angelucci

Measuring Voters’ Knowledge of Political News∗
Preliminary
Charles Angelucci† Andrea Prat‡

Columbia University
April 2020
Abstract
We propose a methodology to measure voters’ knowledge of news about recent
political events that combines a transparent protocol for identifying stories, an
incentivized quiz to elicit knowledge, and estimation of a model of individual
knowledge that includes story difficulty, partisanship, and memory decay. We
apply our methodology to measure knowledge of news about the US Federal
government in a monthly sample of 1,000 US voters repeated eight times. People
in the most informed tercile are 72% more likely than people in the bottom tercile
to know the main story of the month. We also document large inequalities across
socioeconomic groups, with the best-informed group over 16 percentage points
more likely to know the typical news story compared to the least-informed group.
We find that voters are 7-19% less likely to know stories that reflect poorly on their
preferred political party. Time also matters, with each month passing lowering
the odds of knowing the typical news story by 5 percentage points. We repeat
our study on news about the Democratic Party primaries.
∗
We would like to thank Laurent Bouton, Julia Cagé, Teresa Esteban Casanelles, Simone Galperti,
Leo Goldman, Amit Khandelwal, José L. Montiel Olea, Suresh Naidu, Jacopo Perego, Jett Pettus,
Vincent Pons, Miklos Sarvary, Andrey Simonov, Sevgi Yuksel, and seminar audiences at Columbia
University, the University of Exeter, Harvard University, the NYC Media Seminar, the 4th Summer
Workshop in Political Economy at EIEF, and the Rational Inattention and Political Economy
Conference at UC San Diego for useful suggestions and comments. We are grateful to Szymon Sacher
for outstanding research assistance. We gratefully acknowledge funding received from The Media and
Technology Program at the Columbia Business School
†
3022 Broadway, New York, NY 10027. Email: [email protected]
‡
3022 Broadway, New York, NY 10027. Email: [email protected]
1
1 Introduction
The media plays an important role in providing voters with the information they need
to keep government accountable. Informed voters are aware of what the government
does and are thus in a position to punish or reward the incumbent at the next election.
The central role played by the media in maintaining government accountability is well-
documented by a growing body of literature in political economy. For example, in
the US, Snyder and Strömberg [2010] find that political districts with greater media
coverage elect representatives who work harder to promote their constituents’ interests.
Similarly, in Uganda, Reinikka and Svensson [2005] document that schools in areas with
greater newspaper coverage are better run. This logic applies to new media: Gavazza
et al. [2018] show that the expansion of broadband internet in the UK crowded out
local news and ultimately reduced local public spending.1,2
A government that is aware of the link between information and voting behavior is
also more likely to cater to the better-informed voters. This proposition has received
empirical support: for example, Strömberg [2004] shows that US counties with higher
radio ownership received greater funding from the Federal Government during the New
Deal. The logic can be formalized in a simple model of retrospective voting [Strömberg,
2001, Prat and Strömberg, 2013]. An incumbent politician knows that voters care about
her policy choices. If different social groups have different levels of information, better
informed groups will be more responsive to the incumbent’s behavior and the latter
1
Other papers showing an effect of news coverage on political outcomes include Eisensee and
Strömberg [2007], Ferraz and Finan [2008], Gerber et al. [2009], Enikolopov et al. [2011], Banerjee
et al. [2012], Kendall et al. [2015], Labonne et al. [2019], Arias et al. [2018], and Arias et al. [2019].
See Strömberg [2015] for a survey of the literature linking mass media and political outcomes.
2
Media bias can also affect political outcomes [e.g., DellaVigna and Kaplan, 2007, Martin et al.,
2017]. On the relationship between media bias and political outcomes see also Gentzkow et al. [2015].
2
will design policies that cater to them.3 Inequalities in knowledge of the news are thus
likely to exacerbate other existing types of inequalities [Delli Carpini and Keeter, 1996].
Voters’ knowledge of political news is, therefore, a key ingredient of many political
economy models. Those theories do not just consider average knowledge but also how
knowledge is distributed across topics and voters. Indeed, there exists a sizeable body
of work that measures voter knowledge, with some of it focusing on news knowledge.
Polling and research organizations regularly report survey results on voter knowledge
[e.g., Pew, 2017, Eurobarometer, 2017].4 On the academic side, the public opinion
literature has provided a number of measures for political knowledge. Price and Zaller
[1993] measure recall of 16 news stories. Examples of survey questions include: “Do you
remember any recent stories about Marine Colonel Oliver North receiving a sentence for
his conviction in the Iran-Contra Affair? [If yes:] Do you recall anything about what
sentence he received?” and “Do you recall any stories about a U.S. Supreme Court
decision this summer on abortion? [If yes:] Do you remember what the court decided?”
They find that respondents’ background level of political knowledge is the strongest
predictor of current news recall across a wide range of topics.
The canonical work in this area is Delli Carpini and Keeter [1996], who collate
about 3,700 questions asked in various surveys from 1940 to 1993, with the objective to
measure the American public’s level of political knowledge. They divide questions
into five categories, one of which is domestic politics. In the last year for which
they have information (1990), the statements are: “Who will pay for S&L bailout?”;
3
A simple model (developed in Online Appendix A) shows that if ρ̄g is the average news knowledge
level in social group g, a re-election seeking incumbent will choose her behavior as if maximizing a
welfare function where each group’s weight is proportional to its news knowledge level.
4
The American National Election Studies (ANES) also include two questions on political
knowledge: ‘Which party had most members of congress before the election?’ and ‘Which party
had most members of congress after the election?’.
3
“Why is the Hubel telescope in the news?”; “Did Bush veto a plant closing bill?”;
“What is the illiteracy rate in US?”; “What is the percentage of population that is
Hispanic/Black/Jewish?”5
In recent years, news knowledge has been examined from the perspective of fake
news. Some commentators have argued that misinformation spread through social
media has played an important role in elections and referenda around the world [e.g.,
Levitin, 2016, Stengel, 2019]. Alcott and Gentzkow [2017] measure consumption and
recall of fake news in the 2016 election, and Barrera Rodriguez et al. [2018] investigate
the role played by fake news and fact-checking on French voters’ beliefs and political
preferences. Lazer et al. [2018] discuss the prevalence and impact of the phenomenon
and potential interventions. More recently, Allcott et al. [2019] measure the effect of
Facebook on news knowledge.6 To measure knowledge, they include a list of 15 true and
false statements and ask respondents to select which, in their opinion, are true. The
true statements are borrowed from recent articles published in the New York Times,
the Wall Street Journal, Fox News, CNN, MSCNBC, and US News & World Report.
The false statements are either modifications of existing articles from the same sources
or recent fake news identified by third-party fact-checkers. Allcott et al. [2019] show
that Facebook usage tends to increase knowledge of the news.
Any news knowledge measurement exercise faces an initial challenge: what set of
knowledge items should voters be tested on? As the examples above illustrate, this
challenge is hard because the set of possible news items is unstructured, heterogeneous,
5
Relatedly, Prior and Lupia [2008] measure political knowledge by administering surveys that
include 14 questions about facts relevant to the 2004 presidential election. They find that typical survey
methods (quick, unincentivized questions) likely underestimate voters’ true knowledge of politics.
6
See also Chen and Yang [2019] on the relationship between consumption of uncensored information
and knowledge of current events in China.
4
and virtually unbounded. To the best of our knowledge, the literature approaches this
challenge by letting the researchers select the set of news stories over which survey
respondents are quizzed. While this methodology is natural, it has drawbacks in
terms of interpretation and replicability. Only the researcher knows what universe of
knowledge items he or she considered and what criterion he or she used to select within
that universe the items that ended up on the survey. Only that particular researcher
is in a position to replicate the knowledge selection process at a different time.
This paper contributes to the literature in three ways. First, we develop and
implement a simple news selection protocol that is outside the control of the researcher.
Second, this leads to a stable news-generating stochastic process that we use to estimate
a structural model of voter knowledge. Third, the structural model allows us to
disentangle the effect of some of the factors that affect knowledge such as news characteristics,
individual characteristics, partisanship, and time passing.
Our news selection process consists of two steps: (i) Selection of the universe of
relevant news items. The protocol selects a news source, sets an inclusion criterion, and
identifies the set of stories that satisfy that criterion. The researcher has no hand on the
content and wording of the stories. (ii) Selection of the knowledge items to be included
in the test. The protocol specifies a process to select a subset of (i). This step may rely
on the subjective judgment of other agents, but the process must be codified. For (i),
this paper uses the set of all Reuters news wires devoted to US national politics. For (ii),
we assemble a panel of journalists and ask them to select – within the subset identified
in (i) – the three most important stories of the month about the Federal Government.
We then conduct surveys to measure US voters’ knowledge of these stories.
The importance of a story is clearly a subjective matter, and any attempt to measure
importance ultimately relies on someone’s judgment. The proposed approach does not
5
aim at universality or objectivity but just transparency. The subjectivity in our protocol
can be ascribed to well-defined set of actors: a large for-profit organization like Reuters
and a panel of professional journalists. The protocol provides a way of identifying the
most important Federal Government stories according to mainstream journalism.
We exploit the protocol in a number of ways. Chiefly, we repeat the survey for
eight months on eight different panels of approximately 1,000 US voters. On several
occasions, we also included 1- and 2-month-old stories, to measure knowledge decay over
time. Finally, we extend the protocol to news about the Democratic Party presidential
primaries, chosen among the same set of Reuters news wires about national politics
and ranked by the same panel of journalists.
Once news stories about the Federal Government are selected, we measure knowledge
in a financially incentivized survey in many ways similar to those used by, for instance,
Allcott et al. [2019], Guess [2015], Prior et al. [2015], Bullock et al. [2015], and Chen
and Yang [2019].7 Respondents are selected by YouGov, a polling company, to produce
a nationally representative sample of US voters. As part of the survey, respondents take
multiple quizzes. In each quiz, we present our respondents with six items: the three
most important knowledge items of the month according to our panel of journalists as
well three plausible but false statements. Consistent with our approach to real news,
we rely on the panel of journalists to create the false statements. The fake statements
cover the Federal Government and are written in the same journalistic style as the true
knowledge items. Survey respondents are given 60 seconds to select the 3 statements
which, to the best of their recollection, are true. They receive a monetary reward in
7
On the role of partisanship and incentives to recall information accurately see Prior et al. [2015]
and Bullock et al. [2015]. Both papers show that monetary incentives lead to less party cheerleader
behavior in answering survey questions. On the effects of monetary incentives in surveys that measure
political knowledge see also Prior and Lupia [2008].
6
case all three true knowledge items are chosen.8
The survey data is used to estimate the distributions of parameters of a news
knowledge model. In our model knowledge is a continuous variable: when a respondent
is confronted with a news story (true or false), she assigns a probability of truth between
zero and one that depends on (i) features of the story like salience and partisanship
(e.g., whether the story reflects favorably on the Republican Party) and (ii) features
of the respondent like knowledge and ideology. The respondent uses these assigned
probabilities to select the three stories he or she thinks are most likely to be true.
The model yields a discrete choice specification that can be estimated with standard
Bayesian techniques. While every news story is different and may be harder or easier,
the stochastic generating process for both true and fake stories is exogenously given.
The main object of interest is the posterior distribution of the respondent-level knowledge
parameter, but we also obtain estimates for the salience and partisanship of each story,
as well as the effect of time passing on news knowledge.
In our main analysis, we measure voters’ knowledge of news stories about the
Federal Government. An agent’s knowledge of a particular news story is the estimated
probability the agent assigns to that story being true. Our findings can therefore be
reported at different knowledge levels. If for now we define “knowledge” as attributing
a chance equal to at least 75% that a news story is true, according to our estimates
the average voter knows approximately half of the three most important news stories
of the month. About 66% of US voters know the most important (according to the
journalists) story of the month, and the share of US voters who know the second and
8
This approach implicitly defines knowledge as awareness of a fact. A deeper notion of knowledge
entails understanding that fact. One may be aware that President Trump was impeached without
truly understanding what the impeachment process is. One limitation of our approach is that we only
attempt to measure this more superficial form of knowledge.
7
third most important stories of the month falls to 50% and 38%, respectively.
Significant heterogeneity across voters exists. For instance, the average voter in the
top-third of the distribution knows roughly 1.7 out of 3 news stories. By contrast, the
average voter in the bottom-third of the distribution knows roughly 1.1 news stories.
Further, we find a relatively large effect of partisanship/congruence on knowledge, with
respondents being 7-19% more likely to know news stories that reflect favorably on their
preferred political party. Time also significantly affects knowledge of political news: we
document that one month of time (two months of time) reduces by 4 percentage points
(7 percentage points) the share of voters who know a given story. We also measure
inequalities in news knowledge across socioeconomic groups (defined by age, gender,
race, and income). According to our estimates, the best-informed group (wealthy and
college-educated white men aged 47 and more) is over 16 percentage points more likely
to know the typical news story compared to the least-informed group (low-income and
high-school educated young women).9 Finally, we investigate the relationship between
news diets and knowledge of the news. We find that news consumption (defined both in
number of news outlets and time usage) positively predicts knowledge of the news. We
also show that knowledge of the news varies by news outlet, with, for example, Google
News users being 1.8 percentage points more likely than Facebook users to know the
typical news story of the month (among voters who rely on 2 or fewer news outlets).
We provide a number of extensions. First, we illustrate the replicability of our
methodology by focusing on a different set of knowledge items. For three months in a
row, we rely on our panel of journalists to select the 3 most important stories of the
month regarding the Democratic Party presidential primaries. In addition to showcasing
9
As noted by Prior [2014], text surveys may exaggerate knowledge inequalities by omitting visual
clues (e.g., by not including pictures of actors mentioned in the news and included in our surveys).
8
the robustness of our method, this extension allows us to measure how much attention
voters pay to a key political event. Second, we replicate our main analysis about the
Federal Government by relying on a different sample of respondents recruited through
Amazon Mechanical Turk (MTurk).
The rest of the paper is structured as follows. Section 2 reviews the news-generating
process and the survey design. Section 3 describes the model as well as our estimation
approach. Section 4 reports our main results. Section 5 presents various extensions of
our analysis as well as robustness checks. Section 6 concludes.
2 Design
The key components in our analysis are knowledge quizzes, in which respondents
are rewarded if they succeed in choosing the true knowledge items included in a list
containing both true and false knowledge items. We review the protocol we have
employed to generate the true and false knowledge items. We also describe the information
we have collected through the surveys.
2.1 News Generating Process
We design a protocol to identify, each month, the three most important news stories
about the US Federal Government.
Universe of Relevant Knowledge Items. We rely on Reuters’ publicly-available

