Trump Corrections

Taking Fact-checks Literally But Not Seriously?
The Effects of Journalistic Fact-checking on

Factual Beliefs and Candidate Favorability⇤
Forthcoming at Political Behavior
Brendan Nyhan† Ethan Porter‡

University of Michigan George Washington University
Jason Reifler§ Thomas J. Wood¶

University of Exeter The Ohio State University
January 11, 2019
Abstract
Are citizens willing to accept journalistic fact-checks of misleading claims from can-
didates they support and to update their attitudes about those candidates? Previous
studies have reached conflicting conclusions about the effects of exposure to counter-
attitudinal information. As fact-checking has become more prominent, it is therefore
worth examining how respondents respond to fact-checks of politicians — a question
with important implications for understanding the effects of this journalistic format
on elections. We present results to two experiments conducted during the 2016 cam-
paign that test the effects of exposure to realistic journalistic fact-checks of claims
made by Donald Trump during his convention speech and a general election debate.
These messages improved the accuracy of respondents’ factual beliefs, even among
his supporters, but had no measurable effect on attitudes toward Trump. These re-
sults suggest that journalistic fact-checks can reduce misperceptions but often have
minimal effects on candidate evaluations or vote choice.
⇤
We thank Kim Gross, John Pfaff, and D.J. Flynn for comments and Kyle Dropp for fielding study 1. This
research received funding support from the European Research Council (ERC) under the European Union’s
Horizon 2020 research and innovation program (grant agreement No 682758). We also received support
from the School of Media and Public Affairs at George Washington University. All errors are our own.
†
Professor, Ford School of Public Policy; bnyhan@umich.edu
‡
Assistant Professor, School of Media and Public Affairs; evporter@gwu.edu
§ Professor, Department of Politics; J.Reifler@exeter.ac.uk
¶
Assistant Professor, Department of Political Science; wood.1080@osu.edu
Journalistic fact-checking is an important new form of political news coverage (e.g., Spivak
2011; Graves 2016). However, little is known about its effects on citizens. Do they accept
fact-checks that conflict with their political affiliations or shrug off those that contradict
the claims of their preferred candidates? These questions have important implications
for debates over citizen competence and the quality of governance in democracies (e.g.,
Hochschild and Einstein 2015; Jamieson 2015).
Concerns about people’s willingness to accept unwelcome factual information like
counter-attitudinal fact-checks have become so widespread that Oxford Dictionaries named
post-truth the word of the year after the 2016 U.S. elections (BBC 2016). These concerns
are well-justified. Some research indicates, for instance, that people can be highly resis-
tant to journalistic fact-checks. Nyhan and Reifler (2010) find that corrective information
in mock news articles frequently fails to reduce salient misperceptions and can even in-
crease the prevalence of misperceptions among ideologically vulnerable groups compared
to those who read an article with no correction — a “backfire effect.” Other studies us-
ing relatively balanced formats have also found stiff resistance to uncongenial journalistic
fact-checks (e.g., Nyhan, Reifler, and Ubel 2013; Garrett, Nisbet, and Lynch 2013; Jarman
2016). Citizens may be especially resistant to unwelcome fact-checks during campaigns,
which frequently stimulate motivated reasoning (e.g., Lenz 2012).
By contrast, other studies find that fact-checking and other types of factual information
can partly overcome directionally motivated reasoning and reduce, or “debunk,” misper-
ceptions (e.g., Weeks 2015; Nyhan and Reifler N.d.; Wood and Porter 2018; Chan et al.
2017). Notably, Wood and Porter (2018) examine 52 issue areas and observe no evidence
of backfire effects. They do, however, find widespread evidence of motivated reasoning —
for approximately 80% of issues tested, responsiveness to corrective information varied
by ideology (unlike Nyhan and Reifler N.d.).
We thus confront conflicting expectations about how people might respond to journal-
1
istic fact-checks during a general election campaign. Citizens may resist fact-checks that
conflict with their partisan or ideological commitments and maintain (or even strengthen)
their misperceptions. Alternately, people might accept journalistic fact-checks and update
their beliefs to be at least somewhat more accurate.
We present results of two studies conducted during the 2016 U.S. presidential cam-
paign that help illuminate this debate. Both studies examine actual misstatements that
were made by candidates and their proxies and fact-checked by the media during the cam-
paign — a time when partisan commitments are activated and the influence of partisan
leaders is likely to be especially strong (e.g., Zaller 1992; Lenz 2012). Specifically, Study
1 is a preregistered survey experiment that evaluates the effects of a journalistic fact-check
of misleading claims about crime made by Donald Trump at the GOP convention. To
increase the realism of the study’s evaluation of the effects of journalistic fact-checking
in a campaign, Study 1 also includes experimental conditions in which a political elite
attempts to denigrate and undermine the fact-check in question. Study 2 tests the effect
of fact-checking a claim Trump made about unemployment during the first presidential
debate among subjects experimentally induced to have watched the debate.
The design of these studies was intended to address two potential explanations for
conflicting findings in the literature. In many cases, respondents in survey experiments
may lack motivation to engage in effortful resistance to unwelcome factual information
about relatively obscure topics or lack sufficient context to form counter-arguments. The
lack of such motivations or context could explain the null or mixed results described above.
To address these concerns, we administered both studies at the height of a presidential
general election campaign and designed them to maximize partisan directional motiva-
tions. In Study 1, we tested corrections of a presidential candidate’s convention accep-
tance speech shortly after it was delivered; in study 2, we tested corrections of a candidate
on the night of a presidential debate among a sample that was encouraged to watch the
2
debate. In addition, while many facts are not the subject of political controversy, we tested
fact-checks of claims that one candidate (Trump) used to criticize the other (Clinton), a
context in which directionally motivated reasoning may be common.
Second, given that counter-arguing of unwelcome political information may be greater
when relevant considerations are available, the first experiment experimentally manipu-
lates the availability of messages denigrating the fact-check. In this study, we randomly
paired a fact-check of a Trump statement with a statement by Trump campaign chairman
Paul Manafort denying the correction and provided an additional Manafort statement as-
cribing a political motivation to the FBI, the source of the data in the fact-check. These
messages should reduce the cognitive demands of counter-argument for low-information
respondents and provide additional considerations upon which high-information Trump
supporters can draw, increasing the realism of the information environment in which the
fact-check is delivered and the likelihood of observing motivated resistance.
The design of these studies also allows us to investigate other important theoretical
questions about the effects of fact-checking. First, we consider whether people are willing
to not only revise their factual beliefs in response to a fact-check but to change their atti-
tudes toward the candidate who has made a claim that has been fact-checked — an effect
that would likely increase the reputational threat that fact-checking poses to politicians
3
(see, e.g., Nyhan and Reifler 2015).1 In addition, our first study tests whether people are
willing to accept attitude-inconsistent information but instead change their interpretations
of that information in a directionally motivated manner. For instance, Gaines et al. (2007)
find that Democrats and Republicans updated their beliefs about the Iraq war relatively ac-
curately over time; it was interpretations of those facts that diverged along partisan lines.
Khanna and Sood (2018) similarly find that incentivized respondents provide more correct
answers but perceived greater unfairness in the study when doing so.
Our results indicate that exposure to fact-checks reduced misperceptions among sup-
porters of both major party presidential candidates but did not affect attitudes toward those
candidates. In Study 1, exposure to fact-checking reduced misperceptions about crime
rates even when respondents were provided a message by a Trump staffer disparaging
the fact-check. Similarly, providing a fact-check of Trump just after a debate in Study 2
reduced misperceptions about unemployment in Michigan and Ohio, even among Trump
supporters. In short, journalistic fact-checks can overcome directionally motivated reason-
ing and bring people’s beliefs more in line with the facts even when the counter-attitudinal
information is disparaged by a co-partisan. However, neither Clinton nor Trump support-
ers changed their attitudes towards either candidate after receiving fact-checks, suggesting
that voters’ preferences during a presidential election are not contingent on their percep-
1 Wintersieck (2017) is a notable exception. There are crucial differences between our
study and hers, however. First, whereas Wintersieck looks at candidates deemed “honest”
by fact-checkers, our studies examine statements flagged by fact-checkers for being false.
Second, while Wintersieck focuses on a statewide election and recruits student subjects at
a university, we enroll broader pools of participants in two experiments about a national
election. This sampling distinction is particularly relevant here given that students might
be more disposed to engage in the effortful cognition required to counterargue unwelcome
information such as fact-checks (Krupnikov and Levine 2014).
4
tions of the factual accuracy of the candidates. In other words, factual corrections can
achieve the limited objective of creating a more informed citizenry but struggle to change
citizens’ minds about whom to support.
Theory and hypotheses
Our theoretical approach builds on research suggests that people have competing goals
in information processing (e.g., Kunda 1990; Molden and Higgins 2005). When people
have stronger accuracy goals (e.g., they are provided a reward for each correct answer
as in Hill 2017), respondents will process information more dispassionately and seek to
maximize the accuracy of their beliefs. When directional motivations are more salient
(e.g., party identity or a preferred candidate), respondents will instead tend to process
information in a manner that is consistent with that preference rather than maximizing
accuracy or considering information in a dispassionate manner (Taber and Lodge 2006;
Bolsen, Druckman, and Cook 2014).
These findings suggest that directionally motivated reasoning may be an especially
salient factor when people process factual information about controversial political issues
in a highly partisan context such as a presidential election campaign. As discussed above,
however, researchers have reached mixed conclusions about the extent to which direc-
tionally motivated reasoning affects belief updating in response to fact-checking in such
contexts (e.g., Nyhan and Reifler 2010; Nyhan, Reifler, and Ubel 2013; Garrett, Nisbet,
and Lynch 2013 versus Weeks 2015; Nyhan and Reifler N.d.; Wood and Porter 2018; Chan
et al. 2017; Porter and Kirby 2018). Other research suggests that directional motivations
may have greater influence on interpretations of contested facts (Gaines et al. 2007) than
on factual beliefs themselves; we also consider this possibility.
To adjudicate among these claims, we test three preregistered hypotheses based on
5
the empirical literature and theoretical considerations discussed above.2 In each case,
we determine whether information is pro- or counter-attitudinal by whether respondents
indicate supporting Trump or Hillary Clinton on a pre-treatment vote choice measure.
First, we propose to test whether fact-checking is ineffective at increasing the accuracy
of factual beliefs among people for whom it is counter-attitudinal (H1; see Nyhan and
Reifler 2010) or whether fact-checks increase belief accuracy but acceptance is greater
among those for whom the information is pro-attitudinal (H2; see Wood and Porter 2018)3
H1 (motivated resistance): Respondents will resist unwelcome facts about controversial

issues. As a result, people exposed to journalistic fact-checks that are counter-
attitudinal will not come to hold more accurate views. In some cases, their views
could even become more inaccurate. (Evaluated in studies 1 and 2.)
H2 (differential acceptance): Respondents will accept journalistic fact-checks that are

