

Total

Non-U.S.

mainstream

Non-U.S.

non-

mainstream

Coronavirus/COVID-19

2,594,510

90,600

2,503,910

6,823,410

453,900

6,369,510

Vaccines

33,980

2,375

31,605

69,600

3,257

66,343

Vaccines + Sarah Gilbert Etc.

28,740

1,371

27,369

54,860

2,299

52,561

Increases Whole Time Period

325,550

15,200

310,350

666,895

41,386

625,509

Decreases Whole Time Period

87,550

2,462

85,088

99,630

3,067

96,563

Increases 4/24-6/27 Period

103,700

3,581

100,119

314,548

16,660

297,888

Decreases 4/24-6/27 Period

33,000

676

32,324

53,850

1,297

52,553

Reopening

412,780

19,300

393,480

680,052

31,630

648,422

Masks

386,890

23,600

363,290

670,994

43,090

627,904

Masks and Trump

56,579

8,756

47,823

46,187

2,339

43,848

Benefits Masks

51,700

4,436

47,264

61,680

2,687

58,993

Social Distancing

378,940

19,600

359,340

811,503

55,610

755,893

Benefits Social Distancing

60,450

3,975

56,475

86,249

4,163

82,086

Hydroxychloroquine

21,440

2,273

19,167

33,005

2,746

30,259

Hydroxychloroquine and Trump

10,640

1,636

9,004

12,503

929

11,574

Notes: Article counts come from a LexisNexis for the period January 1

, 2020 to July 31

, 2020. The left most column indicates the search terms used (see

methodology documents for exact searches). The article can be counted in multiple rows if the article contains both sets of terms.

Appendix 1

Dataset Construction Details and Search Terms Used

Data Set Construction

Our dataset was assembled from Nexus Lexis articles. We utilized the following instructions:

1. Click on the link (links were derived from search terms at the bottom of this document)

(setting pages to display 50 at a time instead of 10)

2. Click the dropdown on the left that says location by publication

3. Click edit settings

4. In results display settings, switch it from 10 to 50.

5. Scroll to the bottom and hit save (you may have to do this every time for each link, not entirely sure how it “saves”

(downloading)

6. Before downloading, double check that you are sorting by relevance, and the slider is set to group duplicates

7. Click the little box beside the folder to select the whole page

8. Go to the next page and do the same

9. Click the download button which looks like it’s a downwards pointing arrow

10. In the dialog box, make sure the format is RTF and “save as individual files” these likely won’t be done already.

11. Download, and repeat until reaching 2500/link. In the final dataset this number may be less due to duplicates.

Lexus Nexus Article Search Process

vaccines

inc/dec

reopening

coronavirus or COVID-19 and

ATLEAST5(vaccine)

coronavirus or COVID-19 and cases and increase or

decrease

COVID-19 or coronavirus and

reopening

American mainstream sources in our dataset consisted of:

US Mainstream

Sources

International

Mainstream Sources

Fox

AFR

IndianExpress

Hindu

MSNBC

Analysis

MetroUK

Sun (England)

ABC

AsiaPacific

Newcastle

SunHerald

CBS

AustralianFin

Northern Territory

Sydney Morning

Herald

CNN

BrisbaneTimes

SundayAge

Times of India

NBC

CTV

SundayHerald

TorontoStar

NPR

CanberraTimes

SydneyMorning

WestAZ

LAtimes

DailyMirror

Advertiser

WAToday

Newsweek

Geelong Advertiser

TheAge

Telegraph

Politico

HeraldSun

TheAustralian

Guardian

TheHill

HinduTimes

AustralianMag

NYtimes

Hobart

Courier

NYPost

IllawarraMercery

EveningStandard

USAToday

IndiaToday

GlobeMail

Appendix Table 1 (not for Publication):

Negativity by Specific Media Source

(1)

(2)

Prob Article is Negative—U.S.

Sources

Share of Negative Words (Standardized)—U.S.

Sources

Fox

0.396***

0.681***

(0.0131)

(0.0384)

MSNBC

0.283***

0.294*

(0.0532)

(0.156)

ABC

0.367***

0.745***

(0.0169)

(0.0495)

CBS

0.357***

0.707***

(0.0210)

(0.0616)

CNN

0.394***

0.833***

(0.00757)

(0.0222)

NBC

0.101**

0.365**

(0.0492)

(0.144)

NPR

0.255***

0.572***

(0.0125)

(0.0367)

LATimes

0.413***

1.088***

(0.0175)

(0.0513)

Newsweek

0.174***

0.170

(0.0483)

(0.142)

Politico

0.341***

1.045***

(0.0243)

(0.0712)

TheHill

0.563***

1.167***

(0.149)

(0.436)

NYTimes

0.256***

1.126***

(0.0109)

(0.0320)

NYPost

0.149***

0.788***

(0.0463)

(0.136)

USAToday

0.293***

0.884***

(0.0240)

(0.0704)

Constant

0.838***

-2.169***

(0.139)

(0.408)

Observations

10,156

R-squared

0.394

0.298

Omitted category consists of all U.S. sources not named above. Regressions are estimated using a linear probability model. Robust standard

errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Appendix Table 2 (not for publication):

Relationship Between Negativity and Political Leanings of Audience

(1)

Probability Article is Negative

(Controlling for Politics)

conservative_trusted_source

0.0507

(0.0404)

Constant

0.300**

(0.111)

Observations

11,505

R-squared

0.355

Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Appendix Figure 1 (not for publication):

Probability Article is Negative and Audience Political Leanings