wire stories about US national politics to approximate the universe of relevant knowledge
9
items.10 This choice allows us to focus on essential and basic facts covered by mainstream
media. Each wire story is composed of a headline, a brief summary, a picture, and a
longer article. There are approximately 80 wire stories a week.
Generating 3 True and 3 False Knowledge Items. We rely on a panel of 3

professional journalists recruited through the Columbia School of Journalism. To avoid
recency effects, each week, each journalist is asked to select and rank the 5 most
important wire stories of the week according to him/her.11 Specifically, journalists
are provided with each wire story’s headline and brief summary. Journalists are also
given the url to the longer articles. Because multiple wire stories can deal with the same
underlying “meta story” (e.g., “Coronavirus” or “Trump Impeachment Enquiry”), we
ask the journalists to select only one wire story per meta story.12 At the end of every
month, we take the four previous weeks’ selected wire stories (up to 3×4×5 of them)
and pool them into their relevant meta stories (since different weeks’ wire stories can
deal with the same underlying event). We filter out the wire stories that do not cover
the Federal Government (by far, most stories deal with the Federal Government).13,14
We then present each meta story and associated wire stories to our panel and ask
them to select the three most important meta stories of the month.15 Once the three
10
Reuters’ wires dedicated to US national politics can be found at
https://www.reuters.com/news/archive/politicsNews.
11
Although we give significant discretion to our jury members in selecting the most important
stories, we ask them to adopt US-centered criteria of importance. All jury members are US citizens.
12
Whether two Reuters wire stories belong to the same underlying meta story is often easy to
determine. In the rare occasions where the boundaries of a meta story are blurry, journalists are
allowed to communicate and resolve ambiguity.
13
We adopt the US definition of the “Federal Government” as being composed of the legislative,
executive, and judicial branches.
14
The few stories that do not cover the Federal Government deal with the presidential primaries.
In Section 5, we replicate our analysis by focusing on the Democratic Party presidential primaries.
15
Specifically, we ask each panel member to rank these stories and we then aggregate these choices.
We rely on randomization to break eventual ties.
10
stories are selected, each story is allocated to a journalist who is asked to write a short
statement summarizing the story or summarizing the most important fact behind the
story (e.g., The U.S Senate acquitted Trump of impeachment charges).
Our main instrument to estimate voters’ knowledge of political news consists of
asking them to select three out of six statements. Three of these statements correspond
to the three true statements described in the previous paragraph. The remaining three
statements are false short statements about the Federal Government. We relied on our
panel of journalists to produce these plausible but ultimately false short statements
about the Federal Government. Among other pre-specified rules, journalists were
instructed to write false statements of roughly equal length as the true statements, and
in the same journalistic style.16 For each survey, a journalist was randomly assigned to
select the three fake statements from the list produced by the whole panel.17
Why did we rely on a panel of human journalists to identify top stories, rather than
use some more “objective” machine learning algorithm? One could for instance select
the most clicked stories in aggregators like Google News or the most popular articles
on mainstream media like the New York Times, or use some ranking that is based on
those numbers. But obviously such approach would rely on subjective judgment too,
that of Google News users or New York Times readers, who are likely to be different
in terms of knowledge, partisanship, and taste from other voters. Note that whatever
makes Google News users or New York Times readers more likely to click on a story is
16
We also instructed the panel to avoid writing negations of events that really took place, to avoid
writing statements that could be perceived as related to the real statements, to avoid using numbers
and figures, and to primarily use past tenses. We conducted Google searches to ensure our fake stories
did not actually occur.
17
Notice that we could have relied on fake news that actually circulated (‘real’ fake news), by
for instance using third-party fact-checkers. Although it would be interesting to use our method to
quantify the extent to which voters believe in fake news, in this paper we limit ourselves to measuring
voters’ knowledge of real news.
11
likely to affect their knowledge of that story too, thus biasing all the rest of the analysis.
2.2 Survey Design
This paper exploits data gathered from 8 online surveys we conducted through polling
company YouGov.18 The first survey took place on the 17th of December 2018 and
the last survey on the 17th of February 2020.19 For each survey, we asked YouGov to
enroll a representative sample of the US citizen adult population.20 All surveys were
administered to 1,000 individuals, except for one survey which was administered to 1,500
individuals.21 YouGov is able to draw respondents from a pool of 2 million members in
the US, and it provides a wide array of high-quality background information concerning
each survey respondent (demographics, income, education, ideology, party affiliation,
interest in politics, etc.). Importantly, this information was collected by YouGov long
before respondents took our survey. Responses regarding political preferences or general
attitudes are therefore unaffected by our survey. Additionally, we asked a series of
questions regarding news consumption habits. Our survey took respondents on average
5-6 minutes to complete. Participants received about $1.9 on average (paid via gift
cards) in exchange for completing the survey. Payments included a 50¢ show up fee
18
See https://today.yougov.com/find-solutions/omnibus/ for information on YouGov.
19
Notice our time period does not coincide with a presidential election. Recent research suggests
that it is information acquired over long periods of time (as opposed to during the weeks immediately
preceding an election) that determine most voters’ beliefs [Le Pennec and Pons, 2019].
20
To construct the sample, YouGov employs a two-step procedure. In the first step,
a random sample is drawn from the population (using either Census information or
the American Community Survey). This sample is referred to as the target sample.
In the second step, a matching technique is utilized to match each member of the
target sample with members of YouGov’s pool of respondents. For further details, see
https://smpa.gwu.edu/sites/g/files/zaxdzs2046/f/downloads/YG˙Matching˙and˙weighting˙basic˙description.pdf.
21
We also instructed YouGov to avoid enrolling individuals who participated in prior editions of the
survey (this restriction was lifted during our eighth survey).
12
Statistic YouGov ACS 2018
Median Age 49.00 47.00
% Female 0.52 0.51
% White 0.69 0.73
% Black 0.11 0.13
% 4yr College Degree 0.30 0.31
% Unemployed 0.07 0.06
% Married 0.48 0.48
% Family Inc <30k 0.28 0.17
% Family Inc 30k - 60k 0.20 0.23
Table 1: Socioeconomic Characteristics
and bonuses equal to $1 for each quiz correctly answered.22

Table 1 provides basic descriptive statistics regarding the socioeconomic characteristics
of the survey respondents who participated in all eight surveys.23 It also reports the
corresponding statistics for the population of US adult citizens according to the 2018
American Community Survey of the Census Bureau (ACS).24 All dimensions appear
closely aligned with the general population, with the exception of family income (with
low income brackets being over-represented in the YouGov sample).
Table 2 reports information on the party affiliation of our survey respondents, and
compares it with the statistics provided by Pew [2018].25 For the purposes of this paper,
we pool the respondents who report that they “Lean Democrat” (“Lean Republican”)
22
Our description of the survey is based on the last survey we administered. Some modifications
were introduced as we conducted more surveys. We highlight these modifications when relevant.
23
Respondents who took multiple surveys are counted only once, and the characteristics we use
(e.g., age) are those relevant when they took their first survey.
24
To obtain the 2018 ACS go to https://www.census.gov/programs-surveys/acs.
25
YouGov asks respondents to select one option among “Strong Democrat”, “Not very strong
Democrat”, “Lean Democrat”, “Independent”, “Lean Republican”, “Not very strong Republican”,
“Not sure”, “Don’t know”. About 4% of respondents report either “Not Sure” or “Don’t Know”.
Because our model incorporates political preferences, we pool these respondents with the respondents
who report being “Independent”.
13
Party Affiliation YouGov Pew 2018
% Democrat 45 48
% Republican 35 39
% Independent 16 7
% Other 4 6
Table 2: Party Affiliations
with the respondents who support the Democratic Party (Republican Party). The
proportions are roughly comparable, with the exception of Independents who appear
somewhat over-represented in the YouGov sample.
Our survey was composed of two main parts: (i) a series of questions about media
consumption habits and (ii) a series of questions about recent political news.
2.2.1 Media Consumption Habits
All survey respondents were asked to provide information regarding their recent consumption
of news about US national politics. Specifically, we asked respondents to report whether
they had acquired information about national politics during the previous 7 days, and
whether they acquired it online, by watching television, by listening to the radio, and/or
by reading a print newspaper. We use the resulting information to create the dummy
variables Televisioni , Printi , Radioi , Onlinei . We also create the discrete variable
Mediai , defined as the sum of these 4 dummy variables. For all survey respondents
who selected one or more types of media (e.g., television and online), we further
asked them to report the news sources they relied on to obtain information about
national news (e.g., CNN and Facebook). We used the resulting information to create
the discrete variable News Sourcesi .26 Finally, survey respondents were also asked to
26
Many news sources are available across media (e.g., CNN is available both on television and
online). We consolidated news sources as appropriate.
14
Media 1.77
Television, % 62
Print, % 21
Radio, % 30
Online, % 63
News Sources 4.48
Total Time (minutes) 346.97
Table 3: Media Consumption Habits
report the amount of time they dedicated to getting information about national politics
(again, during the previous 7 days). We used this information to code the variable
Timei . Tables B.1 and B.2 in Online Appendix B present the language used in the
corresponding survey questions.
Table 3 reports summary statistics regarding our main media consumption variables.
Our average survey respondent relies on roughly 1.8 media, and television and internet
are by far the most popular media (both are consumed by roughly 60% of our survey
respondents). Further, the average respondent relies on roughly 4.5 news sources to
obtain information and consumes over five hours a week of national news.
2.2.2 Knowledge of the News
All surveys included one or two knowledge quizzes about current news stories (less than
four weeks old). In a number of surveys, we also included one-month and two-month
old knowledge quizzes to the study the effect of time.27 Overall, we included 11 distinct
knowledge quizzes in our eight surveys. Table B.13 in Online Appendix B reports how
the various quizzes were allocated to the various surveys we administered. Each quiz is
27
In the last survey, each respondent took one quiz about the Federal Government, one quiz about
the Democratic Party presidential primaries, and one quiz containing two-month old events about either
the Democratic Party presidential primaries or the Federal Government (respondents were randomly
allocated to a topic).
15
composed of 6 short statements (where the ordering of the statements was randomized
across respondents). Survey respondents were told the list contained exactly 3 true
statements and 3 false statements. Respondents were asked to select which 3, to the
best of their ability, were the correct statements. To avoid individuals from obtaining
information somewhere else, respondents were given 60 seconds to make their selection
(a timer was added from the second survey onward to help respondents estimate the
amount of time they had left). Whereas no monetary incentives were given during the
first survey (in addition to the base compensation), from the second survey onward
we offered an extra $1 (paid via a giftcard) to all respondents who selected all three
correct statements. All survey respondents were revealed the correct answers once
they took the quiz. Tables B.3-B.10 in Online Appendix B include all quizzes that we
administered through our series of surveys. Across all surveys and quizzes, our average
survey respondent selected approximately 2.17 true statements.28
In the last four surveys, we also asked our survey respondents to report their feelings
towards the six statements contained in the quiz they completed. Specifically, for each
true statement, respondents were asked how favorably, in their opinion, the statement
reflected on the Republican Party. Similarly, for each false statement, respondents
were asked how favorably, in their opinion, the statement would have reflected on
the Republican Party had it been true. Respondents were allowed to select one option
among “very unfavorable”, “unfavorable”, “neither unfavorable nor favorable”, “‘favorable”,
and “very favorable”. We used the resulting information about the average respondent’s
feeling toward statement j to create the continuous variable bj ∈ [−∞, ∞].29 Across
28
Presumably because of the 60-second limit, some respondents ended up selecting strictly fewer
than 3 statements. A tiny share of respondents also selected strictly more than 3 statements. Overall,
about 19% of respondents selected a number of statements different from 3. We exclude these
respondents from our analysis and discuss the potential biases this exclusion introduces when relevant.
29
To construct it, we first map the answers such that “neither unfavorable nor favorable” is
16
all surveys and quizzes, the average true statement has b = −0.11, that is, the average
survey respondent felt that the average true statement reflected slightly unfavorably
on the Republican Party. Similarly, across all surveys and quizzes, the average false
statement received a score of b = 0.16, that is, the average survey respondent felt that
the average false statement reflected slightly favorably on the Republican Party. Tables
B.11 and B.12 in Online Appendix B present the language used in the corresponding
survey questions.
2.2.3 Discussion
We could have designed our survey in a number of alternative ways. For example, we
could have made it such that each respondent’s task consisted of either (i) determining
whether each statement in a list of 6 statements was true or false or (ii) choosing
the 3 true statements included in a list of 3 pairs containing each 1 true and 1 false
statement. The first alternative is formally identical to our setting if respondents are
told that exactly half of the 6 statements are true and if they are allowed to read all 6
statements before making up their mind. Precisely anchoring respondents’ beliefs about
the number of real and false statements is desirable for the purposes of estimating our
model (see Footnote 32). Moreover, not allowing respondents to read all 6 statements
before determining which 3 are more likely to be true (that is, forcing respondents to
declare statements as true or false sequentially) would prevent them from fully utilizing
their knowledge, which we are attempting to measure. In other words, under the
first alternative to our setting, we would make assumptions and choices that would
render it formally identical to our current setting. Moreover, the second alternative is
represented by 0, “very unfavorable” is represented by -1 and “very favorable” is represented by 1.