counter-attitudinal and update their beliefs to become more accurate, though the
extent to which they update their beliefs may vary based on their prior attitude.
(Evaluated in studies 1 and 2.)
2 Our preregistration for Study 1 documents our hypotheses and analysis plan (URL
omitted for peer review). Unless otherwise noted, all Study 1 analyses are consistent with
this document. Study 2 was conducted too rapidly to be preregistered (it was fielded im-
mediately after the debate) but our analysis follows Study 1 to the greatest extent possible.
3 As discussed above, previous findings are mixed on both hypotheses. For H1, see
Nyhan and Reifler (2010) (backfire on two of five studies) versus Wood and Porter (2018)
(no cases of backfire). For H2, compare Wood and Porter (2018), which finds a consistent
pattern of ideological differentials in belief updating, with Nyhan and Reifler (N.d.), which
finds no evidence of differential acceptance when fact-checks are pro-attitudinal.
6
We also consider the effects of fact-checks on “interpretations” of factual claims. We de-
fine the term consistent with Gaines et al. (2007) who use it to include evaluations (“the
crime rate is high”), explanations (“crime has increased because of a decline in moral val-
ues”), and inferences (“Obama’s acceptance of this crime rate reveals indifference to our
plight”). In this context, respondents may be willing to accept a fact-check which contra-
dicts a favored politician but form an attitude-consistent interpretation of the information
that they have accepted. For instance, if conservative respondents are asked to accept a
contradiction of Trump’s claim about increase in violent crime, they might ascribe the
decline in crime to a factor that is consistent with their beliefs (e.g., tougher policing and
longer sentences). We therefore propose to test whether interpretations of factual claims
are formed and updated in a directionally motivated manner (H3; see Gaines et al. 2007):
H3 (attitude-consistent interpretation): Respondents will accept journalistic fact-checks

that are counter-attitudinal, but interpret them in an attitude-consistent manner.
(Evaluated in study 1 only.)
Finally, as noted above, we also evaluate a key research question — will people not only
revise their factual beliefs, but alter their attitudes toward a candidate who has made a
false or unsupported claim? We therefore also measure attitudes toward the candidate,
including vote preference.4 The voluminous literature on candidate evaluation and vote
choice in a presidential election provides unclear expectations. For instance, Funk (1999)
finds that perceived traits such as honesty condition overall evaluation of candidates in the
expected direction — candidates who are perceived as having integrity are more warmly
regarded. However, Rahn et al. (1990) and Pierce (1993) find no relationship between
perceived traits and vote choice even among sophisticated participants who are thought to
4 Findings for two other preregistered research questions are described below and in the
online appendix.
7
be capable of connecting potentially distal considerations. Accordingly, we have unclear
expectations for the effect of fact-check exposure on subsequent candidate evaluations.
Stimuli
To maximize the external validity of our experiments, we adopted actual candidate state-
ments from the 2016 presidential election. In our experiments, we focus on misleading
claims made by Donald Trump in two high-profile candidate appearances — a suggestion
of dramatically rising crime in his nomination acceptance speech (study 1) and a claim
about the loss of manufacturing jobs during one of the presidential debates (study 2).
Both claims were fact-checked by journalistic outlets at the time. The treatments we use
in the studies aim to match the fact checks published for corresponding misstatements.5
For instance, our fact-check of Trump’s convention statements about the prevalence of
violent crime in Study 1 is very similar to the approach used by journalistic fact checkers.
Here is the Associated Press fact-check of the crime claims Trump made in his convention
speech (Sullivan and Day 2016):
THE FACTS: Violent crime has dropped dramatically since the early 1990s.
According to FBI data , the national violent crime rate last peaked in 1991 at
758 reported violent crimes per 100,000 people. In 2014, the latest year for
which full data is available, the rate was 366 per 100,000 people.
Our fact-check for the same issue in Study 1 is very similar:

5 As we describe below, we also seek to maximize the realism of the treatments we use
to test the effects of elite messages denigrating a fact-check. Study 1 tests the effects of
exposure to actual statements made by Paul Manafort, Trump’s campaign chairman at the
time, challenging the fact-checking of Trump’s convention speech.
8
According to FBI’s Bureau of Justice Statistics, the violent crime rate has
fallen dramatically and consistently over time. According to their estimates,
the homicide rate in the U.S. in 2015 was half that recorded in 1991.
Study 2 considers the effect of fact-checking Trump’s claims about job loss in Michi-
gan and Ohio during the first presidential debate. Politico’s “wrongometer” wrote the
following in response to that claim:
In fact, the state’s unemployment rate has declined in recent years, according
to the Bureau of Labor Statistics. That figure now stands around 4.5 percent,
down from a 14.9 percent unemployment rate in June 2009 (Politico 2016).
Similarly, the New York Times wrote the following about Trump’s claim:
Ohio and Michigan have, indeed, suffered major manufacturing job losses
over the past generation. But in the past year, Ohio has gained 78,300 jobs,
and Michigan has gained 75,800 jobs. In August, the unemployment rate was
4.9 percent in Michigan and 4.7 percent in Ohio, both in line with the national
rate (New York Times 2016).
Our fact-check for this same issue (described further below) reads as follows:
In fact, according to the Bureau of Labor Statistics, unemployment has fallen

in both Michigan and Ohio. Both states each saw 70,000 new jobs over the
last year.
We believe our treatments are faithful representations of the type of journalistic fact-
checking that is now widely disseminated by the media. In other words, our experiment
9
is not a test of authoritative and logically infallible attempts to debunk false claims. It is
instead an analysis of how journalistic fact-checks affect mass beliefs and attitudes.6
Study 1: Crime perceptions
Study 1 tested the effects of journalistic fact-checks about changes in levels of crime over
time. Respondents were randomly assigned to one of five conditions. In a control condi-
tion, respondents read an article without any political content. Participants in the treatment
conditions read a mock news article featuring misleading claims by Donald Trump about
crime rates based on an actual article that originally appeared on CNN.com. In one treat-
ment condition, participants read an article that omits any corrective information, allowing
us to test the effect of exposure to a candidate’s statement alone. Others were assigned to
read versions of the article that include neutral corrective information in the style of jour-
nalistic fact-checking, allowing us to test its effects versus a no-correction version of the
6 It is important to note that journalistic fact-checks do not always logically contradict a
speaker (e.g., Marietta and Bowser 2015; Uscinski and Butler 2013). Fact-checkers often
seek instead to address possible inferences that listeners might draw from a candidate’s
statement. For instance, Trump’s nomination speech described an “epidemic” of violent
crime. He did not directly state that crime has increased, but a listener might infer as much
(indeed, Trump made clear statements about increasing crime rates at other times). Like
other journalistic fact-checks, our treatment thus cites FBI data on the long-term decline
in violent crime. Similarly, Trump’s debate statement emphasized factory jobs leaving
Ohio and Michigan. While he did not directly say that employment in Michigan and Ohio
is suffering because of trade policy, he implied that widespread job loss was taking place.
Consequently, our fact-check, like several in the media, provided data on changes in jobs
and unemployment in those states.
10
article as well as the control condition. One condition tested the fact-check alone, while
two other conditions tested whether elite political actors can cultivate resistance to factual
information. In these conditions, the fact-check was followed by a statement from a Trump
surrogate disparaging the validity of the information or a statement by the surrogate dis-
paraging the validity of the information and attributing a political motive to the source of
the information.
Experimental design and instrument
After a series of demographic and attitudinal questions, participants were assigned to one
of five conditions.7 The treatments were based on actual news events during the 2016
Republican National Convention. In his speech, Trump described an America ridden by
increasing crime (uncorrected claim). As the media pointed out, however, these depictions
were contradicted by FBI crime data showing a long-term secular decline (journalistic
fact-check). When Paul Manafort, Trump’s then-campaign chairman, was pressed about
this discrepancy (Schleifer 2016), he questioned the validity of FBI statistics (rejection
of the fact-check) and suggested the FBI could not be trusted because it did not recom-
mend indicting Hillary Clinton in her email scandal (conspiracy theory/fact-check source
derogation).
The specific treatments shown to participants are as follows:
• Control: A birdwatching article.
• Rising crime message: A news article summarizing Trump’s claims
• Rising crime message + fact-check: A news article summarizing Trump’s claims

with a fact-check citing FBI statistics.
7 The full instrument is in Online Appendix A.
11
• Rising crime message + fact-check + denial: A news article summarizing Trump’s
claims with a fact-check citing FBI statistics and a statement from Manafort rejecting
the statistics.
• Rising crime message + fact-check + denial + source derogation: A news article

summarizing Trump’s claims about crime with a fact-check citing FBI statistics, a
statement from Manafort rejecting the statistics, and Manafort’s contention that the
FBI was not to be trusted.
Outcome variables
After treatment, subjects were asked two article recall questions to measure receipt of
treatment. We then measured several outcome variables of interest. To assess motivated
resistance, we asked people’s beliefs about changes in the crime rate. We also asked
about perceptions of the accuracy of federal crime statistics and the treatment article and
whether/how the treatment article was biased. In addition, respondents were asked to
choose among interpretations of crime trends, which were coded for belief consistency.8
Finally, we measured evaluations of Trump and other politicians.
Results
Responses were collected on September 30, 2016 by Morning Consult (n = 1, 203) and on
Mechanical Turk (n = 2, 983).9 To assess our hypotheses, we performed a series of OLS
8 Per our preregistration, respondents who indicated crime was up due to inequality or
unemployment were coded as -1 (liberal), those who said crime was up due to moral de-
cline or down due to tougher policing were coded as 1 (conservative), and other responses
were coded as 0.
9 Demographic and balance data for both samples are provided in Online Appendix C.
12
regressions with robust standard errors.10 All treatment effect estimates are unweighted
intent to treat effects.11 In the pre-treatment vote choice questions, a small percentage of
respondents did not support Trump or Clinton; per our preregistration, we exclude them
from subsequent analyses.12
To evaluate motivated resistance, we test whether the marginal effect of exposure to
the fact-check conditions on misperceptions about crime is null or positive for Trump sup-
porters and whether the difference in effects is significant compared to Clinton supporters.
To evaluate differential acceptance, we test whether the marginal effect of exposure to the
conditions including a fact-check is negative for Trump supporters and whether the differ-
ence in treatment effects is significant versus Clinton supporters. We also measure source
derogation and counterargument to understand how respondents respond to fact-checks.
Table 1 presents the effects of the manipulation on beliefs about changes in crime where
the control condition is the excluded category. These estimates are calculated among
Trump and Clinton supporters (the latter are thus the excluded category for the Trump
support indicator). The table also reports auxiliary quantities representing differences
versus the uncorrected statement condition.
Our results indicate that journalistic fact-checking had a pronounced effect on fac-
10 All analyses in this section are consistent with our preregistration unless otherwise
indicated. OLS models are replicated using ordered probit where applicable in Online
Appendix C.
11 Mean scores on two attention checks were 1.62 and 1.92 for controls and 1.59 and
1.87 for the treatment groups on Morning Consult and Mechanical Turk, respectively. (See
Online Appendix A for wording.) We therefore deviate slightly from our preregistration
to omit consideration of response time as a measure of attention.
12 We report equivalent but more complex models estimated on the full sample in Online
Appendix C.
13
Table 1: Message exposure effects on beliefs about changes in crime
Morning Consult Mechanical Turk