Then, for each statement, we take the average of this measure across respondents, and rescale the
resulting variable to have a standard deviation equal to 1.
17
dominated by our quiz format because it would generate a distribution in the number
of true statements selected by individuals with given knowledge levels with greater
variance, which would make it harder for our model to measure knowledge precisely.
Last, an even more obvious alternative to our quiz design consists of asking respondents
to report directly the confidence they ascribe to each statement being true. Precisely
eliciting such beliefs is notoriously difficult and we prefer to indirectly infer them by
asking respondents to solve a task.
Finally, our method measures aided recall. We measure respondents’ ability to assess
the plausibility of various knowledge items. An alternative approach would consist of
(i) informing respondents of the existence of a list of 3 true knowledge items about
recent political events selected by a panel of mainstream journalists and (ii) asking
respondents to guess these knowledge items.30 Although this alternative approach may
also be used to measure respondents’ knowledge of the news, our model would have to
be extended to take a number of indirect steps into account (e.g., modeling the ability
to second-guess the journalists’ opinions).
3 Model
We develop our model in three steps. We begin with formulating the basic general
problem an agent faces when she is trying to assign a probability of truth to a statement,
which is a standard application of Bayesian binary hypothesis testing. In the second
step we consider an agent who is asked to pick the statement that is most likely to
be true out of a set of statements and we show that, under standard assumptions, the
problem corresponds to a familiar parameterized discrete choice problem. Finally, we
30
We thank Miklos Sarvary for this comment.
18
apply this theoretical framework to the survey instrument we are using to arrive at the
econometric model that we will be using in the rest of the paper. In the last subsection,
we clarify the link between our model and the existing statistical literature.
3.1 The News Knowledge Problem
Suppose agent i is trying to establish the truth of statement j, which we call τj ∈

{0, 1}, where 0 represents a false statement and 1 a true statement. The agent observes
a signal yij about the statement. For simplicity, assume the signal is continuously
distributed and has full support on the real line. The signal’s conditional distribution
depends on the truth τj , on the agent’s knowledge precision θi , on the statement’s
characteristics aj (e.g., straightforwardness, salience, or familiarity), and on the number
of months t since the story became true:
h i
f yij |τj ,θi , aj , t
The agent is also endowed with a prior probability that the statement is true, which
depends on the statement’s partisanship bj , and on the agent’s party affiliation γi :
h i
g τj |γi , bj
The agent’s posterior probability that the statement is true is given by:
h i h i
h i f yij |τj = 1, θi , aj , t g τj = 1 | γi , bj
Pr τj = 1|yij = h i h i h i h i
f yij |τj = 1, θi , aj , t g τj = 1|γi , bj + f yij |τj = 0, θi , aj , t g τj = 0|γi , bj
Suppose we wish to know whether the agent believes the statement is true with at
19
least probability h ∈ (0, 1). The relevant condition is:
h i h i
f yij |τj = 1, θi , aj , t g τj = 0|γi , bj h
h i ≥ h i , (1)
f yij |τj = 0, θi , aj , t g τj = 1|γi , bj 1 − h
or:
h i h i
ln f yij |τj = 1, θi , aj , t − ln f yij |τj = 0, θi , aj , t
h i h i
≥ ln g τj = 0|γi , bj − ln g τj = 1|γi , bj + H,
where H = ln h/(1 − h) . The left-hand side of the inequality is a function of random

h i
variable yij . As yij is in turn distributed according to f yij |τj ,θi , aj , t , we can write
the left-hand side as xij , a real-valued random variable distributed according to some
h i
f˜ xij |τj , θi , aj , t . The first part of the right-hand side is a deterministic function of γi
h i
and bj , which we write as g̃ γi , bj . Thus, the agent assigns at least probability h to
statement j being true if:
h i
xij ≥ g̃ γi , bj + H. (2)
Let F̃ be the cumulative distribution function of f˜. For any level h, the probability
h i
that the agent assigns at least probability h to statement j is 1 − F̃ g̃ γi , bj + H .
This expression is a characterization of the agent’s belief in the truth of statement j in
terms of the threshold h and the underlying parameters γi and bj .
3.2 A Discrete Choice Model
We now make a number of functional form assumptions that lead to a tractable and
familiar logit specification. Assume that the random variable on the left-hand side of
20
(2) can be written as:

xij = 2τj − 1 aj θi δ −t − εij ,
where εij follows an extreme value distribution of type I. Recall that we interpret θi ≥ 0
as agent i’s knowledge precision and aj ≥ 0 as the straightforwardness (the contrary of
difficulty) of the news story. The parameter δ captures the effect of time passing, with
t = 0, 1, · · ·.
Also assume the prior term can be written as
h i
g̃ γi , bj = −αbj γi .
Again, recall that we interpret bj ∈ (−1, 1) as the partisanship of the news story: a high
(low) bj denotes a story that reflects favorably (unfavorably) on the Republican Party.
Similarly, γi ∈ {−1, 0, 1} denotes agent i’s party affiliation, where γi = 1 (γi = −1)
means that agent i identifies with the Republican Party (Democratic Party) and γi = 0
means that agent i identifies as Independent. The term bj γi captures the tendency of
voters to believe statements that agree with their ideology and the parameter α ≥ 0
measures the strength of this effect.
This formulation is equivalent to agent i assigning to statement j a plausibility value

zij = 2τj − 1 aj θi δ −t + αbj γi − εij

The plausibility value is a random variable with support (−∞, ∞), and mean 2τj − 1 aj θi δ −t +
αbj γi .
Now suppose the agent is given a set J of statements and asked to pick the one
that is most likely to be true. Each statement j is characterized by its truth τj ,
21
its straightforwardness aj , and its partisanship bj . The error term is i.i.d. across
statements. The agent will select the statement with the highest plausibility value zj .
This is similar to a standard logit discrete choice model and it leads to the following:
Proposition 1 The probability agent i believes statement j is the most likely to be true
among the set J of statements is

−t
exp 2τj − 1 aj θi δ + αbj γi
πij = P . (3)
k∈J exp (2τk − 1) ak θi δ −t + αbk γi
The comparative statics of the expressions above are intuitive.31 If j is a true

(false) statement, πij is increasing (decreasing) in i’s knowledge precision θi and j’s
straightforwardness aj . As the agent becomes infinitely knowledgeable (θi → ∞) or the
statement becomes infinitely easy (aj → ∞), the probability tends to 1 if the statement
is true and to zero if it is false.
Recall that the location parameter of the standard Gumbel distribution does not
h i
affect πij . To fix a value for the location parameter, we assume that E εij is such that
the probability that the agent places at least a 0.5 probability on a totally uninformative
h i
statement (aj = 0) on which she has no prior view (bj = 0) is 0.5 (i.e., Pr εi,j ≤ 0 =
0.5). This requires setting the location parameter of the Gumbel distribution at −λ,
where λ is Euler’s constant (approximately 0.57), and it implies that the CDF of εij
−λ
is Φ εij = e−e . Thus, the probability that agent i assigns at least probability h to
31
The expression above holds under the assumption that the random variable εij is independent
across the six statements. In practical terms, this means that the statements are not related in ways
that make their plausibility value correlated. An obvious violation occurs when two statements refer to
related stories “President Trump visited France” and “President Trump met with Emmanuel Macron.”
We believe the independence condition is satisfied in practice within every round as both the true
stories and the fake stories are designed to belong to distinct meta-stories (see Section 2).
22
statement j being true is given by:

−t
exp 2τj − 1 aj θi δ + αbj γi
ρij (h) = . (4)
h
exp 2τj − 1 aj θi δ −t + αbj γi + 1−h
3.3 Econometric Model
In our survey quizzes, respondents are given 6 statements (ordered randomly). They
are told that exactly 3 statements are true and they receive $1 if they successfully
select these 3 true statements. This creates some mechanical correlation between
answers. For instance, if I think that one statement is true and I know that only three
statements are true, then I must be more pessimistic about the other statements. This
mechanical correlation is fully incorporated in the estimation procedure. Intuitively, the
information that exactly three statements are true does not affect the optimal strategy
of the respondent: pick the three statements that are most likely to be true. More
formally, proceed as follows. Assume each respondent maximizes the probability of
receiving the monetary reward. Let T ≡ (τ1 , τ2 , τ3 , τ4 , τ5 , τ6 ) ∈ {0, 1}6 be the set of all
possible ‘truth vectors’ over the six statements. The respondent first observes all six
signals, and is thus capable of computing the posterior probability associated to any
element of T . Let T3 denote the subset of T whose elements sum up to exactly 3, and
T\3 its complement. Using the posterior probabilities obtained after observing his/her

6 signals, the respondent computes the probabilities Pr (T3 ) and Pr T\3 . Next, the
respondent incorporates the fact that exactly 3 statements are true by using Bayes
rule (and that the ordering of the statements was randomized according to a discrete
uniform distribution). Specifically, he/she selects the 3 statements j, j 0 , j 00 with the
23
highest associated:
h i h i h i h i h i h i
Pr yij Pr y ij 0 Pr y ij 00 1 − Pr y ij 000 1 − Pr y ij 0000 1 − Pr y ij 00000
. (5)
Pr (T3 )
This is formally equivalent to the respondent choosing the 3 statements with the 3
highest associated values. For the purposes of our estimation exercise, we rely on
the probability of selecting any 3 statements {j, j 0 , j 00 } for all possible orderings of
the plausibility values associated to the statements j, j 0 , and j 00 . Given our logit
specification, the probability of selecting statements {j, j 0 , j 00 } in this exact order is
given by: πij∈J · πij 0 ∈J\{j} · πij 00 ∈J\{j,j 0 } .
Our objective is to estimate, for each respondent i, a posterior distribution of
knowledge precision θi ∈ R and, for each statement j (whether true or false), posterior
distributions of aj ∈ R. In addition, we estimate the posterior distributions of population
parameters δ ∈ R+ and α ∈ R.32,33
In what follows let g ∈ G denote a socioeconomic group, where groups are defined
as intersections of 4 demographic characteristics: Age (below/above median), Gender,
Family Income (below/above median), and race (white and minority).
We estimate the model by Bayesian methods, specifically Hamiltonian Monte Carlo
[Hoffman and Gelman, 2014] implemented in Stan [Carpenter et al., 2017]. To that
end, we specify common prior distributions θi ∼ N (µg , σ 2 ) and aj ∼ N (0, 1), with
32
Notice that, to be meaningful, the economic model requires condition θ ≥ 0. In our estimation
exercise, we do not impose this constraint. As we report below, the posterior distribution of θ we
recover has negligible mass below 0.
33
A common problem with this family of models [e.g., Bock, 1972, and see discussion of the
literature below] is that θ and a are identified through their product, so that there always exists one
additional degree of freedom. This problem is solved by “anchoring” one of the two variables to some
arbitrary scale. Consistent with our Bayesian approach, in our analysis the anchoring is achieved by
assuming that a is distributed according to a standard normal.
24
hyperpriors µg ∼ N (0, 10), ∀g, and σ ∼ exp( 14 ).34 The remaining prior distributions
are specified as α ∼ N (0, 10) and δ ∼ N (1, 1). Notice that we allow for varying group-

level means for the prior distribution of θ (i.e., θi|i∈g ∼ N µg , σ ).35
The key identifying assumption is that the processes that generate the a’s and
the θ’s are stochastically independent.36 While some months our panel of journalists
selects real and fake stories that are easier or harder and YouGov selects better or worse
respondents (though that is less likely, given our sample size), what is required is that
these two sources of variations are not systematically correlated.
We propose a three-step procedure to estimate the parameters of the model. In step
1, we arbitrarily fix θi = 1, ∀i ∈ I, and estimate the remaining parameters. In step 2,
we fix aj to equal its posterior mean from Step 1 and estimate µg , ∀g, and σ. Finally,
in Step 3, we fix µg , ∀g, and σ at their posterior means from Step 2 and reestimate
(θi )i∈I , aj , α, and δ.
We conclude with some final remarks. In the last four surveys, we separately asked
the respondents to report their sentiment towards all true and false statements. We
can thus directly use this information to create the variable bj . Because we did not
ask these questions in the first four surveys, we first estimate our model by relying on
an alternative measure of congruence. Specifically, for each statement we compute the
34
P6 P6
Following Bock [1972] we impose the restriction that j=1 aj = 0 by fixing a6 = − j=1 aj .
In the absence of this restriction, one could add any constant to all the a’s without affecting the
probability of selecting a given statement j.
35
An alternative approach consists of assuming a common mean for the prior distribution of θ across
all individuals and groups. Such an approach would be rather conservative when quantifying knowledge
differences across groups. Given the limited data available at the individual level, the posterior
distributions of individual knowledge θi have a relatively large variance. As a direct consequence,
the common prior assumption would pull individual estimates toward the mean. Nevertheless, results
under this alternative approach are very similar to those with group-level means, with the exception
of our results on inequalities (with smaller differences across groups).
36
This type of mutual dependence between questions is obviously different from the mechanical
dependence discussed above.
25
difference between the share of Republicans and Democrats who selected that statement
and normalize that variable to have a standard deviation equal to 1. Although this
approach suffers from a possible reverse causality problem, we first use it in our main
analysis. Later on, we will restrict our attention to the last four surveys and rely
exclusively on the separately-observed measure of congruence.
3.4 Literature Discussion
The model we develop here is loosely related to Item Response Theory (IRT), a set
of statistical models that are used to analyze test results with the objective of inferring
the difficulty of the test questions and the traits of the test takers [Van der Linden
and Hambleton, 1997]. However, we face two important differences with standard
approaches in this literature.
In standard IRT applications such as the Rasch model [Rasch, 1960], the researcher
can rank alternatives a priori (usually because an answer can only be right or wrong).
Here, instead we cannot a priori rank different statement bundles that contain different
subsets of true statements. Suppose that A, B, and C are true statements and D, E,
and F are false statements: it is not ex ante clear whether choosing, say, (A, B, D) is
better than choosing (A, C, E). We are closest to an extension of IRT called Nominal
Response Model (NRM), developed by Bock [1972], which allows items to be ranked in
a partially unknown manner.
However, we cannot use any of the IRT models, including NRM, directly because of
one important difference. The objective of all IRT tests is to measure the underlying skill
of test takers. Instead, we are interested in measuring two factors: the underlying skill
of our respondent (the precision of their signal) and the effect of partisan congruence.
26
The latter effect is well-known to be important in political knowledge but it is not
salient in educational testing.
We therefore must augment Bock [1972] by developing a model where individuals
have two traits, skill and ideology, and news stories have two characteristics, difficulty
and partisanship. The combination of ideology and partisanship determines response
rates in a non-monotonic way: it increases or decreases the probability that a person
chooses a certain true or false statement depending on the congruence between the
person’s ideology and the statement’s partisanship.
4 Analysis
4.1 Knowledge of the News
Within our framework, the probability that individual i with knowledge precision
θi assigns a probability equal to or higher than h to statement j being true is equal to
ρij (h) (see (4)). Our first results shed light on the average voter’s knowledge of political
news. For each statement j and individual i, and for any confidence level h, we can
compute the posterior distribution of ρi,j (h) as well as its average. In particular, let
F (θ) represent the posterior distribution of θ in the sample. One can then compute
R 1 P P
θ∈R ρ(h|θ)dF (θ), whose empirical analog is given by IN i n ρ(h|θi,n ) (where I is the
number of individuals and N is the number of draws from the posterior distribution
of θi ).37 Figure 1 plots the average value of ρi,j (h) for all values of h ∈ [0, 1], by
distinguishing between the top 3 stories of the month about the Federal Government.
37
We refer to the average voter for simplicity. Formally speaking, though, we compute the average
probability that a voter selected at random according to F (θ) assigns probability h or higher to a
statement being true.
27
Figure 1: Knowledge of the News
Recall that the ranking of news stories by importance is provided by our panel of
journalists. Even within a given rank (say, first story of the month), however, the
properties of the news stories –as captured by aj – may vary from one month to the
next. To address this issue, within each rank, we take the average of the mean of the
posterior distribution of aj across stories. We also suppose this fictitious typical story
to be neutral in its partisanship (i.e., we set b = 0).
Table 4 reports the average voter’s knowledge of the first, second, and third news
story of the month about the Federal Government, for various intervals of confidence.
To report our results in a way that is easier to comprehend, it is useful to focus on a
particular level of confidence h. In what follows, we say that an individual knows a
(true) statement if he/she assigns a probability h ≥ 0.75 to the statement being true.
Similarly, we say that an individual is uncertain about the veracity of a news story if she
assigns a probability of truth between 0.25 and 0.75, and that she believes the story to
be false if she assigns a probability of truth lower than 0.25. Accordingly, the top panel
of Table 4 reports the corresponding figures. For the first news story of the month,
the probability that the average voter knows the story is equal to 66%. Similarly, the
28
probability that the average voter is uncertain (i.e., h ∈ (0.25, 0.75)) is equal to 27%,
and the probability that the average voter believes the story to be false is 6%. These
numbers change as we move from the first to the second and third stories of the month.
For example, the probability that the average voter knows the second and third typical
story falls to 50% and 38%, respectively. Reassuringly, therefore, the ranking of news
stories by our panel of journalists is reflected in voters’ knowledge of these stories.
Naturally, saying that a voter “knows” a news story if she assigns a probability at
least as high as 0.75 to the story being true is arbitrary. The second and third panels of
Table 4 report similar figures for alternative definitions of knowledge. For example, in
the second panel, we report that the average voter is 74% likely to attribute 2 to 1 odds
to the first story of the month being true. The corresponding figures for the second and
third news stories of the month are 60% and 48%, respectively. Last, the third panel
of Table 4 reports the likelihood that the average voter attributes a probability greater
than or equal to h = 0.5, 0.6, 0.7, 0.8, 0.9 to the first, second, and third news stories of
the month being true. Strikingly, the average voter has a 43% chance of being very
confident (h ≥ 0.9) about the first story of the month.
An alternative approach to expressing voters’ knowledge of political news consists of
computing the expected number of news stories – among the top 3 stories of the month
– known by voters. In addition to being directly interpretable, this way of measuring
knowledge is also particularly amenable to quantifying differences across voters. In
what follows, we rank individuals by their associated knowledge precision θi and report
results for the average individual belonging to the bottom-third, middle-third, and top-
third of the knowledge distribution. In particular, Table 5 reports the probability that
the average member of these three groups knows the typical first, second, and third
29
Confidence First story Second story Third story
0 - 0.25 0.06 0.11 0.16
0.25 - 0.75 0.27 0.39 0.46
0.75 - 1 0.66 0.5 0.38
0 - 0.33 0.09 0.15 0.22
0.33 - 0.66 0.17 0.25 0.3
0.66 - 1 0.74 0.6 0.48
0.5 - 1 0.84 0.74 0.64
0.6 - 1 0.78 0.66 0.55
0.7 - 1 0.71 0.56 0.44
0.8 - 1 0.61 0.43 0.32
0.9 - 1 0.43 0.26 0.17
Table 4: Knowledge of the News
news story of the month.38 The reported numbers are suggestive of relatively large
heterogeneity in knowledge across voters. For example, whereas the average voter in
the top-third of the distribution has a 73.8% chance of knowing the first news story of
the month, the average voter in the bottom-third has only a 42.9% chance of knowing
the story. Using these numbers, one computes that – of the top three news stories of the
month – the average voter in the bottom-third of the distribution knows approximately
1.1 stories, the average voter in the middle-third knows approximately 1.4 stories, and
the average voter in the top-third knows close to 1.7 stories.
We conclude this subsection by reporting the posterior distribution of θ that we
recover (see Figure 2). One somewhat striking feature of F (θ) is its relatively low mass
close to zero. Our estimates suggest that very few individuals are uninformed or close
to uninformed. This finding may is easily explained by some basic patterns in the raw
data. Across all quizzes, only 3% of respondents selected 0 true statements and only
38
Again, by typical story we mean a story whose associated parameter a corresponds to the average
value of the mean of the posterior distributions of aj across relevant stories.
30
Knowledge tier
Lower Middle Higher
First story 0.431 0.61 0.737
Second story 0.355 0.465 0.563
Third story 0.292 0.337 0.382
Table 5: Average Probability that an Individual in a Given Tier Knows a Statement
14% selected 1 true statement. By way of comparison, an uninformed individual, with

no choice but to randomize uniformly, chooses 1.5 correct statements on average.39,40
The same individual has a probability equal to 0.05 to choose 0 true statements and
a probability equal to 0.45 to choose one true statement. The theta distribution that
fits the data cannot place a large weight on individuals that have little or no ability to
discern truth.
Figure 2: The posterior distribution of θ
39
For simplicity, we here ignore partisan biases. Incorporating these would not change the logic of
our argument, which is that our average respondent does not appear to be uninformed.
40
Moreover, because each individual completes only but a few quizzes, the variance of the
distribution Fi (θ) is relatively large, so that the common prior assumption tends to pull all individuals
upward. Further, the restriction to respondents who selected exactly 3 statements may also in part
explain the relatively small mass around 0.
31
4.2 Heterogeneity across News Stories
Next we explore various dimensions of heterogeneity across news stories. Table