Trump support 0.52*** 0.81***
(0.16) (0.11)
Uncorrected statement -0.34* -0.05
(0.18) (0.09)
Uncorrected ⇥ Trump support 0.35 0.16
(0.23) (0.16)
Fact-check -0.94*** -0.97***
(0.19) (0.10)
Fact-check ⇥ Trump support 0.09 0.11
(0.26) (0.18)
Fact-check denial -0.74*** -0.94***
(0.19) (0.09)
Fact-check denial ⇥ Trump support 0.21 0.38**
(0.27) (0.17)
Denial/source derogation -1.12*** -0.77***
(0.19) (0.10)
Denial/source derogation ⇥ Trump support 0.61** 0.29*
(0.27) (0.17)
Constant 3.64*** 3.01***
(0.12) (0.06)
Fact-check uncorrected statement
Clinton supporters -0.61*** -0.92***
(0.19) (0.10)
Trump supporters -0.86*** -0.97***
(0.18) (0.15)
Denial uncorrected statement
Clinton supporters -0.41** -0.89***
(0.19) (0.09)
(0.19) (0.14)
Denial/derogation uncorrected statement
(0.19) (0.10)
(0.20) (0.14)
R2 0.14 0.19
N 990 2430
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided). OLS models with robust standard
errors. All independent variables are dichotomous indicators. Dependent variable is a five-
point scale of perceived change in crime incidence where higher values indicate perceived
greater incidence. Participants are respondents from Morning Consult or Mechanical Turk
who supported Hillary Clinton or Donald Trump in the 2016 election (reference category
for Trump support is thus Clinton support).
14
tual beliefs. Though Trump’s supporters were more likely than Clinton’s to believe that
crime had increased or not declined significantly over the previous ten years, corrective
information reduced misperceptions among supporters of both candidates. Specifically,
respondents exposed to FBI statistics about decreased crime reported significantly lower
misperceptions compared to the uncorrected statement conditions in both samples regard-
less of the candidate they supported (p < .01). Exposure to Trump’s claim did not further
increase misperceptions among his supporters.
Our results are inconsistent with motivated resistance. We do observe some evidence
of differential acceptance, however. Both the fact-check denial on Mechanical Turk (p <
.05) and the denial/source derogation in both samples (p < .05 in Morning Consult, p <
.10 in Turk) reduce crime misperceptions less among Trump supporters than Clinton sup-
porters.13
These findings are illustrated in Figure 1, which presents mean crime perceptions by
condition and candidate support from the Morning Consult data. Mean beliefs about crime
change among Trump supporters declined from 4.17 (out of 5) in the control and uncor-
rected statement conditions to 3.31 in the fact-check condition, 3.62 in the fact-check and
denial condition, and 3.65 in the fact-check/denial/source derogation condition. Similar
declines are observed among Clinton supporters.
To understand these responses, we examine how the treatments affect judgments of
the accuracy and fairness of the articles. Table 2 demonstrates that fact-checks provoke
different perceptions of accuracy and fairness among Clinton and Trump supporters. More
specifically, this table presents results from statistical models available in Table C6 in
Online Appendix C that estimate the effect of the manipulation on the perceived fairness
and accuracy of the stimulus article and the perceived accuracy of federal crime statistics
13 These quantities are estimated with respect to the control condition. These differences
are not significant relative to the uncorrected condition.
15
Figure 1: Crime perceptions by treatment condition and candidate support
Mean beliefs about crime change by candidate preference and experimental condition.
Survey data from Morning Consult.
for Clinton and Trump supporters relative to the uncorrected condition.

Relative to the uncorrected statement condition, Clinton supporters view the article as
more accurate and fair when a fact-check is present. In three of four comparisons, they
even view federal crime statistics as more accurate when Trump’s staffer questions them.
By contrast, Trump supporters view the article as less accurate and fair when it includes a
fact-check — a contrast with their reported factual beliefs, which became more accurate.
Trump supporters are also less likely to view federal crime statistics as accurate when they
are invoked in a fact-check, especially when questioned by a Trump staffer (p < .01 in
each denial condition versus the uncorrected condition).
To illustrate these findings, we plot the means of the perceived accuracy of the stimu-
lus article and federal crime statistics by condition and candidate support for the Morning
Consult data in Figure 2. When Trump supporters receive a fact-check, they are less likely
to see the article or federal crime statistics as accurate (mean of 2.7 for both) compared
to when they receive Trump’s uncorrected statement (means of 2.8 and 3.0, respectively).
16
Figure 2: Perceived accuracy of article and federal crime statistics by treatment condition
and candidate support
Mean perceived accuracy and fairness of the stimulus article and perceived accuracy of
federal crime statistics by candidate preference and experimental condition. Survey data
from Morning Consult.
17
Table 2: Message exposure effects on perceptions of accuracy and fairness
Article accurate Article fairness Statistics accurate

MC MT MC MT MC MT
Fact-check uncorrected
Clinton supporters 0.40*** 0.65*** 0.26** 0.56*** 0.04 0.08*
(0.11) (0.05) (0.11) (0.06) (0.09) (0.04)
Trump supporters -0.17* -0.21*** -0.24*** -0.30*** -0.33*** -0.11
(0.10) (0.08) (0.09) (0.07) (0.10) (0.07)
Denial uncorrected
Clinton supporters 0.43*** 0.69*** 0.28** 0.47*** 0.12 0.14***
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
Trump supporters -0.19* -0.06 -0.37*** -0.26*** -0.46*** -0.21***
(0.11) (0.08) (0.11) (0.07) (0.11) (0.07)
Denial/derogation uncorrected
Clinton supporters 0.47*** 0.68*** 0.24** 0.41*** 0.15* 0.11**
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
Trump supporters -0.16 -0.10 -0.23** -0.22*** -0.34*** -0.16**
(0.10) (0.07) (0.10) (0.07) (0.10) (0.07)
* p < 0.10, ** p < 0.05, *** p < .01. Auxiliary quantities from OLS models with robust
standard errors reported in Table C6. All independent variables are dichotomous indica-
tors. For each dependent variable, higher values indicate greater accuracy and fairness
(each measured on a four-point scale). Participants are respondents from Morning Con-
sult (MC) or Mechanical Turk (MT) who supported Hillary Clinton or Donald Trump in
the 2016 election (reference category for Trump support is thus Clinton support).
The opposite pattern is frequently observed among Clinton supporters, who view the ar-
ticle as more accurate when a fact-check is included (mean of 2.9 versus 2.5). These re-
sponses do not vary in the presence of fact-check denial or source derogation; the response
seems driven by the presence of a fact-check.14
The differences in perceptions of the articles that we observe can be interpreted as
consistent with H3. However, as Table C11 in Online Appendix shows, we do not find
evidence that people interpret the changes in crime they perceive in a viewpoint-consistent
manner (e.g., “[t]ougher policing and longer prison sentences” for Trump supporters who
think crime has decreased).
Finally, we estimate the marginal effect of fact-checking on evaluations of Trump
14 Findings are similar for perceived article bias (see Online Appendix C).
18
among Clinton and Trump supporters in Table 3. Specifically, we asked subjects to eval-
uate both candidates on a 1–5 scale ranging from “Very unfavorable” to “very favorable.”
We find no significant effects of the fact-check on favorability toward Trump regardless of
respondents’ candidate preference.15
In sum, the evidence from Study 1 shows that, even if people are inclined to take a
skeptical view of a fact-checking article and the data underlying it, fact-checks can still
spur people to hold more factually accurate beliefs. However, these changes in belief
accuracy do not seem to lead to corresponding changes in attitudes toward the candidate
being fact-checked.
Study 2
One limitation of Study 1 is that we examined the effects of a fact-check several weeks after
the misstatement it targeted was made. It is possible that subjects had already been exposed
to fact-checking of the misstatements we studied and that this exposure limited the effects
of the fact-check on attitudes toward the candidate. The design of Study 2 addresses this
limitation because it was conducted immediately after the first 2016 presidential debate.
In the first wave of Study 2, 1,546 participants from Mechanical Turk were asked stan-
dard political and demographic questions as well as questions about their access to cable
television and media consumption. They were then instructed to watch the debate and told
15 In Table C19 in Online Appendix C, we show that the manipulation had no effect on
favorability toward Clinton or Barack Obama either.
19
Table 3: Message exposure effects on Trump favorability