6 lists all true statements that were included in the various quizzes and surveys.
Similarly, Table 7 lists all false statements that were included in the various quizzes and
surveys. For each statement, the tables report the date, the share of survey respondents
who selected the statement when completing the quiz (“raw mean”), the mean of the
posterior distribution of aj , the predicted share of respondents who – according to our
model’s estimates – will select the statement when completing the quiz, as well as the
probability that the average voter assigns probability h to statement j being true (where
we distinguish between three intervals of confidence h).
As suggested by the tables, there exists significant heterogeneity across news stories
(within both the true and false statements). Some statements were selected by virtually
all our survey respondents and others were selected only by a tiny share of respondents.
Recall that the parameter aj captures how responsive the likelihood of selecting statement
j is to knowledge θ. What the tables suggest is that some true statements are much more
easily detectable as true by knowledgeable respondents than others.41 Similarly, some
false statements are much more easily detectable as false by knowledgeable respondents
than others.
Next, the tables report, for each statement, the predicted probability that the
average voter selects it when completing the quiz (taking into account the characteristics
of the remaining 5 statements that were included in the same quiz). As suggested by
the numbers, our model approximates the actual data well, irrespective of whether a
statement is chosen by few or many respondents.
41
For two statements, being more knowledgeable was seemingly a disadvantage.
32
Finally, the tables suggest that there exists significant heterogeneity across news
stories regarding respondents’ knowledge. For example, the average voter has a 74%
probability of knowing the (true) story “The US Senate acquitted Trump of Impeachment
Charges.” By contrast, it knows the (true) story “Supreme Court granted a request
by President Trump’s administration to fully enforce a new rule that would curtail
asylum applications by immigrants at the U.S.-Mexico border” only with probability
35%, despite 69% of our sample selecting the statement when completing the quiz.
This last news story – with its misleadingly high share of selections – illustrates how
our structural approach takes into account the various properties of all the knowledge
items included in the quiz to identify voters’ actual knowledge of each single item.
In particular, our model often finds a significant difference between the probability of
knowing a story and the probability of selecting a story when completing a quiz (as the
example above illustrates).
Reassuringly, none of the false statements we included in our quizzes are widely
believed to be true. In fact, the vast majority of our false statements are believed to be
true by fewer than 15% of respondents and none are believed to be true by more than
29% of respondents. Some false stories are more widely believed that some true stories.
4.3 Effect of Partisanship
The model we estimate allows for multiple dimensions of heterogeneity across news
stories. One dimension of particular interest is the extent to which a story reflects
favorably on the Republican Party: Is voters’ knowledge of political news skewed
towards those stories that reflect most favorably on their preferred political party [e.g.,
33
ρ
Statement Month Raw Mean a Prob of selecting < 0.25 ∈ (0.25, 0.75) > 0.75
At a closed-door meeting at the White House, top envoy to China delivered evidence Sep 19 0.36 -0.41 0.35 0.35 0.48 0.18
of rising Farm Belt frustration over bio-fuel policy.
The U.S. Supreme Court gave itself another chance to make a definitive ruling on Dec 18 0.41 -0.26 0.39 0.31 0.49 0.2
electoral map disputes
Vice President Mike Pence visited Nebraska to take stock of the devastation unleashed March 19 0.52 0.31 0.59 0.19 0.49 0.32
across the U.S. Midwest by floods.
The Trump administration credited cooperation from Mexico and Central American Sep 19 0.63 0.62 0.67 0.15 0.45 0.4
countries in cracking down on migrants.
President Trump proposed plan to make U.S. immigration more merit-based. May 19 0.66 0.47 0.67 0.17 0.47 0.36
Trump and Democrats agree to pursue $2 trillion Infrastructure Plan April 19 0.67 0.47 0.68 0.17 0.47 0.36
U.S. lawmakers to unveil revised criminal justice bill in push for final passage Nov 18 0.68 0.45 0.68 0.17 0.47 0.36
Supreme Court granted a request by President Trump’s administration to fully enforce Sep 19 0.69 0.44 0.7 0.17 0.47 0.35
a new rule that would curtail asylum applications by immigrants at the U.S.-Mexico
border.
U.S. Senate hands Trump rebuke on Saudi Arabia Nov 18 0.7 0.56 0.72 0.15 0.46 0.39
Mexico agreed to take more migrants seeking asylum in the United States while they May 19 0.7 0.62 0.73 0.15 0.45 0.4
await adjudication of their cases.
Republican lawmakers in the House of Representatives condemned President Trump’s Sep 19 0.75 0.74 0.73 0.13 0.43 0.44
decision to withdraw troops from Syria.
Homeland Security Secretary Nielsen resigns amid Trump anger over border April 19 0.78 0.79 0.8 0.13 0.42 0.45
President Donald Trump vetoed the measure passed by Democrats and Republicans March 19 0.8 0.49 0.67 0.16 0.47 0.37
in Congress to end his emergency declaration on building a border wall with Mexico.
Special Counsel Robert Mueller did not find the Trump 2016 campaign knowingly March 19 0.82 1.03 0.86 0.1 0.38 0.52
conspired with Russia.
Democratic lawmakers called for further investigation into a revelation that in 2016 Dec 18 0.84 0.98 0.87 0.11 0.39 0.5
Paul Manafort gave polling data to a man linked to Russian intelligence
Rod Rosenstein, U.S. deputy attorney general who appointed Special Counsel Robert April 19 0.84 1.1 0.89 0.1 0.37 0.53
Mueller, submits resignation
The House of Representatives passed legislation seeking to rein in President Trump’s Oct/Nov 19 0.84 0.99 0.87 0.11 0.39 0.51
ability to deploy U.S. forces to fight abroad
Attorney General William Barr said that President Trump’s attacks on prosecutors, Oct/Nov 19 0.87 1.18 0.9 0.09 0.35 0.55
the judge and jurors in the trial of Roger Stone undermined the Justice Department’s
work
Former Trump lawyer Michael Cohen sentenced to three years prison Nov 18 0.88 1.29 0.92 0.08 0.33 0.58
Alabama’s governor signed a bill to ban nearly all abortions in the state. May 19 0.9 1.46 0.94 0.07 0.3 0.62
Whistle-blower report complains of White House cover-up on Trump-Ukraine scandal. Sep 19 0.9 1.41 0.95 0.08 0.31 0.61
The U.S. Government was partially shut down in fight over Trump’s border wall with Dec 18 0.94 1.46 0.95 0.07 0.31 0.62
Mexico
A whistleblower filed a complaint against President Trump, leading to an Sep 19 0.94 1.97 0.98 0.05 0.23 0.72
impeachment inquiry.
The U.S Senate acquitted Trump of impeachment charges Oct/Nov 19 0.95 2.1 0.99 0.05 0.21 0.74
Table 6: True Statements
Bénabou and Tirole, 2002]?42 If so, to what extent? The model we estimate assumes
that all voters are possibly biased along partisan lines in their baseline knowledge of
the news, and that the extent of the bias (captured by the parameter α) is identical
across voters.
We elicited respondents’ feelings towards the news only from the 5th survey onward
(see Section 2). To use all 8 surveys, we must thus proxy stories’ partisanship differently.
We proxy the extent to which a news story reflects favorably on the Republican Party
42
Throughout, we rely on the bipartisan nature of American politics to assume that a story that
reflects favorably on the Republican party must reflect unfavorably on the Democratic Party. Similarly,
we assume that a story that “neither reflects favorably nor unfavorably” on the Republican Party does
not reflect either favorably or unfavorably on the Democratic Party either.
34
ρ
A Tape surfaced of President Trump supporting abortion Oct/Nov 19 0.07 1.87 0.05 0.71 0.24 0.05
President Trump’s Tax Returns showed billions given to various charities. Sep 19 0.09 2.33 0.03 0.78 0.18 0.04
Mitt Romney decided to run for president against Trump in the 2020 race after Oct/Nov 19 0.11 1.43 0.07 0.61 0.31 0.08
breakout role in impeachment
2020 Presidential Candidate Elizabeth Warren took millions in Wall Street campaign March 19 0.13 1.33 0.11 0.59 0.33 0.08
contributions.
Trump administration to continue to allow U.S. research using fetal tissue from May 19 0.13 1.45 0.1 0.62 0.31 0.07
abortions.
President Trump took a week-long break from Campaigning to Deal with Coronavirus Oct/Nov 19 0.15 0.98 0.12 0.5 0.39 0.11
Outbreak
Trump secures funding for border wall in meeting with top Democrats Nov 18 0.17 1.24 0.13 0.57 0.34 0.09
ISIS beheaded three Americans in response to Al-Baghdadi’s death. Sep 19 0.17 1.24 0.1 0.57 0.34 0.09
Attorney General Barr released text message from Special Counsel prosecutor Robert May 19 0.19 1.01 0.16 0.51 0.39 0.11
Mueller: ’We’re taking down Trump.’
Trump fired Federal Reserve Chairman Jerome Powell for raising interest rates Dec 18 0.2 1.17 0.15 0.55 0.35 0.09
Trump releases redacted version of his taxes to Congress April 19 0.21 1 0.16 0.51 0.39 0.11
Soybean farmers marched on Washington over Chinese tariffs’ impacts Dec 18 0.22 0.83 0.21 0.46 0.42 0.12
China blacklists Apple and Microsoft amid escalating trade war. Sep 19 0.23 0.83 0.23 0.46 0.42 0.12
Saudi Crown Prince To Address Senate In Effort To Clear His Name In Journalist?s Nov 18 0.25 0.69 0.23 0.42 0.44 0.14
Murder
Clinton Foundation loses nonprofit status April 19 0.25 0.63 0.24 0.41 0.45 0.14
The Virginia Bar Association disbars Attorney General Barr for lying to Congress April 19 0.25 0.73 0.22 0.43 0.43 0.13
President Donald Trump diverted Puerto Rico aid to fund the border wall with March 19 0.29 0.68 0.23 0.42 0.44 0.14
Mexico.
Federal Judge rules public funding for Planned Parenthood unconstitutional Nov 18 0.32 0.37 0.32 0.34 0.48 0.18
President Trump announces he will resume peace talks with Iran at UN General Sep 19 0.37 0.39 0.36 0.34 0.48 0.18
Assembly.
Trump Threatened To Raise Border Wall Cost To $7 Billion If Stall By Democrats Dec 18 0.39 0.18 0.42 0.29 0.5 0.21
Continues
China and the United States agreed on a new comprehensive trade deal. Sep 19 0.41 -0.23 0.49 0.2 0.49 0.3
U.S. Border Patrol facility admitted to measles outbreak among migrant children in May 19 0.42 0.09 0.41 0.27 0.5 0.23
custody.
House Republicans Unveil Legislation To Significantly Limit Funding To Planned March 19 0.44 -0.18 0.53 0.21 0.5 0.29
Parenthood Centers Nationwide.
Vaping case to make its way to Supreme Court. Sep 19 0.44 0.22 0.42 0.3 0.49 0.21
Table 7: False Statements
by using the difference between the share of Republican respondents and the share of
Democratic Respondents who selected the story when completing the quiz. Moreover,
we normalize this measure to have a variance equal to 1. We then rank the statements
according to their partisanship measure bj , and select statements within given percentile
ranks: the 10th, 25th, 50th, 75th, and 90th percentile. Statements with low (high)
values of bj are likely favorable to the Democratic (Republican) party.
Figure 3 plots the posterior distribution of the population parameter α. The
congruence parameter is rather tightly estimated away from zero, suggesting the presence
of a partisanship effect. Table 8 reports, for various percentiles in the distribution of
bj , the probability that a supporter of given party attributes a given probability to a
statement being true (lower than 25%, between 25% and 75%, and higher than 75%).
35
Figure 3: The posterior distribution of the congruence parameter α
As news stories reflect less favorably on the Republican Party, the share of Republican
respondents who attribute a probability of truth greater than or equal to 75% falls.
Not surprisingly given that we assumed α to be a population parameter, the effect is
symmetric for Democratic respondents. To quantify the magnitude of this effect, we
define Partisan Gap as the difference in the average ρ (h) across supporters of a given
party, between Republican and Democratic party, normalized by the corresponding
value for the Independent respondents. By this metric, for example, supporters of the
Republican Party are 18.39% more likely than supporters of the Democratic Party to
know a story located on the 90th percentile of the distribution (i.e., a statement that
reflects rather positively on the Republican Party). Similarly, Republican respondents
are 17.15% less likely to know stories that reflect poorly on the Republican Party (i.e.,
stories located on the 10th percentile).
The approach highlighted above suffers from a possible reverse causality problem.
We thus replicate our analysis on the last four surveys, using the measure of bj separately
elicited from our survey respondents (see Section 2). Table 9 reports the corresponding
36
Congruence Confidence
0 − 0.25 0.25 − 0.75 0.75 − 1
Republican 0.11 0.39 0.5
Strongly Pro-Republican (90th pct) Democrat 0.14 0.44 0.42
Partisan Gap -30.4 -11.16 18.39
Moderatly Pro-Republican (75th pct) Democrat 0.13 0.43 0.44
Partisan Gap -12.73 -4.7 7.72
Neutral (50th pct) Democrat 0.12 0.42 0.46
Partisan Gap 4.59 1.7 -2.79
Moderatly Pro-Democrat (25th pct) Democrat 0.12 0.41 0.48
Partisan Gap 12.62 4.66 -7.66
Strongly Pro-Democrat (10th pct) Democrat 0.11 0.4 0.5
Partisan Gap 28.33 10.41 -17.15
Table 8: Partisan Knowledge of the News 1/2
results. The magnitude of the congruence effects are smaller but economically significant.
For example, the Partisan Gap is equal to 7.83% for a news story that reflects favorably
on the Republican Party, it is small (-3.03%) for a neutral news story, and it is equal
to -18.93% for a news story that reflects unfavorably on the Republican Party. For
completeness, Table 20 in the Appendix reports our main results regarding the average
voter’s knowledge of the news. Overall, our main findings appear unaffected when
restricting our attention to the data from the last four surveys and using the direct
measure of bj .
37
0 − 0.25 0.25 − 0.75 0.75 − 1
Partisan Gap -14.06 -5.77 7.83
Partisan Gap -10.21 -4.19 5.68
Partisan Gap 5.44 2.23 -3.03
Partisan Gap 19.92 8.17 -11.08
Partisan Gap 34.15 13.92 -18.93
Table 9: Partisan Knowledge of the News 2/2
4.4 Effect of Time
In our framework, the probability that a voter knows attributes a probability of

truth equal to h or higher also depends on the number of months that have elapsed
since the story came out. This is captured by the population parameter δ in (4). Figure
4 plots the posterior distribution of the decay parameter δ. It is tightly estimated away
from 1, suggesting an effect of time passing on voters’ knowledge of the news.
In Table 10, we report the average probability that a voter attaches various levels
of confidence h to the average story being true as a function of the number of months
that have elapsed. For example, the probability that a voter attributes a chance equal
to or higher than 75% that a typical story is true is 44% when the story is less than 4
weeks old, but the corresponding figure falls to 40% when the story is between 4 and 8
38
Figure 4: Posterior Distribution of Parameter δ
weeks old, and to 37% when the story is between 8 and 12 weeks old. In other words,
time seems to have a rather sizable effect on the odds of knowing a story. Although
determining the exact underlying mechanism is beyond the purview of this paper, the
effect of limited memory and motivated beliefs in combination with decreasing media
coverage are likely significant drivers of our findings [e.g., Zimmermann, 2020].
Time Passed (Months)