(0.14) (0.10)
Uncorrected statement -0.03 -0.07
(0.13) (0.06)
(0.20) (0.14)
Fact-check 0.10 -0.12*
(0.13) (0.06)
Fact-check ⇥ Trump support -0.11 -0.03
(0.20) (0.14)
Fact-check denial 0.10 -0.05
(0.13) (0.07)
Denial ⇥ Trump support 0.01 -0.09
(0.20) (0.14)
Denial/source derogation -0.13 -0.05
(0.11) (0.07)
Derogation ⇥ Trump support 0.17 -0.06
(0.21) (0.15)
Constant 1.42*** 1.36***
(0.09) (0.05)
Clinton supporters 0.14 -0.05
(0.13) (0.06)
Trump supporters -0.10 -0.21
(0.14) (0.13)
Clinton supporters 0.13 0.02
(0.13) (0.06)
Trump supporters 0.01 -0.19
(0.16) (0.14)
Clinton supporters -0.09 0.02
(0.11) (0.06)
(0.18) (0.14)
R2 0.62 0.58
N 989 2430
* p < 0.10, ** p < 0.05, *** p < .01. OLS models with robust standard errors. All
independent variables are dichotomous indicators. The dependent variable is a five-point
scale where higher values indicate greater favorability. Participants are respondents from
Morning Consult or Mechanical Turk who supported Hillary Clinton or Donald Trump in
the 2016 election (reference category for Trump support is thus Clinton support).
20
they would be invited to take a survey immediately after its conclusion.16 As soon as the
debate ended, participants were invited to complete a survey that included questions about
candidates’ debate performances and respondents’ general attitudes towards the candidates
(Wave 2). It included the following statement from Trump:
Our jobs are fleeing the country to Mexico... they’re building some of the
biggest, most sophisticated plants. Not so much in America. Thousands of
jobs leaving Michigan, Ohio...their companies are just leaving, they’re gone.17
The claim that large numbers of jobs were leaving Michigan and Ohio at the time due to
factories being moved abroad is inaccurate. As fact-checkers pointed out on the night of the
debate, employment had not fallen in either state. The New York Times fact-check pointed
to the number of new jobs created in each state over the previous year (New York Times
2016). A fact-check by National Public Radio (NPR) of this Trump statement directed
readers to the Bureau of Labor Statistics (BLS) for data confirming the fact check—and
rebutting Trump (National Public Radio 2016).
After seeing Trump’s claim, respondents were randomly assigned with probability .5
to receive the following fact-check: “In fact, according to the Bureau of Labor Statistics,
16 Because the broader experiment in which Study 2 was embedded was designed to
examine how post-debate news coverage affected debate perceptions, participants were
assigned to one of five content consumption conditions that were orthogonal to the ran-
domization we examine here (C-SPAN with no post-debate coverage, Fox News with or
without post-debate coverage, or MSNBC with or without post-debate coverage). We ex-
cluded subjects who did not have access to cable and block-randomized by party and pre-
ferred cable channel. For additional details, consult (reference omitted for peer review).
17 The instrument was prepared before transcripts were available, so the statement in our
study differs slightly from the official transcript.
21
unemployment has fallen in both states. Both states each saw 70,000 new jobs over the
last year.” Like The New York Times’ fact-check on the night of the debate, we pointed
to the number of new jobs created in each state over the previous year; and like NPR’s
fact-check, we explicitly based the claim of the fact check on BLS data.
All respondents then were asked: “Over the last few years, has unemployment gone
up or down in Michigan and Ohio?” Subjects could respond on a five-point scale, from
“Gone down a lot” to “Gone up a lot,” with “Stayed the same” as the middle category. (See
Online Appendix A for exact wording.) This wave was closed by noon the next day. Five
days later, we recontacted participants and measured perceptions of the debate’s winner
and attitudes toward the candidates (Wave 3).18
Study 2 allows us to test for both motivated resistance (H1) and differential acceptance
(H2). Under H1 (motivated resistance), Trump supporters exposed to the fact-check would
not only resist, but come to hold more inaccurate views than uncorrected Trump support-
ers. The fact-check treatment undermined Trump’s claims about the economy and his
opposition to foreign trade, which were central to his candidacy. The claim’s political im-
portance and its clear contradiction by government data makes it a useful test of partisans’
18 See Online Appendix C for details on participant demographics and experimental
balance. Though we cannot rule out the possibility of post-treatment bias (Montgomery,
Nyhan, and Torres 2018), we find no significant effect of treatment assignment at wave 2
on wave 3 participation in a simple OLS model (b = 0.05, p > .10).
22
responsiveness to fact-checking.19 Under H2 (differential acceptance), Trump and Clinton
supporters who are exposed to the fact-check would both integrate this information, but
Trump supporters would be less accepting of it.
Analysis
We begin with respondents’ beliefs about unemployment in Michigan and Ohio after the
fact-check in Wave 2. Following Study 1, we estimate OLS models with robust standard
errors that include indicators for fact-check exposure and Trump support in Wave 1 and an
interaction term.20 As in Study 1, we restrict our analysis to respondents who reported sup-
porting Clinton or Trump in Wave 1. The outcome measure is coded so that higher values
indicate belief that unemployment stayed the same or increased rather than decreased.
We present results in Table 4. The first model shows that the fact-check decreased mis-
perceptions about unemployment in Michigan and Ohio among both Clinton and Trump
supporters (p < .01). As with Study 1, the evidence from Study 2 does not support moti-
vated resistance (H1). We also find no evidence of differential acceptance, contradicting
H2. The second model shows that these results are consistent when we control for the
19 Such fact-checks are common. For instance, more than 60% of the claims rated by
PolitiFact and the Washington Post Fact Checker were found to be mostly or totally false by
both fact-checkers (Lim 2018). Moreover, fact-checkers consider it part of their mission to
check claims against official data sources and frequently do so. Graves (2016, 85) writes,
for instance, that “Fact-checkers always seek official data and often point to examples
like this [a fact-check assessing claims about government spending and job growth using
federal data] to explain what they do.”
20 The design does not include a control condition or fact-check denial and denial/source
derogation conditions. The omitted category is an uncorrected statement.
23
Table 4: Message exposure effects on unemployment beliefs
(1) (2)
Trump support 0.44⇤⇤⇤ 0.45⇤⇤⇤
(0.13) (0.13)
Fact-check 0.34⇤⇤⇤ 0.34⇤⇤⇤
(0.11) (0.11)
(0.18) (0.18)
Condition: MSNBC (post-debate coverage) 0.06
(0.14)
Condition: MSNBC (no post-debate coverage) 0.03
(0.14)
Condition: Fox (post-debate coverage) 0.13
(0.14)
Condition: Fox (no post-debate coverage) 0.04
(0.15)
Constant 2.70⇤⇤⇤ 2.71⇤⇤⇤
(0.07) (0.12)
(0.11) (0.11)
(0.14) (0.14)
R2 0.05 0.05
N 825 825
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided). OLS models with robust standard
errors. Respondents are Mechanical Turk workers who supported Hillary Clinton or Don-
ald Trump and were assigned to watch the first presidential debate in 2016 (reference
category for Trump support is thus Clinton support). All independent variables are di-
chotomous indicators. The dependent variable is a five-point scale of perceived change in
unemployment in Ohio and Michigan where higher values indicate a perceived increase
in unemployment. The omitted category for the media manipulation is C-SPAN with no
post-debate coverage.
orthogonal media manipulation.

Finally, Table 5 considers fact-check effects five days later (Wave 3). We again use OLS
models with robust standard errors to estimate whether exposure to a fact-check affected
perceptions that Trump won the debate, evaluations of his debate performance, and vote
choice. We again include indicators for fact-check exposure and Trump support and an in-
teraction term. We see no evidence that the fact-check affected these outcomes except for
24
perceptions that Trump “won” the debate, which declined slightly among his supporters.
However, vote choice was not affected. Echoing Study 1, fact-checking reduced misper-
ceptions but had no discernible effects on participants’ candidate preferences, including
supporters of the candidate who had been fact-checked.
Table 5: Fact-check effects at Wave 3
Trump won debate Trump favorability Trump vote

Trump support (W1) 0.67⇤⇤⇤ 0.47⇤⇤⇤ 0.89⇤⇤⇤
(0.04) (0.02) (0.03)
Fact-check 0.02 0.01 0.001
(0.03) (0.02) (0.02)
Fact-check ⇥ Trump vote (W1) 0.17⇤⇤⇤ 0.06 0.01
(0.06) (0.04) (0.04)
Constant 0.02 0.16⇤⇤⇤ 0.03⇤
(0.02) (0.02) (0.02)
Clinton supporters -0.02 -0.01 -0.001
(0.03) (0.02) (0.02)
Trump supporters -0.15** -0.05 -0.01
(0.05) (0.03) (0.03)
R2 0.47 0.53 0.81
N 527 527 527
* p < 0.10, ** p < 0.05, *** p < .01. OLS models with robust standard errors. All in-
dependent variables are dichotomous indicators. The dependent variables are measured
on a five-point scale where higher values indicate greater belief that Trump won the de-
bate, a more favorable attitude toward Trump, or a greater intention to vote for Trump,
respectively. Respondents are Mechanical Turk workers who supported Hillary Clinton or
Donald Trump in the 2016 election (reference category for Trump support is thus Clinton
support). All outcome measures were collected in wave 3.
Conclusion
In two studies of fact-checking conducted during the 2016 presidential election, we find
that people express more factually accurate beliefs after exposure to fact-checks. These ef-
fects hold even when fact-checks target their preferred candidate. In Study 1, we exposed
participants to variants of an article covering claims Donald Trump made about crime.
25
Trump supporters were willing to accept a fact-check of those claims and update their be-
liefs, though we observe some evidence of differential acceptance (i.e., they revised their
beliefs less than Clinton supporters). Similarly, the accuracy of Trump supporters’ be-
liefs about unemployment increased in Study 2 after seeing a fact-check of a claim Trump
made during the first debate. However, exposure to journalistic fact-checks did not affect
attitudes toward him in either study. Ultimately, we find no evidence that changes in fac-
tual beliefs in a claim made by a candidate affect voter preferences during a presidential
election.
Our results on interpretations were mixed. Study 1 participants evaluated the fairness
and accuracy of the stimulus article and the accuracy of the federal crime statistics it cited
in a directionally motivated fashion. However, they did not adopt viewpoint-consistent in-
terpretations of the change in crime. Further research is needed to understand how respon-
dents interpret counter-attitudinal fact-checks and other forms of corrective information.
These results help theoretically inform our understanding of both motivated reasoning
and factual belief updating. As with other recent research (Wood and Porter 2018), we find
little evidence of a backfire effect on respondents’ factual beliefs. However, our results
also do not suggest that respondents accept fact-checks uncritically; exposure to counter-
attitudinal information decreases perceptions of the accuracy of our stimulus article and
the source of counter-attitudinal information in Study 1. These findings suggest motivated
reasoning can coexist with belief updating, a possible explanation for divergent findings
in the literature (e.g., Flynn 2016); they need not be mutually exclusive phenomena (see
also Coppock and Guess N.d. and Khanna and Sood 2018).
Of course, these studies have several limitations. First, we did not test a fact-check
of a Clinton misstatement and cannot evaluate how her supporters would have reacted.
Second, Trump was infamous for extreme exaggerations and misstatements, which may
have made some respondents receptive to fact-checking but also prepared his supporters
26
to rationalize their continued support for him. Also, we measured feelings toward the
candidates post-treatment and use them as an outcome measure. It is possible that that the
strongest Trump supporters respond to fact-checks differently than less ardent supporters,
but we are unable to test this conjecture.
Further research is also necessary to determine the extent to which our results gen-
eralize to other contexts or forms of fact-checking. Our results suggest that fact-checks
are unlikely to meaningfully diminish the strong attachments people have to their party’s
candidate in a campaign context with strong partisan cues. Their effects may be stronger,
however, in elections with weaker partisan cues and less well-known figures on the ballot
(e.g., Wintersieck 2017). In addition, fact-checks could have stronger attitudinal effects
on feelings toward politicians the more temporally distant they are from an election. In
addition, the scope of our finding is limited to the effects of a fact-check of a single mis-
statement, the format most often employed by journalists. Correcting a series of inaccurate
claims might have had larger effects on candidate evaluations. Other research might also
consider the effects of joint fact-checking of both major-party candidates or of positive
fact-checks corroborating the accuracy of a candidate’s claim. In addition, future research
should examine affective and emotional responses to candidate misstatements of fact as
well as the fact-checks that set the record straight.
Finally, as with all studies of this sort, we cannot completely rule out acquiescence
bias or demand effects. Recent research indicates that demand effects in Internet-based
survey experiments are relatively rare, however (Mummolo and Peterson N.d.).
Despite these limitations, our results provide compelling evidence that citizens can
accept the conclusions of journalistic fact-checks of misstatements even when they are
made by one’s preferred candidate during a presidential election. This information had
little effect on people’s attitudes toward the candidate being corrected, however. In other
words, Trump supporters took fact-checks literally, but not seriously enough to affect how
27
they felt toward their preferred candidate.
References
BBC. 2016. “‘Post-truth’ declared word of the year by Oxford Dictionaries.” Novem-
ber 16, 2016. Downloaded February 6, 2017 from http://www.bbc.com/news/
uk-37995600.
Bolsen, Toby, James N Druckman, and Fay Lomax Cook. 2014. “The influence of partisan
motivated reasoning on public opinion.” Political Behavior 36 (2): 235–262.
Chan, Man-pui Sally, Christopher R. Jones, Kathleen Hall Jamieson, and Dolores Albar-
racín. 2017. “Debunking: A Meta-Analysis of the Psychological Efficacy of Messages
Countering Misinformation.” Psychological Science 28 (11): 1531–1546.
Coppock, Alexander, and Andrew Guess. N.d. “Does Counter-Attitudinal Information