Confidence 0 1 2
0 - 0.5 0.13 0.15 0.16
0.25 - 0.75 0.43 0.45 0.47
0.75 - 1 0.44 0.4 0.37
Table 10: Effect of Time Passing
4.5 Inequalities
As mentioned above, there exists an important literature documenting the relationship

between media coverage and voters’ information and, in turn, the relationship between
39
voters’ information and the attention received from politicians. One important channel
through which this accountability channel operates is through voting. If voters are
aware of the policies and actions implemented by politicians, the latter have greater
incentives to cater to voters’ preferences to increase their odds of reelection. Our
analysis so far has mostly documented the level of knowledge about political news
exhibited by the average voter. Investigating the distribution of knowledge across
socioeconomic groups is also of interest. As politicians are likely aware of the link
between information and voting, they have incentives to skew their policies towards the
better informed groups of voters.
To illustrate some of these dynamics, in Online Appendix A we develop a simple
model of retrospective voting inspired by Strömberg [2001], Prat and Strömberg [2013],
and Matějka and Tabellini [2017]. In the model, various groups of voters differ in their
policy preferences ug (·), their size sg , and information levels ρ̄g (the share of informed
individuals in group g). We show that an incumbent politician seeking reelection has
ρ̄g
incentives to allocate weights equal to s
ρ̄ g
on the various groups of voters, where
ρ̄ denotes the average voter’s level of information. By contrast, a utilitarian social
planner would allocate weights equal to sg . In other words, the incumbent politician
places greater weight on the better informed groups of voters.
In this section, we quantify the extent of knowledge inequalities across socioeconomic
groups. Table 11 reports for the 16 socioeconomic groups our model explicitly identifies
– the intersections of Age, Gender, Race, and Income (see Section 3), the probability
that an average member of a particular group assigns a probability equal to or greater
than 0.75 to the typical news story of the month being true.43 Our results suggest
43
Again, by typical news story we mean a news story whose associated parameter a is the average
of the means of the posterior distributions of all our parameters aj . We also suppose this typical news
story to be neutral (i.e., we set b = 0).
40
significant differences across groups of voters. To take an extreme example, the average
nonwhite, female voter age 47 or less with a below-median income has a 36% probability
of knowing the typical news story about the Federal Government. By contrast, the
average white, male voter age 48 or more with an above-median income has a 52%
probability of knowing the same story.
Age > Female White Income ρ < 0.25 ρ ∈ (0.25, 0.75) ρ > 0.75
47 60k+
1 0.16 0.46 0.38
2 x 0.15 0.45 0.40
3 x 0.13 0.43 0.44
4 x x 0.12 0.42 0.46
5 x 0.17 0.47 0.36
6 x x 0.16 0.46 0.38
7 x x 0.15 0.45 0.40
8 x x x 0.14 0.44 0.42
9 x 0.14 0.44 0.42
10 x x 0.11 0.41 0.48
11 x x 0.11 0.41 0.47
12 x x x 0.10 0.38 0.52
13 x x 0.14 0.45 0.41
14 x x x 0.12 0.41 0.47
15 x x x 0.12 0.42 0.46
16 x x x x 0.12 0.41 0.47
Table 11: Knowledge of Political News across Socioeconomic Groups
Next, we explore the explanatory role played by socioeconomic factors in a regression

format.44 Column (1) in Table 21 (see Appendix A) looks at the effects of various
socioeconomic factors on the probability that a voter knows the typical news story about
44
Recall that our model allows for different group-level means µg across 16 groups defined by Age,
Gender, Race, and Income. This approach tends to give these four socioeconomic characteristics greater
weight in explaining θi (and thus ρi (h)) compared to other characteristics (e.g., interest in politics,
or media usage). For this reason, the coefficients associated to Age, Gender, Race, and Income are
relatively large in our regression analysis. By and large, not prioritizing these four variables would still
lead to larger coefficients on socioeconomic variables, but the differences would diminish noticeably.
41
the Federal Government. All our coefficients are estimated very precisely. Age is the
most important characteristic, with voters age 47 or more being 5.6 percent points more
likely to know the typical story. Intuitively, college education and income also positively
predict knowledge, by 0.8 and 2.7 percentage points respectively. By contrast, women
and racial minorities are associated with lower knowledge of the news. Women are 3.1
percentage points less likely to know the typical story about the Federal Government.
Hispanics and African-Americans are 3.7 and 4.1 percentage points less likely to know
the typical news story, respectively. Column (2) adds political affiliations (where the
excluded category are ‘Independents’) and Column 3 adds general engagement with
party politics (partisanship). The coefficients on political parties are small and they
switch sign depending on whether partisanship is included. Partisanship increases the
odds of knowing the typical news story by 0.4 percentage points. Table 22 (where
Column (1) reproduces Column (3) in Table 21) includes media consumption habits.
In both Columns (2) and (3) the number of news outlets and time usage (in minutes)
are significantly positively associated with knowledge of the news, and the coefficients
on the socioeconomic factors are largely unchanged by the inclusion of these news
consumption habits (as well as extra media controls in Column (3)). Finally, Table 23
(which reproduces Table 21’s Column (3) and Table 22’s Column (3)) adds Political
Interest as a control variable. Political Interest has been highlighted by previous work
as an important factor in determining knowledge. Our results are consistent: we find
that general interest in politics increases the odds of knowing the typical story about
the Federal Government by 1.5 percentage points.
Lasso Regression. We also employ standard Lasso regression methods to shed light
on the most important determinants of news knowledge. Table 24 reports our results,
42
where the dependent variable is the probability that a voter knows the typical news
story. Independent variables include all socioeconomic factors, political characteristics,
all media consumption variables, and all two-way interactions. The penalty λ is varied
to include one variable at a time. As reported in Table 24, Age is the most important
predictor of knowledge, followed by Black, Female, Family Income, and Political Interest.
We return to our simple theoretical framework to illustrate the relevance of our
findings from a political economy angle. In Figure 5, the grey bars correspond to the
size of various age groups in our sample. By contrast, the blue bars represent the actual
weights an incumbent seeking reelection would allocate these various groups, say when
designing a policy that affects voters of different ages differently. For example, according
to our estimates, the incumbent will behave as if voters age 69 or more represent 18%
of all voters, even though they represent only 15% of voters. This occurs because, as
discussed, age is positively associated with knowledge of the news.
Figure 5: Inequalities in Knowledge of the News
43
4.5.1 Media Consumption
Survey respondents were also asked about which media outlets they rely on to get
their national news. In this section we briefly explore the extent to which reliance
on various media outlets predict knowledge of the news. Although there is ultimately
little we can say in terms of causality, understanding how knowledge correlates with
media outlets is nonetheless interesting. Table 25 explores in a regression format the
relationship between various media outlets and the probability that a voter knows
the typical news story about the Federal Government. In addition to including the
baseline socioeconomic factors and media consumption habits (number of sources and
time usage) analyzed above, we also include a series of dummies capturing voters’
various degrees of reliance on the 10 most important media outlets in terms of attention
share [Prat, 2018]. These are Fox, CNN, ABC, Facebook, NBC, CBS, Google, New
York Times, MSNBC, and YouTube. For example, “3 or more sources: fox” takes value
1 if an individual relies on 3 or more news sources, including Fox. Similarly, “2 or
less sources: cbs” takes value 1 if an individual relies on strictly less than three news
sources, and one of these sources is CBS. When consumed in addition to 2 or more news
outlets, 6 news outlets are not statistically associated with news knowledge. Reading
the New York Times (in addition to 2 or more other media outlets) is significantly
positively associated with news knowledge. By contrast, CBS, Google, and YouTube are
significantly negatively associated with news knowledge. When consumed exclusively,
or in addition to a single other news outlet, CNN, ABC, and Facebook are significantly
negatively associated with news knowledge. By contrast, the BBC and Google positively
predict knowledge of the news. Table 26 conveys similar information, by reporting the
average probability of knowing a story for various news diets.
44
5 Extensions and Robustness Checks
We replicate our analysis in two ways. First, we switch topic and measure voters’
knowledge of the Democratic Party primaries. Second, we return to stories about the
Federal Government but rely on a different sample of respondents to measure knowledge.
5.1 Democratic Party Presidential Primaries
In this extension, we apply our news generating process to select knowledge items
pertaining to the Democratic Party presidential primaries. Our objective is twofold.
We illustrate the robustness of our method, which can be used to measure voters’
knowledge of distinct types of topics. Also, we shed light on voters’ knowledge of a
key US electoral institution. Exactly as before, we estimate the model highlighted in
Section 3 to obtain the posterior distributions of the various parameters of interest. The
model is estimated using the quizzes about the Democratic Party presidential primaries
exclusively (included in the last 3 surveys). Even though an individual completed
quizzes about both the primaries and the Federal Government, we rely only on his/her
performance when completing the quizzes about the primaries to estimate individual
parameters.45,46 In other words, we allow an individual’s knowledge precision θi to vary
by topic.
Tables (12) and (13) replicate Tables (6) and (7) for our measurement of voters’
knowledge of the Democratic primaries. Again, there exists significant heterogeneity
across stories. Whereas only 23% of voters knew the story: Democrats in Presidential
45
This was true also in our analysis of voters’ knowledge of political news covering the Federal
Government: we relied exclusively on individuals’ performance when completing the quizzes about the
Federal Government.
46
Because we included the quizzes about the primaries in the last three of our surveys, we are able
to separately measure bj . We therefore present results that rely on our direct measure of bj only.
45
debate hint at no swift end to China tariffs, 60% of them knew the story: The Democratic
presidential nominating race got off to a chaotic start in Iowa, as the results of the
state’s caucuses were delayed for hours. For both of these stories (and others), our
model predicts significantly different probabilities of knowing the story versus selecting
it when completing the survey.
Democrats in Presidential debate hint at no swift end to China tariffs. Sep 19 0.45 -0.09 0.44 0.27 0.5 0.23
Democratic groups launched a multi-million digital ad effort to fight President Trump. Sep 19 0.6 0.34 0.6 0.19 0.49 0.33
In a recent debate, all of the Democratic presidential candidates agreed universal Sep 19 0.79 0.79 0.79 0.13 0.43 0.44
healthcare is a top priority.
Former New York Mayor Michael Bloomberg has been considering whether to run for Sep 19 0.83 1.02 0.84 0.11 0.39 0.5
president.
Elizabeth Warren catches up with Joe Biden in a national opinion poll. Sep 19 0.84 1.02 0.86 0.11 0.39 0.5
Two billionaire Democratic presidential hopefuls, Michael Bloomberg and Tom Steyer, Oct/Nov 19 0.84 1.06 0.85 0.1 0.38 0.52
collectively spent more in 2019 than the rest of the Democratic candidates combined
Bernie Sanders won New Hampshire’s Democratic presidential primary Oct/Nov 19 0.84 1.12 0.87 0.1 0.37 0.53
Presidential candidate Elizabeth Warren proposed a Medicare for All plan that she Sep 19 0.89 1.35 0.91 0.08 0.33 0.59
said would not require raising middle-class taxes.
The Democratic presidential nominating race got off to a chaotic start in Iowa, as the Oct/Nov 19 0.89 1.41 0.92 0.08 0.32 0.6
results of the state’s caucuses were delayed for hours
Table 12: True Statements
Pete Buttigieg chose Kamala Harris as his Vice-Presidential pick Oct/Nov 19 0.1 1.63 0.07 0.65 0.29 0.06
Bernie Sanders admitted to taking Wall Street campaign contributions Oct/Nov 19 0.11 1.36 0.09 0.59 0.33 0.08
Hillary Clinton endorsed presidential candidate Tulsi Gabbard despite previous spat. Sep 19 0.13 1.78 0.07 0.68 0.26 0.06
Black face photo shows up in Joe Biden’s past. Sep 19 0.15 1.26 0.14 0.56 0.35 0.09
Voting Intentions Poll showed Bloomberg above Biden with white, working class Sep 19 0.18 0.75 0.21 0.43 0.44 0.13
voters.
Andrew Yang Endorsed Amy Klobuchar, saying she is Most Honest in the Race Oct/Nov 19 0.21 0.6 0.2 0.39 0.46 0.15
Elizabeth Warren plan would slash 70% of mining jobs. Sep 19 0.37 0.28 0.37 0.31 0.49 0.2
Pete Buttigieg received a significant donation, pushing him to the front of the Sep 19 0.37 0.18 0.37 0.29 0.5 0.22
fundraising race among all Democratic candidates as of early November.
Kamala Harris attacks Cory Booker over Newark’s water problem. Sep 19 0.41 0.18 0.41 0.29 0.5 0.22
Table 13: False Statements
Table 14 reports the probability that a voter knows (for various intervals of confidence
h) the typical first, second, and third story of the month about the Democratic Party
presidential primaries.47 As before, the ranking is provided by our panel of journalists.
For example, the average voter is 53% likely to assign a probability to the first story
47
By typical we mean a story whose associated parameter a corresponds to the average of the means
of the posterior distributions of all our parameters aj (for a given rank: either first, second, or third).
46
of the month being true equal to or greater than 0.75. The corresponding figure falls
to 50% and 33% when we move to the second and third most important stories of the
month. Overall, therefore, it seems that the average voter is more likely to know the
typical story about the Federal Government than the typical story about the Democratic
primaries. This difference seems to be driven largely by the first story of the month
about the Federal Government (see Table 4).