Cause Backlash? Results from Three Large Survey Experiments.” Forthcoming, British
Journal of Political Science.
Flynn, D.J. 2016. “The Scope and Correlates of Political Misperceptions in the Mass
Public.” Unpublished paper, Dartmouth College.
Funk, Carolyn L. 1999. “Bringing the candidate into models of candidate evaluation.” The
Journal of Politics 61 (3): 700–720.
Gaines, Brian J., James H. Kuklinski, Paul J. Quirk, Buddy Peyton, and Jay Verkuilen.
2007. “Same facts, different interpretations: Partisan motivation and opinion on Iraq.”
Journal of Politics 69 (4): 957–974.
Garrett, R. Kelly, Erik C. Nisbet, and Emily K. Lynch. 2013. “Undermining the corrective
28
effects of media-based political fact checking? The role of contextual cues and naïve
theory.” Journal of Communication 63 (4): 617–637.
Graves, Lucas. 2016. Deciding what’s true: The rise of political fact-checking in American
journalism. Columbia University Press.
Hill, Seth J. 2017. “Learning together slowly: Bayesian learning about political facts.”
The Journal of Politics 79 (4): 1403–1418.
Hochschild, Jennifer L., and Katherine Levine Einstein. 2015. Do Facts Matter? Informa-
tion and Misinformation in American Politics. Norman, OK: University of Oklahoma
Press.
Jamieson, Kathleen Hall. 2015. “Implications of the Demise of ‘Fact’ in Political Dis-
course.” Proceedings of the American Philosophical Society 159 (1): 66–84.
Jarman, Jeffrey W. 2016. “Motivated to Ignore the Facts: The Inability of Fact-Checking
to Promote Truth in the Public Sphere.” In Truth in the Public Sphere, ed. Jason Hannan.
Lexington Books.
Khanna, Kabir, and Gaurav Sood. 2018. “Motivated Responding in Studies of Factual
Learning.” Political Behavior 40 (1): 79–101.
Krupnikov, Yanna, and Adam Seth Levine. 2014. “Cross-sample comparisons and external
validity.” Journal of Experimental Political Science 1 (1): 59–80.
Kunda, Ziva. 1990. “The case for motivated reasoning.” Psychological bulletin 108 (3):
480–498.
Lenz, Gabriel S. 2012. Follow the leader? How voters respond to politicians’ performance
and policies. Chicago, IL: University of Chicago Press.
29
Lim, Chloe. 2018. “Checking How Fact-Checkers Check.” Research & Politics 5 (3).
Marietta, Morgan, David C. Barker, and Todd Bowser. 2015. “Fact-Checking Polarized
Politics: Does the Fact-Check Industry Provide Consistent Guidance on Disputed Re-
alities?” The Forum: A Journal of Applied Research in Contemporary Politics 13 (4):
577-596.
Molden, Daniel C., and E. Tory Higgins. 2005. “Motivated thinking.” In The Cambridge
handbook of thinking and reasoning, ed. Keith James Holyoak and Robert G. Morrison.
Cambridge University Press.
Montgomery, Jacob M., Brendan Nyhan, and Michelle Torres. 2018. “How Condition-
ing on Posttreatment Variables Can Ruin Your Experiment and What to Do about It.”
American Journal of Political Science 62 (3): 760–775.
Mummolo, Jonathan, and Erik Peterson. N.d. “Demand Effects in Survey Experiments:
An Empirical Assessment.” SSRN. Downloaded September 30 2018 from https://
papers.ssrn.com/sol3/papers.cfm?abstract_id=2956147.
National Public Radio. 2016. “Fact Check: Trump And Clinton De-
bate For The First Time.” September 26, 2016. Downloaded February
15, 2017 from http://www.npr.org/2016/09/26/495115346/
fact-check-first-presidential-debate.
New York Times. 2016. “Our Fact Checks of the First Debate.” September 26th, 2016.
Downloaded July 27, 2018 from https://www.nytimes.com/2016/09/27/
us/politics/fact-check-debate.html.
Nyhan, Brendan, and Jason Reifler. 2010. “When Corrections Fail: The persistence of
political misperceptions.” Political Behavior 32 (2): 303–330.
30
Nyhan, Brendan, and Jason Reifler. 2015. “The Effect of Fact-Checking on Elites: A Field
Experiment on US State Legislators.” American Journal of Political Science 59 (3):
628–640.
Nyhan, Brendan, and Jason Reifler. N.d. “Do People Actually Learn From Fact-
Checking? Evidence from a longitudinal study during the 2014 campaign.” Unpub-
lished manuscript. Downloaded June 28, 2017 from http://www.dartmouth.
edu/~nyhan/fact-checking-effects.pdf.
Nyhan, Brendan, Jason Reifler, and Peter A. Ubel. 2013. “The hazards of correcting myths
about health care reform.” Medical Care 51 (2): 127–132.
Pierce, Patrick A. 1993. “Political sophistication and the use of candidate traits in candi-
date evaluation.”.
Politico. 2016. “Trump wrong on Michigan job losses.” September 26, 2016.
Downloaded November 11, 2017 from https://www.politico.com/
blogs/2016-presidential-debate-fact-check/2016/09/
trump-wrong-on-michigan-job-losses-228707.
Porter, Ethan, Thomas J. Wood, and David Kirby. 2018. “Sex Trafficking, Russian Infiltra-
tion, Birth Certificates, and Pedophilia: A Survey Experiment Correcting Fake News.”
Journal of Experimental Political Science 2 (5): 304–331.
Rahn, Wendy M., John H. Aldrich, Eugene Borgida, and John L Sullivan. 1990. “A Social
Cognitive Model of Candidate Appraisal.” In Information and Democratic Processes,
ed. John A. Ferejohn and James H. Kuklinski. University of Illinois Press.
Schleifer, Theodore. 2016. “Paul Manafort doubts FBI statistics af-

ter agency spared Hillary.” CNN, July 12, 2016. Downloaded February
31
13, 2017 from http://www.cnn.com/2016/07/21/politics/
paul-manafort-fbi-statistics-hillary-clinton/.
Spivak, Carey. 2011. “The Fact-Checking Explosion.” American Journalism Review 32:
38–43.
Sullivan, Eileen, and Chad Day. 2016. “AP FACT CHECK: Crime stats don’t back
Trump’s dire view.” Associated Press, July 13, 2016. Downloaded October 22, 2018
from https://apnews.com/3e132f145e0c44cf96cb7f4fd448b34a.
Taber, Charles S., and Milton Lodge. 2006. “Motivated Skepticism in the Evaluation of
Political Beliefs.” American Journal of Political Science 50 (3): 755–769.
Uscinski, Joseph, and Ryden Butler. 2013. “The Epistemology of Fact Checking.” Critical
Review 25 (2): 162–180.
Weeks, Brian E. 2015. “Emotions, partisanship, and misperceptions: How anger and anx-
iety moderate the effect of partisan bias on susceptibility to political misinformation.”
Journal of Communication 65 (4): 699–719.
Wintersieck, Amanda L. 2017. “Debating the truth: The impact of fact-checking during
electoral debates.” American Politics Research 45 (2): 304–331.
Wood, Thomas, and Ethan Porter. 2018. “The Elusive Backfire Effect: Mass Attitudes’
Steadfast Factual Adherence.” Political Behavior.
Zaller, John. 1992. The Nature and Origins of Mass Opinion. Cambridge University Press.
32
Online Appendix A
Study 1 instrument
The introductory sections of the Morning Consult and Mechanical Turk questionnaires
differed slightly, which we noted was a possibility in our preregistration. Here we present
the pre-experimental components of both questionnaires. The experimental component is
identical between the two studies. To conserve space, it is therefore presented only once.
Mechanical Turk questionnaire introductory section
[Consent text]
Do you consent to participate in the interview?
-Yes
-No
What is your age?

-Under 18
-18-24
-25–34
-35–44
-45–54
-55–64
-65–74
-75–84
-85 or older
In which state do you currently reside? [Drop down menu]
What is your gender?

-Male
-Female
-Other
If the 2016 presidential election were held today and the candidates were Democrat Hillary
A-1
Clinton and Republican Donald Trump, for whom would you vote?
-Hillary Clinton
-Donald Trump
-Don’t know
-No opinion
What is your annual household income?

-Under 20 thousand dollars
-20 to under 35 thousand
-35 to under 50 thousand dollars
-100 thousand or more dollars
[If “100 thousand or more dollars” is selected] Is that...

-250 thousand or more
Generally speaking, do you think of yourself as...

-Republican
-Democrat
-Independent
-Something else
[If “Democrat” is selected] Would you call yourself...

-A strong Democrat?
-A not very strong Democrat?
[If “Republican” is selected] Would you call yourself...

-A strong Republican?
-A not very strong Republican?
A-2
[If “Independent” or “Something else” is selected] Do you think of yourself as closer to...
-the Democratic Party
-the Republican Party
-Neither
Thinking about politics these days, how would you describe your political viewpoint?
-Very liberal
-Liberal
-Slightly liberal
-Moderate
-Slightly conservative
-Conservative
-Very conservative
-Don’t know
What is the last grade or class you completed in school?

-None, or grade 1–8
-High school incomplete (grades 9–11)
-High school diploma or equivalent, no further schooling
-Technical or vocational school after high school
-Some college, no degree
-Associate’s or two-year college degree
-Four-year college degree
-Graduate or professional school after college, no degree
-Graduate or professional degree
Are you yourself of Hispanic origin or descent, such as Mexican, Puerto Rican, Cuban, or
or some other Spanish background?
-Yes
-No
Which term best describes your race or background?

-American Indian
-Asian American
A-3
-Black
-White
-Other
Morning Consult questionnaire introductory section
What is your age?

-Dropdown menu ranging from “Under 18” to “100 or over”
What is your gender?

-Male
-Female
What is the last grade or class you completed in school?

-None, or grade 1–8
-High school incomplete (grades 9–11)
-High school diploma or equivalent, no further schooling
-Technical or vocational school after high school
-Some college, no degree
-Associate’s or two-year college degree
-Four-year college degree
-Graduate or professional school after college, no degree
-Graduate or professional degree
Are you yourself of Hispanic origin or descent, such as Mexican, Puerto Rican, Cuban, or
or some other Spanish background?
-Yes
-No
Which term best describes your race or background?

-American Indian
-Asian American
-Black
-White
A-4
-Other
Generally speaking, do you think of yourself as...

-Republican
-Democrat
-Independent
-Something else
[If “Democrat” is selected] Would you call yourself...

-A strong Democrat?
-A not very strong Democrat?
[If “Republican” is selected] Would you call yourself...

-A strong Republican?
-A not very strong Republican?
[If “Independent” or “Something else” is selected] Do you think of yourself as closer to...
-the Democratic Party
-the Republican Party
[Omitted Morning Consult questions not present in the MT questionnaire]
In which state do you currently reside? [Drop down menu]
If the 2016 presidential election were held today and the candidates were Democrat Hillary
Clinton and Republican Donald Trump, for whom would you vote?
-Hillary Clinton
-Donald Trump
-Don’t know/No opinion
A-5
What is your annual household income?
-Under 20 thousand dollars
-100 thousand or more dollars
[If “100 thousand or more dollars” is selected] Is that...

-250 thousand or more
Thinking about politics these days, how would you describe your political viewpoint?
-Very liberal
-Liberal
-Slightly liberal
-Moderate
-Slightly conservative
-Conservative
-Very conservative
-Don’t know
[Consent text]
Do you consent to participate in the interview?
-Yes
-No
Experimental component begins here. The Morning Consult and Mechanical Turk ques-
A-6
tionnaires are identical from this point forward.
We would like to know how much you worry about various problems facing the country.
How much do you personally worry about crime and violence?

-A great deal
-A fair amount
-Only a little
-Not at all
How much do you personally worry about the availability and affordability of health care?
-A great deal
-A fair amount
-Only a little
-Not at all
We would like to know how serious you think various problems facing the country are.
Overall, how would you describe the problem of crime in the United States — is it ex-
tremely serious, very serious, moderately serious, not too serious, or not serious at all?
-Extremely serious
-Very serious
-Moderately serious
-Not too serious
-Not serious at all
Overall, how would you describe the energy situation in the United States — is it extremely
serious, very serious, moderately serious, not too serious, or not serious at all?
-Extremely serious
-Very serious
-Moderately serious
-Not too serious
-Not serious at all
A-7
Overall, how would you describe the problem of obesity in the United States — is it ex-
-Extremely serious
-Very serious
-Moderately serious
-Not too serious
-Not serious at all
We would like to know how important various political issues are to you personally.
How important is the issue of crime to you personally?

-Extremely important
-Very important
-Moderately important
-Slightly important
-Not at all important
How important is the issue of health care to you personally?

-Very important
-Slightly important
How important is the issue of global warming/climate change to you personally?

-Very important
-Slightly important
These next few questions ask about politics in Washington, D.C. and are intended to help
us learn what types of information are known to the public. It is important to us that you
do NOT use outside sources like the Internet to search for the correct answer. Will you
A-8
answer the following questions without help from outside sources?
-Yes
-No
For how many years is a United States Senator elected - that is, how many years are there
in one full term of office for a U.S. Senator?
-Two years
-Four years
-Six years
-Eight years
-None of these
-Don’t know
How many times can an individual be elected President of the United States under current
laws?
-Once
-Twice
-Three times
-Unlimited number of terms
-Don’t know
How many U.S. senators are there from each state?

-One
-Two
-Four
-Depends on which state
-Don’t know
Who is the Prime Minister of the United Kingdom?

-Nicola Sturgeon
-Jeremy Corbyn
-Nick Clegg
-Theresa May
-Don’t know
A-9
For how many years is a member of the United States House of Representatives elected -
that is, how many years are there in one full term for a U.S. House member?
-One year
-Two years
-Four years
-Six years
-Eight years
-Don’t know
Please read the article on the next page carefully. (You’ll need to answer some questions
about it later in the survey.)
[Article condition 1: control]

What do you need for birdwatching?
-The most basic equipment required for bird watching is your eyes, though you will
soon need to have more items with you if you intend to make this a pastime or serious
hobby. How far you go is a matter of taste and budget.
-The most useful thing that you can carry is a notepad and pencil. Use this to make
a note of location, time, date, weather and habitat. Do a list of the birds that you see and
know. Do a drawing or write down a description of those that you don’t. You can look
them up later in you field guide. Your notebook should become a diary of where you have
been and what you have seen.
-A field guide is a book that provides descriptions of birds to assist you in their identi-
fication. The descriptions use several factors to help you determine the exact bird that you
are looking at. As soon as you see a bird that you do not recognize you will need to have
access to a good field guide. There are many to choose from.
-Binoculars. These are pretty essential and buy the best that you can afford. A good
pair well looked after will last you a lifetime. Take time to choose ones that suit you.
[Article condition 2: Trump rising crime message]

Trump, Accepting GOP Nomination, Promises End to Violent Crime “Epidemic”
Closing the Republican convention in Cleveland today, businessman Donald Trump ac-
A-10
cepted his party’s nomination for president in a speech that made crime a central issue of
his campaign. Trump described himself as the “law and order” candidate and said he was
committed to lowering crime rates. This is a “moment of crisis” for the country, Trump
said, which threatens “our very way of life.”
The Republican nominee, who previously stated that “crime is rising” in the United
States, painted a picture of an “epidemic” of violent crime that he blamed on the Obama
administration’s “rollback of criminal enforcement.”
Trump’s comments capped off the concluding day of the 2016 Republican National
Convention, which took place from July 18-21. By concluding before the Summer Olympics
commence in Rio, both parties’ conventions came earlier in the political season than in pre-
vious years. In 2008 and 2012, both parties staged their conventions following the Games,
which were held in Beijing and London, respectively.
The speech was attended by approximately 2,470 delegates from all 50 states, the Dis-
trict of Columbia and five territories. Each state’s delegate allotment is set by national
party rules and includes at-large delegates, congressional district delegates, and national
party representatives. Apart from the states, the District of Columbia and the five territo-
ries are awarded a specified number of at-large delegates.
[Article condition 3: Trump rising crime message + fact-check]

According to FBI’s Bureau of Justice Statistics, the violent crime rate has fallen dra-
matically and consistently over time. According to their estimates, the homicide rate in
the U.S. in 2015 was half that recorded in 1991.
A-11
vious years. In 2008 and 2012, both parties staged their conventions following the Games,
which were held in Beijing and London, respectively.
The speech was attended by approximately 2,470 delegates from all 50 states, the Dis-
trict of Columbia and five territories.
[Article condition 4: Trump rising crime message + fact-check + rejection of fact-check]

In response, Trump’s campaign manager, Paul Manafort, said the FBI’s statistics should
be viewed skeptically: "People don’t feel safe in their neighborhoods. I’m not sure what
statistics you’re talking about."
vious years. In 2008 and 2012, both parties staged their conventions following the Games.
[Article condition 5: Trump rising crime message + fact-check + rejection of fact-check

+ conspiracy/fact-check source derogation]
A-12
In response, Trump’s campaign manager, Paul Manafort, said the FBI’s statistics should
be viewed skeptically: "People don’t feel safe in their neighborhoods. I’m not sure what
statistics you’re talking about."
Manafort suggested that the FBI might even have an interest in the election’s outcome:
"The FBI is certainly suspect these days after what they just did with Hillary Clinton,"
he added. Manafort was referring to the FBI’s recent decision not to prosecute Hillary
Clinton over the email scandal that has rocked her presidential campaign.
[If assigned to control condition]

What activity was the focus of the article you just read?
-Bird watching
-Hiking
-Boating
-Swimming
[If assigned to control condition]

What item is most useful to carry for that activity according to the article?
-Telescope
-Notepad and pencil
-Hat
-Pencil
[If assigned to non-control condition]

What event was the focus of the article you just read?
-The Democratic convention
-The Republican convention
-A presidential debate
-A vice presidential debate
A-13
[If assigned to non-control condition]
What issue did the article you just read focus on?
-Crime
-Terrorism
-Economy
-Health care
-Environment
We now would like to ask you some factual political questions. Would you say that, com-
pared to ten years ago, the violent crime rate has gone up, stayed about the same, or gone
down?
-Gone up
-Stayed about the same
-Gone down
[If “Gone up” is selected]

Compared to ten years ago, has the violent crime rate gone up a lot or only somewhat?
-Gone up a lot
-Gone up somewhat
[If “Gone up” is selected]

What would you say is the main reason violent crime has gone up?
-The decline in morals and values
-Increased economic inequality and helplessness
-Some other reason
-Don’t know
[If “Gone down” is selected]

Compared to ten years ago, has the violent crime rate gone down a lot or only somewhat?
-Gone down a lot
-Gone down somewhat
[If “Gone down” is selected]
A-14
What would you say is the main reason violent crime has gone down?
-Tougher policing and longer prison sentences
-The improved economy and reduced unemployment
-Some other reason
-Don’t know
In your view, how accurate are federal crime statistics?

-Extremely accurate
-Somewhat accurate
-Not very accurate
-Not at all accurate
In your view, how accurate was the article you read in reporting federal crime statistics?
-Extremely accurate
-Somewhat accurate
-Not very accurate
-Not at all accurate
In your opinion, how fair was the article you read?

-Extremely fair
-Somewhat fair
-Not very fair
-Not at all fair
Would you say the article you read was biased or not biased?
-Yes, biased
-No, not biased
[If “Yes, biased” is selected]

Would you say the article had a liberal bias, a conservative bias, or some other bias?
-Liberal
-Conservative
-Other
A-15
[If “Yes, biased” is selected]
Would you say the article was extremely biased, somewhat biased, or slightly biased?
-Extremely biased
-Somewhat biased
-Slightly biased
In general, do you think the courts in this area deal too harshly or not harshly enough with
criminals?
-Too harshly
-About right
-Not harshly enough
Next we are going to give you a list of government programs. For each one, please indicate
whether you would like to see spending increased or decreased.
Should government spending on law enforcement be increased, decreased, or kept the

same?
-Increased a great deal
-Increased a moderate amount
-Increased a little
-Kept the same
-Decreased a little
-Decreased a moderate amount
-Decreased a great deal
Should government spending on Social Security be increased, decreased, or kept the same?
-Increased a little
-Kept the same
-Decreased a little
A-16
Should government spending on protecting the environment be increased, decreased, or
kept the same?
-Increased a little
-Kept the same
-Decreased a little
Overall, how would you describe the problem of crime in the United States — is it ex-
-Extremely serious
-Very serious
-Moderately serious
-Not too serious
-Not serious at all
Do you approve or disapprove of how Barack Obama is handling his job as president?
-Strongly approve
-Somewhat approve
-Neither approve nor disapprove
-Somewhat disapprove
-Strongly disapprove
Do you feel favorable or unfavorable toward each of these individuals?
Very Somewhat Neither favorable Somewhat Very

favorable favorable nor unfavorable unfavorable unfavorable
Barack Obama
Hillary Clinton
Bernie Sanders
Donald Trump
Ted Cruz
John Kasich
George W. Bush
A-17
Study 2 instrument
[Wave 1 Party ID]
Please describe your political party affiliation.
-Strong Republican
-Not Strong Republican
-Lean Republican
-Undecided, Independent, Other
-Lean Democrat
-Not Strong Democrat
-Strong Democrat
[Wave 2 pre-treatment vote choice]

Which candidate do you plan for in the upcoming presidential election?
-Hillary Clinton
-Donald Trump
-Somebody else
[Wave 2 pre-treatment debate winner]

Who won the presidential debate?
-Clearly Clinton
-Mostly Clinton
-Neither Clinton nor Trump
-Mostly Trump
-Clearly Trump
[Wave 2 pre-treatment debate evaluations]

Use the sliders below to tell us how you felt about the candidates’ performances during the
debate. A score of 0 means you felt very cold and negative about that candidate’s debate
performance, and a score of 100 means you felt very warm and positive about that candi-
date’s debate performance.
-Hillary Clinton’s debate performance