0 - 0.25 0.10 0.11 0.19
0.25 - 0.75 0.37 0.39 0.49
0.75 - 1 0.53 0.50 0.33
Table 14: News about the Democratic primaries
Next, Table 15 documents the effect of partisanship on the odds of knowing stories
about the Democratic presidential primaries. Again, we find evidence of partisanship on
voters’ knowledge of the news, with voters being more likely to know stories that reflect
favorably on their preferred party. Interestingly, though, the effect of partisanship on
voters’ knowledge of the news about the primaries seems to be lower than that at play
regarding the news about the Federal Government.
Last, Table 16 reports the effect of time passing on the odds that voters know
the typical news story about the Democratic primaries. As for news on the Federal
Government, we find a sizable effect of time, with each month reducing the likelihood
that the average voter knows the typical story by about 5 percentage points.
47
0 − 0.25 0.25 − 0.75 0.75 − 1
Partisan Gap -8.04 -3.56 4.24
Partisan Gap -2.49 -1.1 1.32
Partisan Gap 2.11 0.94 -1.12
Partisan Gap 3.65 1.62 -1.93
Partisan Gap 4.94 2.19 -2.61
Table 15: Partisan Knowledge of the News – Democratic Primaries
Time Passed (Months)

Confidence 0 1 2
0 - 0.5 0.11 0.12 0.14
0.25 - 0.75 0.39 0.42 0.45
0.75 - 1 0.5 0.45 0.41
Table 16: Effect of Time Passing – Democratic Primaries
5.2 MTurk Sample
Although the YouGov sample of US adult voters is of high quality, one may wonder
whether some unobservable traits correlated with YouGov membership may drive our
results. To address this concern, we replicated the eighth survey on a sample of 800
US voters recruited through Amazon Mechanical Turk (MTurk). Clearly, recruiting
participants through MTurk may present its own distinct problems. Nevertheless,
48
investigating whether our main results line up is interesting. Table 17 provides summary
statistics for our sample of MTurk participants, and compares them with those of our
YouGov sample. The MTurk sample is significantly younger, better educated, and
poorer. It also contains fewer nonwhite individuals.
We estimate the various posterior distributions for our parameters of interest using
the eighth survey exclusively.48 Table 18 reports the likelihood that an individual drawn
from the MTurk sample knows (for different confidence levels h) the first, second, and
third news story included in our eighth survey. For example, the average individual
knows the first story of the month about the Federal Government with probability
73%. For completeness, we report the corresponding figures for the YouGov sample,
where – for the sake of comparability – we estimated the model using the eighth survey
exclusively. The numbers appear reassuringly similar, and the differences – where
they exist – are plausibly explained by the underlying differences in socioeconomic
characteristics across both samples.
Statistic YouGov Amazon MTurk

Median Age 49.00 37.00
% Female 0.52 0.51
% White 0.69 0.74
% Black 0.11 0.07
% 4yr College Degree 0.30 0.57
% Unemployed 0.07 0.04
% Married 0.48 0.39
% Family Inc <30k 0.28 0.44
% Family Inc 30k - 60k 0.20 0.33
Table 17: Socioeconomic Characteristics
48
This precludes us from estimating the posterior distribution of δ, which we set equal to 1.
49
0 - 0.25 0.05 0.10 0.09
0.25 - 0.75 0.22 0.38 0.35
0.75 - 1 0.73 0.52 0.56
Table 18: MTurk Sample – News Stories about the Federal Government (8th Survey)

0 - 0.25 0.05 0.09 0.10
0.25 - 0.75 0.22 0.34 0.38
0.75 - 1 0.74 0.57 0.52
Table 19: YouGov Sample – News Stories about the Federal Government (8th Survey)
6 Concluding Remarks
This paper develops a new methodology to measure voters’ knowledge of political

news that combines a protocol for identifying stories, an incentivized quiz to elicit news
knowledge, and the estimation of a model of individual knowledge that includes story
difficulty, partisanship, and time passing. We apply this method – and repeat it 8
times – to the 3 most important news of the month about the US Federal Government
according to mainstream media. We find significant heterogeneity across voters in their
knowledge of the news. We also document large differences across groups of voters as
well as sizable effects of partisanship and of time passing.
Our analysis could be extended in several interesting ways. In particular, replicating
our analysis across a number of countries would allow for an international comparison
of news knowledge. Replicating our analysis in the US but at the local level would also
be relevant, as many commentators worry about the existence of local news deserts.
50
References
Hunt Alcott and Matthew Gentzkow. Social Media and Fake News in the 2016 Election.
Journal of Economic Perspectives, 31(2):211–236, 2017.
Hunt Allcott, Luca Braghieri, Sarah Eichmeyer, and Matthew Gentzkow. The Welfare
Effects of Social Media. Working Paper, 2019.
Eric Arias, Horacio Larreguy, John Marshall, and Pablo Querubin. Does the Content
and Mode of Delivery of Information Matter for Political Accountability? Evidence
from a Field Experiment in Mexico. Working Paper, 2018.
Eric Arias, Horacio Larreguy, John Marshall, and Pablo Querubin. Priors Rule: When
Do Malfeasance Revelations Help or Hurt Incumbent Parties. Working Paper, 2019.
Abhijit Banerjee, Selvan Kumar, Rohini Pande, and Felix Su. Do Informed Voters
Make Better Choices? Experimental Evidence from Urban India. Working Paper,
2012.
Oscar Barrera Rodriguez, Sergei M. Guriev, Emeric Henry, and Ekaterina Zhuravskaya.
Facts, Alternative Facts, and Fact Checking in Times of Post-Truth Politics. CEPR
Working Paper, 2018.
Roland Bénabou and Jean Tirole. Self-Confidence and Personal Motivation. Quarterly
Journal of Economics, 117, 2002.
Darrell R. Bock. Estimating Item Parameters and Latent Ability When Responses Are
Scored in Two or More Nominal Categories. Psychometrika, 37:29–51, 1972.
John G. Bullock, Alan S. Gerber, Seth J. Hill, and Gregory A. Huber. Partisan Bias
in Factual Beliefs about Politics. Quarterly Journal of Political Science, 10, 2015.
Bob Carpenter, Andrew Gelman, Matthew D Hoffman, Daniel Lee, Ben Goodrich,
Michael Betancourt, Marcus Brubaker, Jiqiang Guo, Peter Li, and Allen Riddell.
Stan: A probabilistic programming language. Journal of statistical software, 76(1),
2017.
Yuyu Chen and David Y. Yang. The Impact of Media Censorship: 1984 or Brave New
World? American Economic Review, 109(6):2294–2332, 2019.
Stefano DellaVigna and Ethan Kaplan. The Fox News Effect: Media Bias and Voting.
The Quarterly Journal of Economics, 122(3):1187–1234, 2007.
Michael X. Delli Carpini and Scott Keeter. What Americans Know about Politics and
Why It Matters. Yale University Press, 1996.
51
Thomas Eisensee and David Strömberg. News Droughts, News Floods, and U. S.
Disaster Relief. The Quarterly Journal of Economics, 122(2):693–728, 2007.
Ruben Enikolopov, Maria Petrova, and Ekaterina Zhuravskaya. Media and Political
Persuasion: Evidence from Russia. American Economic Review, 101(7):3253–85,
2011.
Eurobarometer. Standard Eurobarometer 87. European Commision, 2017.
Claudio Ferraz and Frederico Finan. Exposing Corrupt Politicians: The Effects of
Brazil’s Publicly Released Audits on Electoral Outcomes. The Quarterly Journal of
Economics, 123(2):703–745, 2008.
Alessandro Gavazza, Mattia Nardotto, and Tommaso Valletti. Internet and Politics:
Evidence from U.K. Local Elections and Local Government Policies. The Review of
Economic Studies, 86(5):2092–2135, 2018.
Matthew Gentzkow, Jesse M. Shapiro, and Daniel F. Stone. Media bias in the
marketplace: theory. In Simon Anderson, Joel Waldfogel, and David Stromberg,
editors, Handbook of Media Economics, chapter 2. North Holland, 2015.
Alan S. Gerber, Dean Karlan, and Daniel 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.
Andrew M. Guess. Measure for Measure: An Experimental Test of Online Political

Media Exposure. Political Analysis, 23(1):59–75, 2015.
Matthew D Hoffman and Andrew Gelman. The no-u-turn sampler: adaptively setting
path lengths in hamiltonian monte carlo. Journal of Machine Learning Research, 15
(1):1593–1623, 2014.
Chad Kendall, Tommaso Nannicini, and Francesco Trebbi. How Do Voters Respond to
Information? Evidence from a Randomized Campaign. American Economic Review,
105(1):322–53, January 2015.
Julien Labonne, Cesi Cruz, and Philip Keefer. Buying Informed Voters: New Effects
of Information on Voters and Candidates. Working Paper, 2019.
David M. J. Lazer, Matthew A. Baum, Yochai Benkler, Adam J. Berinsky, Kelly M.

Greenhill, Filippo Menczer, Miriam J. Metzger, Brendan Nyhan, Gordon Pennycook,
David Rothschild, Michael Schudson, Steven A. Sloman, Cass R. Sunstein, Emily A.
Thorson, Duncan J. Watts, and Jonathan L. Zittrain. The science of fake news.
Science, 359(6380):1094–1096, 2018. ISSN 0036-8075. doi: 10.1126/science.aao2998.
52
Caroline Le Pennec and Vincent Pons. Vote Choice Formation and Minimal Effects
of TV Debates: Evidence from 61 Elections in 9 OECD Countries. NBER Working
Paper 26572, 2019.
Daniel J. Levitin. A Field Guide to Lies: Critical Thinking in the Information Age.
New York, New York: Dutton, 2016.
Gregory J. Martin, Maria Petrova, and Ali Yurukoglu. Bias in Cable News: Persuasion
and Polarization. American Economic Review, 107(9):2565–2599, 2017.
Filip Matějka and Guido Tabellini. Electoral Competition with Rationally Inattentive
Voters. Working Paper, 2017.
Pew. From Brexit to Zika: What Do Americans Know? Pew Research Center, 2017.
Pew. Wide Gender Gap, Growing Educational Divide in Voters’ Party Identification.
Pew Research Center, 2018.
Andrea Prat. Media Power. Journal of Political Economy, 126(4):1747–83, 2018.
Andrea Prat and David Strömberg. The Political Economy of Mass Media, volume 2 of
Econometric Society Monographs, page 135–187. Cambridge University Press, 2013.
Vincent Price and John Zaller. Who Gets the News? Alternative Measures of News
Reception and Their Implications for Research. The Public Opinion Quarterly, 57
(2):133–164, 1993.
Markus Prior. Visual Political Knowledge: A Different Road to Competence? The