-Donald Trump’s debate performance
A-18
[Wave 2 Trump statement shown to all respondents]
Our jobs are fleeing the country to Mexico... they’re building some of the biggest, most
sophisticated plants. Not so much in America. Thousands of jobs leaving Michigan,
Ohio...their companies are just leaving, they’re gone.
[Wave 2 fact-check shown to those assigned to fact-check condition]

In fact, according to the Bureau of Labor Statistics, unemployment has fallen in both
Michigan and Ohio. Both states each saw 70,000 new jobs over the last year.
[Wave 2 post-treatment question shown to all respondents]

Over the last few years, has unemployment gone up or down in Michigan and Ohio?
-Gone down a lot
-Gone down a little
-Stayed the same
-Gone up a little
-Gone up a lot
[Wave 3 vote choice]

Which candidate do you plan for in the upcoming presidential election?
-Hillary Clinton
-Donald Trump
-Somebody else
[Wave 3 debate evaluations]

Use the sliders below to tell us how you felt about the candidates’ performances during
the first debate. A score of 0 means you felt very cold and negative about that candidate’s
debate performance, and a score of 100 means you felt very warm and positive about that
candidate’s debate performance.
-Hillary Clinton’s debate performance

-Donald Trump’s debate performance
[Wave 3 debate winner]

Who won the first presidential debate?
A-19
-Hillary Clinton
-Donald Trump
A-20
Online Appendix B: Preregistered research questions
In addition to the preregistered hypotheses we consider in main body of the paper, we also
explore preregistered three research questions (RQs) that address important substantive
and methodological issues where we do not have strong theoretical priors. In this online
appendix, we describe these preregistered RQs and briefly report their results.
The first research question (RQ1) examines whether responsiveness to treatments dif-
fers between a more conventional online survey sample from the polling company Morn-
ing Consult and an Amazon Mechanical Turk sample (Study 1 only). Though studies have
found that Mechanical Turk workers provide valid data (e.g., Berinsky, Huber, and Lenz
2012; Levay, Freese, and Druckman 2016), concerns continue about social desirability
bias when researchers pay and evaluate respondents (e.g., Clifford and Jerit 2015). Re-
spondents may be inclined to endorse the conclusions of fact-checks to demonstrate that
they read the stimuli and/or were paying attention. We therefore consider whether the re-
sults of Study 1 differ between participants surveyed by Morning Consult, which draws its
respondents from SSI and other similar panels, and workers on Mechanical Turk.
The second research question (RQ2) tests whether any of our treatments alter policy
preferences or the perceived importance of the issue in question (Study 1 only). Decreasing
the perceived importance of an issue is one way respondents might cope with uncongenial
information. Previous studies have also found mixed results of factual information on issue
opinions (e.g., Flynn 2016).
The third research questions (RQ3) evaluates whether fact-checks about crime affect
evaluations of the incumbent president (Obama), the candidate whose statement is ques-
tioned (Trump), or the candidate who is invoked in an elite attack on the fact-check (Clin-
ton)?
We first consider RQ1. As Tables 1 and 2 demonstrate, the marginal effects of expo-
sure to journalistic fact-checks were strikingly parallel between the Morning Consult and
Mechanical Turk samples (RQ1). Participants on Mechanical Turk are not differentially
responsive to fact-checks, a finding which is consistent with the results in Mummolo and
Peterson (N.d.).1
Turning to RQ2, we estimate the marginal effect of fact-check exposure on respon-
1 Online Appendix C presents fully pooled results with interaction terms testing for differences in effects
between the Morning Consult and Mechanical Turk samples. Out of 48 coefficients that interact an exper-
imental manipulation with an indicator for a Mechanical Turk respondent, only eight are significant at the
p < .05 level. In particular, only one of 24 coefficients testing for differential responsiveness among Trump
supporters is significant at the p < .05 level.
B-1
dents’ views of the importance of crime and their preferences on the issue for both can-
didates’ supporters. As shown in Table C15 in Online Appendix C, we find no consistent
effect across multiple measures and samples.
Finally, we evaluate RQ3 by estimating the marginal effect of assignment to a fact-
check on respondents’ feelings towards the sitting president (Obama), and the two major-
party nominees (Trump and Clinton). These results are presented in Tables C17 and C19.
There is scant evidence that the fact-checks affected how respondents answered the feeling
thermometer questions.
References
Berinsky, Adam J., Gregory A. Huber, and Gabriel S. Lenz. 2012. “Evaluating online
labor markets for experimental research: Amazon. com’s Mechanical Turk.” Political
Analysis 20 (3): 351–368.
Clifford, Scott, and Jennifer Jerit. 2015. “Do attempts to improve respondent attention
increase social desirability bias?” Public Opinion Quarterly 79 (3): 790–802.
Flynn, D.J. 2016. “The Scope and Correlates of Political Misperceptions in the Mass
Public.” Unpublished paper, Dartmouth College.
Levay, Kevin E., Jeremy Freese, and James N. Druckman. 2016. “The Demo-
graphic and Political Composition of Mechanical Turk Samples.” SAGE Open 6 (1):
2158244016636433.
Mummolo, Jonathan, and Erik Peterson. N.d. “Demand Effects in Survey Experi-
ments: An Empirical Assessment.” Unpublished manuscript. Downloaded June 16,
2017 from https://www.dropbox.com/s/lga66eon2o5npw6/demand_
effects_mummolo_peterson_4_20_2017.pdf?dl=0.
B-2
Online Appendix C: Supplementary tables
Study 1 supplementary results
Table C1: Study 1 demographics/balance (Morning Consult)
Control Uncorrected Fact-check Denial Denial/derogation Total

Age
18-34 29% 25% 23% 27% 27% 26%
35-44 17% 22% 16% 15% 16% 17%
45-64 34% 31% 47% 41% 34% 38%
65+ 21% 22% 14% 17% 23% 19%
Gender
Female 53% 59% 52% 57% 56% 55%
Male 47% 41% 48% 43% 44% 45%
Education
High school or less 19% 17% 15% 11% 15% 15%
Some college/associate 35% 36% 42% 40% 42% 39%
Bachelor’s degree 25% 29% 24% 25% 24% 25%
Graduate degree 21% 18% 19% 25% 19% 20%
Race
Nonwhite 14% 16% 16% 18% 16% 16%
White 86% 84% 84% 82% 84% 84%
Party
Democrat 38% 39% 35% 42% 42% 39%
Republican 31% 29% 30% 24% 28% 28%
Independent/something else 31% 32% 34% 34% 30% 32%
C-1
Table C2: Study 1 demographics/balance (Mechanical Turk)
Control Uncorrected Fact-check Denial Denial/derogation Total

Age
18-34 59% 55% 58% 57% 56% 57%
35-44 21% 23% 23% 22% 24% 23%
45-64 18% 20% 16% 17% 17% 18%
65+ 2% 2% 3% 4% 2% 2%
Gender
Female 50% 55% 53% 54% 49% 52%
Male 50% 45% 47% 46% 51% 48%
Education
High school or less 8% 9% 8% 8% 8% 8%
Some college/associate degree 39% 39% 38% 42% 40% 40%
Bachelor’s degree 37% 32% 37% 35% 35% 35%
Graduate degree 16% 19% 17% 16% 16% 17%
Race
Nonwhite 20% 18% 20% 18% 21% 20%
White 80% 82% 80% 82% 79% 80%
Party
Democrat 44% 43% 43% 41% 41% 42%
Republican 19% 22% 22% 22% 21% 21%
Independent/something else 37% 36% 35% 37% 38% 37%
C-2
Table C3: Study 2 demographics/balance (Mechanical Turk)
Wave 1 Wave 2 Wave 3

Control Treatment Total Control Treatment Total
Age
18-34 56% 52% 53% 52% 50% 55% 52%
35-44 22% 25% 24% 24% 27% 22% 25%
45-64 19% 20% 22% 21% 19% 21% 20%
65+ 3% 3% 2% 3% 3% 2% 3%
Gender
Male 50% 54% 51% 52% 57% 54% 55%
Female 50% 46% 49% 48% 43% 46% 45%
Education
High school or less 10% 9% 11% 10% 10% 10% 10%
Some college 36% 37% 32% 35% 33% 28% 31%
Bachelor’s degree 41% 40% 42% 41% 42% 46% 44%
Graduate degree 14% 14% 15% 14% 15% 16% 16%
Race
Nonwhite 21% 20% 22% 21% 19% 22% 21%
White 79% 80% 78% 79% 81% 78% 79%
Party
Democrat 50% 47% 51% 49% 54% 58% 56%
Republican 25% 29% 27% 28% 24% 24% 24%
Independent 25% 24% 22% 23% 22% 18% 20%
C-3
Table C4: Message exposure effects on beliefs about changes in crime (ordered probit)

(0.14) (0.09)
Uncorrected statement -0.24* -0.03
(0.13) (0.07)
(0.20) (0.13)
Fact-check -0.70*** -0.86***
(0.14) (0.09)
(0.21) (0.16)
(0.15) (0.08)
(0.22) (0.14)
Denial/source derogation -0.82*** -0.65***
(0.15) (0.08)
Denial/source derogation ⇥ Trump support 0.41* 0.25*
(0.22) (0.14)
(0.14) (0.09)
(0.15) (0.13)
(0.15) (0.08)
(0.16) (0.12)
(0.14) (0.08)
Trump supporters -0.42** -0.46***
(0.16) (0.11)
N 990 2430
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided); ordered probit models with robust standard errors (cut-
points omitted). Respondents are Morning Consult respondents and Amazon Mechanical Turk workers who
supported Hillary Clinton or Donald Trump in the 2016 general election (reference category for Trump
support is Clinton support).
C-4
Table C5: Message exposure effects on beliefs about changes in crime (all respondents)

Trump support 0.08 0.38***
(0.20) (0.14)
Clinton support -0.43** -0.44***
(0.21) (0.12)
(0.25) (0.16)
(0.29) (0.20)
Uncorrected ⇥ Clinton support -0.24 0.04
(0.30) (0.18)
Fact-check -1.07*** -0.80***
(0.28) (0.18)
Fact-check ⇥ Trump support 0.22 -0.06
(0.33) (0.24)
Fact-check ⇥ Clinton support 0.13 -0.17
(0.33) (0.21)
(0.29) (0.16)
(0.34) (0.22)
Fact-check denial ⇥ Clinton support 0.01 0.16
(0.35) (0.19)
Denial/source derogation -0.62** -0.82***
(0.26) (0.17)
Denial/source derogation ⇥ Trump support 0.11 0.34
(0.32) (0.22)
Denial/source derogation ⇥ Clinton support -0.50 0.05
(0.32) (0.20)
Constant 4.07*** 3.44***
(0.17) (0.11)
(0.19) (0.10)
(0.18) (0.15)
(0.19) (0.09)
(0.19) (0.14)
(0.19) (0.10)
(0.20) (0.14)
R2 0.13 0.18
N 1200 2983
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided); OLS models with robust standard errors. Respondents
are Morning Consult respondents and Amazon Mechanical Turk workers.
C-5
Table C6: Message exposure effects on perceptions of accuracy and fairness