Journal of Politics, 76(1):41–57, 2014.
Markus Prior and Arthur Lupia. Money, Time, and Political Knowledge: Distinguishing
Quick Recall and Political Learning Skills. American Journal of Political Science, 52
(1):169–183, 2008.
Markus Prior, Gaurav Sood, and Kabir Khanna. You Cannot be Serious: The Impact of
Accuracy Incentives on Partisan Bias in Reports of Economic Perceptions. Quarterly
Journal of Political Science, 10, 2015.
Georg Rasch. Probabilistic Models for Some Intelligence and Attainment Tests. Danish
Institute for Educational Research, 1960.
Ritva Reinikka and Jakob Svensson. Fighting Corruption to Improve Schooling:

Evidence from a Newspaper Campaign in Uganda. Journal of the European Economic
Association, 3(2-3):259–267, 2005. doi: 10.1162/jeea.2005.3.2-3.259.
53
James M. Snyder and David Strömberg. Press Coverage and Political Accountability.
Journal of Political Economy, 118(2):355–408, 2010.
Richard Stengel. Information Wars. How We Lost the Global Battle Against
Disinformation and What We Can Do about It. Grove Press, 2019.
David Strömberg. Mass Media and Public Policy. European Economic Review, 45,
2001.
David Strömberg. Radio’s Impact on Public Spending. The Quarterly Journal of

Economics, 119(1):189–221, 2004.
David Strömberg. Media and Politics. Annual Review of Economics, 7(1):173–205,

2015.
Wim J. Van der Linden and Ronald K. (eds) Hambleton. Handbook of Modern Item
Response Theory. H. New York, NY: Springer, 1997.
Florian Zimmermann. The Dynamics of Motivated Beliefs. American Economic Review,

110(2):337–61, 2020.
54
A Appendix Tables

0 - 0.25 0.06 0.11 0.16
0.25 - 0.75 0.27 0.39 0.46
0.75 - 1 0.66 0.50 0.38
Table 20: Knowledge of the News - Second Approach
55
Dependent variable:
ρij(0.75) (Second Story)
(1) (2) (3)
∗∗∗
Democrat 0.004 −0.001
(0.001) (0.001)
Republican 0.0002 −0.004∗∗∗
(0.001) (0.001)
Partisan 0.006∗∗∗
(0.001)
Age > 47 0.056∗∗∗ 0.056∗∗∗ 0.056∗∗∗
(0.001) (0.001) (0.001)
Income > 60k 0.027∗∗∗ 0.027∗∗∗ 0.028∗∗∗
(0.001) (0.001) (0.001)
College + 0.008∗∗∗ 0.007∗∗∗ 0.007∗∗∗
(0.001) (0.001) (0.001)
Female −0.031∗∗∗ −0.031∗∗∗ −0.031∗∗∗
(0.001) (0.001) (0.001)
Black −0.041∗∗∗ −0.042∗∗∗ −0.042∗∗∗
(0.002) (0.002) (0.002)
Hispanic −0.037∗∗∗ −0.038∗∗∗ −0.037∗∗∗
(0.001) (0.001) (0.001)
Constant 0.422∗∗∗ 0.421∗∗∗ 0.422∗∗∗
(0.001) (0.001) (0.001)
Observations 7,614 7,614 7,379
R2 0.530 0.531 0.531
Notes:∗ p<0.1; ∗∗
p<0.05; ∗∗∗
p<0.01.
Table 21: Socioeconomic Factors 1/3
56
Dependent variable:
ρ(0.75)
(1) (2) (3)
Democrat −0.001 −0.001 0.0003
(0.001) (0.001) (0.001)
Republican −0.004∗∗∗ −0.003∗∗ −0.002
(0.001) (0.001) (0.001)
Partisan 0.006∗∗∗ 0.004∗∗∗ 0.003∗∗∗
(0.001) (0.001) (0.001)
Age > 47 0.056∗∗∗ 0.053∗∗∗ 0.053∗∗∗
(0.001) (0.001) (0.001)
Income > 60k 0.028∗∗∗ 0.027∗∗∗ 0.027∗∗∗
(0.001) (0.001) (0.001)
College + 0.007∗∗∗ 0.006∗∗∗ 0.005∗∗∗
(0.001) (0.001) (0.001)
Female −0.031∗∗∗ −0.030∗∗∗ −0.029∗∗∗
(0.001) (0.001) (0.001)
Black −0.042∗∗∗ −0.042∗∗∗ −0.040∗∗∗
(0.002) (0.002) (0.002)
Hispanic −0.037∗∗∗ −0.037∗∗∗ −0.036∗∗∗
(0.001) (0.001) (0.001)
Sources 3+ 0.005∗∗∗ 0.010∗∗∗
(0.001) (0.001)
Total time 0.00001∗∗∗ 0.00001∗∗∗
(0.00000) (0.00000)
Constant 0.422∗∗∗ 0.418∗∗∗ 0.418∗∗∗
(0.001) (0.001) (0.001)
Extra media controls X
R2 0.531 0.538 0.545
Notes:∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Extra media controls
include: voter registration, Indicators for using tv, print,
online and radio as a news source, as well as dummies for 10
biggest news sources interacted with using at least 3 sources.
57
Dependent variable:
ρ(0.75)
(1) (2) (3)
Democrat −0.001 0.0003 0.001
(0.001) (0.001) (0.001)
Republican −0.004∗∗∗ −0.002 −0.001
(0.001) (0.001) (0.001)
Partisan 0.006∗∗∗ 0.003∗∗∗ 0.001
(0.001) (0.001) (0.001)
Poli Interest 0.014∗∗∗
(0.001)
Age ¿ 47 0.056∗∗∗ 0.053∗∗∗ 0.051∗∗∗
(0.001) (0.001) (0.001)
Income > 60k 0.028∗∗∗ 0.027∗∗∗ 0.026∗∗∗
(0.001) (0.001) (0.001)
College + 0.007∗∗∗ 0.005∗∗∗ 0.005∗∗∗
(0.001) (0.001) (0.001)
Female −0.031∗∗∗ −0.029∗∗∗ −0.028∗∗∗
(0.001) (0.001) (0.001)
Black −0.042∗∗∗ −0.040∗∗∗ −0.038∗∗∗
(0.002) (0.002) (0.002)
Hispanic −0.037∗∗∗ −0.036∗∗∗ −0.035∗∗∗
(0.001) (0.001) (0.001)
Sources 3+ 0.010∗∗∗ 0.007∗∗∗
(0.001) (0.001)
Total time 0.00001∗∗∗ 0.00000∗∗∗
(0.00000) (0.00000)
Constant 0.422∗∗∗ 0.418∗∗∗ 0.414∗∗∗
(0.001) (0.001) (0.001)
Extra media controls X X
R2 0.531 0.545 0.556
Notes:∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Extra media controls
include: voter registration, Indicators for using tv, print,
online and radio as a news source, as well as dummies for 10
biggest news sources interacted with using at least 3 sources.
58
Variable (1) (2) (3) (4)
Age 0.0004 0.0007 0.0008 0.0009
Black -0.0011 -0.0082 -0.0160
Female -0.0097 -0.0133 -0.0170
Family Income < 30k -0.0004 -0.0023
Poli Interest 0.0069
Table 24: LASSO Exercise
59
Dependent variable:
ρ(0.75)
Age > 47 0.054∗∗∗ (0.001)
Income > 60k 0.026∗∗∗ (0.001)
College + 0.005∗∗∗ (0.001)
Female −0.029∗∗∗ (0.001)
Black −0.039∗∗∗ (0.002)
Hispanic −0.036∗∗∗ (0.001)
Sources 3+ 0.010∗∗∗ (0.001)
Total time 0.00001∗∗∗ (0.00000)
3 or more sources: fox −0.001 (0.001)
3 or more sources: cnn −0.001 (0.001)
3 or more sources: abc −0.002 (0.001)
3 or more sources: facebook −0.001 (0.001)
3 or more sources: nbc −0.001 (0.001)
3 or more sources: cbs −0.006∗∗∗ (0.001)
3 or more sources: google −0.003∗∗ (0.001)
3 or more sources: nytimes 0.003∗∗ (0.001)
3 or more sources: msnbc 0.001 (0.001)
3 or more sources: youtube −0.004∗∗ (0.001)
3 or more sources: bbc 0.002 (0.002)
2 or less sources: fox 0.002 (0.002)
2 or less sources: cnn −0.006∗∗ (0.003)
2 or less sources: abc −0.008∗∗∗ (0.003)
2 or less sources: facebook −0.007∗∗∗ (0.002)
2 or less sources: nbc 0.005 (0.003)
2 or less sources: cbs 0.005 (0.003)
2 or less sources: google 0.011∗∗∗ (0.004)
2 or less sources: nytimes 0.009 (0.006)
2 or less sources: msnbc 0.007 (0.005)
2 or less sources: youtube 0.006∗ (0.004)
2 or less sources: bbc 0.026∗∗∗ (0.008)
Constant 0.418∗∗∗ (0.001)
Observations 7,614
R2 0.546
Notes:∗ p<0.1; ∗∗
p<0.05; ∗∗∗
p<0.01.
Table 25
60
Outlet AS p̄ij (0.75)
Sources 3+ Sources 1-2
Fox 0.1 0.46 0.46
CNN 0.07 0.45 0.42
ABC 0.06 0.45 0.42
FB 0.06 0.45 0.42
NBC 0.05 0.45 0.44
CBS 0.04 0.45 0.44
Google 0.03 0.44 0.43
NYT 0.03 0.45 0.42
MSNBC 0.03 0.46 0.45
Youtube 0.03 0.44 0.42
BBC 0.01 0.46 0.44
Table 26: Knowledge of the News and Media Outlets
61

angelucci

Uploaded by

Copyright:

Available Formats

angelucci

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

angelucci

Uploaded by

Copyright:

Available Formats

Measuring Voters’ Knowledge of Political News∗

Charles Angelucci† Andrea Prat‡

2.1 News Generating Process

Universe of Relevant Knowledge Items. We rely on Reuters’ publicly-available

Generating 3 True and 3 False Knowledge Items. We rely on a panel of 3

2.2 Survey Design

Table 1: Socioeconomic Characteristics

and bonuses equal to $1 for each quiz correctly answered.22

Table 2: Party Affiliations

2.2.1 Media Consumption Habits

Table 3: Media Consumption Habits

2.2.2 Knowledge of the News

represented by 0, “very unfavorable” is represented by -1 and “very favorable” is represented by 1.

3.1 The News Knowledge Problem

Suppose agent i is trying to establish the truth of statement j, which we call τj ∈

where H = ln h/(1 − h) . The left-hand side of the inequality is a function of random

3.2 A Discrete Choice Model

The comparative statics of the expressions above are intuitive.31 If j is a true

3.3 Econometric Model

3.4 Literature Discussion

4.1 Knowledge of the News

Table 4: Knowledge of the News

Table 5: Average Probability that an Individual in a Given Tier Knows a Statement

14% selected 1 true statement. By way of comparison, an uninformed individual, with

Figure 2: The posterior distribution of θ

Next we explore various dimensions of heterogeneity across news stories. Table

4.3 Effect of Partisanship

Table 6: True Statements

Table 7: False Statements

Table 8: Partisan Knowledge of the News 1/2

Table 9: Partisan Knowledge of the News 2/2

4.4 Effect of Time

In our framework, the probability that a voter knows attributes a probability of

Time Passed (Months)

Table 10: Effect of Time Passing

As mentioned above, there exists an important literature documenting the relationship

Table 11: Knowledge of Political News across Socioeconomic Groups

Next, we explore the explanatory role played by socioeconomic factors in a regression

Figure 5: Inequalities in Knowledge of the News

5.1 Democratic Party Presidential Primaries

Table 12: True Statements

Table 13: False Statements

Confidence First story Second story Third story

Table 14: News about the Democratic primaries

Table 15: Partisan Knowledge of the News – Democratic Primaries

Time Passed (Months)

Table 16: Effect of Time Passing – Democratic Primaries

5.2 MTurk Sample

Statistic YouGov Amazon MTurk

Table 17: Socioeconomic Characteristics

Confidence First story Second story Third story

This paper develops a new methodology to measure voters’ knowledge of political

Eurobarometer. Standard Eurobarometer 87. European Commision, 2017.

Andrew M. Guess. Measure for Measure: An Experimental Test of Online Political

David M. J. Lazer, Matthew A. Baum, Yochai Benkler, Adam J. Berinsky, Kelly M.

Andrea Prat. Media Power. Journal of Political Economy, 126(4):1747–83, 2018.

Markus Prior. Visual Political Knowledge: A Different Road to Competence? The

Ritva Reinikka and Jakob Svensson. Fighting Corruption to Improve Schooling:

David Strömberg. Radio’s Impact on Public Spending. The Quarterly Journal of

David Strömberg. Media and Politics. Annual Review of Economics, 7(1):173–205,