MC MT MC MT MC MT
Trump supporter -0.15 -0.00 -0.07 -0.04 -0.18** -0.08
(0.13) (0.09) (0.09) (0.07) (0.08) (0.05)
Uncorrected statement -0.17 0.05 -0.37*** -0.38*** -0.08 0.02
(0.12) (0.07) (0.10) (0.06) (0.08) (0.04)
Uncorrected ⇥ Trump support 0.48*** 0.50*** 0.41*** 0.49*** 0.20 -0.06
(0.16) (0.12) (0.13) (0.10) (0.12) (0.08)
Fact-check 0.23** 0.71*** -0.10 0.19*** -0.04 0.10**
(0.11) (0.06) (0.09) (0.06) (0.08) (0.04)
Fact-check ⇥ Trump support -0.08 -0.36*** -0.09 -0.37*** -0.18 -0.24***
(0.16) (0.12) (0.14) (0.10) (0.13) (0.08)
Fact-check denial 0.26** 0.74*** -0.09 0.10* 0.04 0.15***
(0.11) (0.06) (0.10) (0.06) (0.08) (0.04)
Denial ⇥ Trump support -0.13 -0.25** -0.24 -0.24** -0.38*** -0.40***
(0.17) (0.12) (0.15) (0.10) (0.14) (0.08)
Denial/source derogation 0.30*** 0.74*** -0.12 0.04 0.06 0.13***
(0.11) (0.07) (0.09) (0.06) (0.08) (0.04)
Denial/source derogation ⇥ Trump support -0.15 -0.28** -0.06 -0.14 -0.29** -0.33***
(0.16) (0.12) (0.14) (0.10) (0.13) (0.08)
Constant 2.68*** 2.25*** 3.06*** 2.99*** 3.05*** 3.00***
(0.09) (0.05) (0.06) (0.04) (0.06) (0.02)
(0.11) (0.05) (0.11) (0.06) (0.09) (0.04)
(0.10) (0.08) (0.09) (0.07) (0.10) (0.07)
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
(0.11) (0.08) (0.11) (0.07) (0.11) (0.07)
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
(0.10) (0.07) (0.10) (0.07) (0.10) (0.07)
R2 0.04 0.14 0.03 0.06 0.08 0.07
N 987 2426 985 2426 987 2429
are Morning Consult respondents (MC) and Amazon Mechanical Turk workers (MT) who supported Hillary
Clinton or Donald Trump in the 2016 election (reference category for Trump support is thus Clinton support).
C-6
Table C7: Message exposure effects on perceptions of accuracy/fairness (ordered probit)

MC MT MC MT MC MT
Trump supporter -0.24 0.01 -0.11 -0.05 -0.32** -0.17
(0.17) (0.13) (0.14) (0.11) (0.15) (0.11)
Uncorrected statement -0.26 0.06 -0.54*** -0.58*** -0.16 0.04
(0.17) (0.09) (0.15) (0.09) (0.15) (0.08)
Uncorrected ⇥ Trump support 0.67*** 0.67*** 0.59*** 0.73*** 0.35 -0.11
(0.23) (0.16) (0.20) (0.15) (0.22) (0.16)
Fact-check 0.30* 0.98*** -0.14 0.30*** -0.07 0.22***
(0.16) (0.09) (0.14) (0.09) (0.15) (0.08)
Fact-check ⇥ Trump support -0.11 -0.54*** -0.15 -0.59*** -0.28 -0.50***
(0.23) (0.16) (0.21) (0.15) (0.23) (0.16)
Fact-check denial 0.34** 1.07*** -0.10 0.15* 0.07 0.34***
(0.16) (0.09) (0.15) (0.09) (0.15) (0.08)
Denial ⇥ Trump support -0.16 -0.42** -0.37 -0.39** -0.61*** -0.83***
(0.24) (0.17) (0.23) (0.15) (0.24) (0.16)
Denial/source derogation 0.40** 1.06*** -0.18 0.06 0.12 0.29***
(0.16) (0.10) (0.14) (0.09) (0.14) (0.09)
Denial/source derogation ⇥ Trump support -0.22 -0.48*** -0.10 -0.24 -0.50** -0.68***
(0.23) (0.16) (0.21) (0.16) (0.22) (0.16)
(0.15) (0.08) (0.16) (0.09) (0.16) (0.09)
Trump supporters -0.22 -0.29** -0.34** -0.45*** -0.54*** -0.22
(0.14) (0.12) (0.14) (0.11) (0.17) (0.15)
Clinton supporters 0.60*** 1.01*** 0.43*** 0.72*** 0.23 0.30***
(0.14) (0.09) (0.16) (0.08) (0.15) (0.09)
Trump supporters -0.23 -0.07 -0.52*** -0.40*** -0.73*** -0.42***
(0.16) (0.12) (0.16) (0.11) (0.18) (0.14)
(0.15) (0.09) (0.16) (0.09) (0.15) (0.10)
(0.15) (0.11) (0.15) (0.11) (0.17) (0.14)
N 987 2426 985 2426 987 2429
points omitted). Respondents are Morning Consult respondents (MC) and Amazon Mechanical Turk work-
ers (MT) who supported Hillary Clinton or Donald Trump in the 2016 election (reference category for Trump
support is thus Clinton support).
C-7
Table C8: Message effects on perceptions of accuracy and fairness (all respondents)

MC MT MC MT MC MT
Trump support 0.29* -0.11 0.11 0.05 0.11 0.03
(0.16) (0.12) (0.11) (0.09) (0.10) (0.07)
Clinton support 0.44*** -0.11 0.18* 0.09 0.29*** 0.11**
(0.17) (0.10) (0.10) (0.07) (0.09) (0.05)
Uncorrected statement 0.42** 0.20* -0.17 -0.14 0.08 -0.06
(0.18) (0.11) (0.12) (0.08) (0.12) (0.07)
Uncorrected ⇥ Trump support -0.11 0.35** 0.21 0.25** 0.03 0.02
(0.21) (0.14) (0.15) (0.12) (0.15) (0.10)
Uncorrected ⇥ Clinton support -0.59*** -0.15 -0.20 -0.24** -0.16 0.08
(0.22) (0.13) (0.15) (0.10) (0.15) (0.08)
Fact-check 0.43** 0.50*** -0.16 0.01 0.07 0.07
(0.18) (0.10) (0.16) (0.09) (0.13) (0.08)
Fact-check ⇥ Trump support -0.28 -0.15 -0.03 -0.19 -0.29* -0.22**
(0.21) (0.15) (0.19) (0.12) (0.16) (0.11)
Fact-check ⇥ Clinton support -0.20 0.21* 0.06 0.17 -0.12 0.03
(0.21) (0.12) (0.18) (0.11) (0.15) (0.09)
Fact-check denial 0.58*** 0.42*** 0.10 -0.10 0.00 -0.00
(0.18) (0.11) (0.13) (0.08) (0.12) (0.07)
Fact-check denial ⇥ Trump support -0.45** 0.07 -0.42** -0.04 -0.35** -0.25**
(0.22) (0.15) (0.18) (0.12) (0.16) (0.10)
Fact-check denial ⇥ Clinton support -0.32 0.33*** -0.18 0.20** 0.03 0.16*
(0.21) (0.12) (0.16) (0.10) (0.15) (0.08)
Denial/source derogation 0.41** 0.51*** -0.33** -0.04 -0.11 -0.05
(0.17) (0.10) (0.14) (0.09) (0.12) (0.07)
Denial/source derogation ⇥ Trump support -0.26 -0.06 0.15 -0.07 -0.12 -0.15
(0.21) (0.14) (0.18) (0.12) (0.16) (0.09)
Denial/source derogation ⇥ Clinton support -0.11 0.22* 0.20 0.08 0.17 0.18**
(0.21) (0.12) (0.17) (0.11) (0.14) (0.08)
Constant 2.24*** 2.36*** 2.88*** 2.90*** 2.76*** 2.89***
(0.14) (0.09) (0.09) (0.06) (0.07) (0.05)
(0.11) (0.05) (0.11) (0.06) (0.09) (0.04)
(0.10) (0.08) (0.09) (0.07) (0.10) (0.07)
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
(0.11) (0.08) (0.11) (0.07) (0.11) (0.07)
(0.10) (0.06) (0.11) (0.06) (0.08) (0.04)
(0.10) (0.07) (0.10) (0.07) (0.10) (0.07)
R2 0.05 0.13 0.04 0.05 0.08 0.06
N 1196 2978 1193 2977 1196 2980
are Morning Consult respondents (MC) and Amazon Mechanical Turk workers (MT).
C-8
Table C9: Message exposure effects on perceptions of article bias

Trump supporter 0.29** 0.19***
(0.15) (0.07)
Uncorrected statement -0.81*** -0.77***
(0.18) (0.08)
Uncorrected ⇥ Trump support 0.36 0.35***
(0.25) (0.12)
Fact-check -0.46*** 0.02
(0.16) (0.07)
Fact-check ⇥ Trump support 0.53** 0.56***
(0.25) (0.14)
Fact-check denial -0.20 -0.13*
(0.17) (0.07)
Denial ⇥ Trump support 0.81*** 0.39***
(0.27) (0.13)
(0.17) (0.08)
Derogation ⇥ Trump support 0.38 0.28*
(0.26) (0.15)
Constant -3.95*** -4.06***
(0.10) (0.04)
Clinton supporters 0.35* 0.78***
(0.20) (0.09)
Trump supporters 0.53*** 0.99***
(0.20) (0.13)
Clinton supporters 0.62*** 0.63***
(0.20) (0.09)
(0.22) (0.12)
Clinton supporters 0.47** 0.48***
(0.21) (0.10)
Trump supporters 0.49** 0.41***
(0.21) (0.14)
R2 0.10 0.11
N 978 2426
supported Hillary Clinton or Donald Trump in the C-9
2016 election (reference category for Trump support is
thus Clinton support). Outcome variable ranges from “Extremely biased conservative” (-3) to “Extremely
biased liberal” (3).
Table C10: Message exposure effects on perceptions of article bias (ordered probit)

Trump supporter 0.25** 0.22***
(0.12) (0.08)
Uncorrected statement -0.66*** -0.77***
(0.15) (0.08)
Uncorrected ⇥ Trump support 0.26 0.31**
(0.20) (0.13)
Fact-check -0.40*** 0.03
(0.14) (0.07)
Fact-check ⇥ Trump support 0.43** 0.52***
(0.20) (0.14)
Fact-check denial -0.19 -0.12*
(0.14) (0.07)
Denial ⇥ Trump support 0.62*** 0.38***
(0.21) (0.14)
(0.14) (0.08)
Derogation ⇥ Trump support 0.28 0.27*
(0.21) (0.16)
Clinton supporters 0.27* 0.80***
(0.16) (0.09)
(0.16) (0.13)
Clinton supporters 0.48*** 0.65***
(0.16) (0.09)
(0.17) (0.13)
Clinton supporters 0.37** 0.49***
(0.16) (0.10)
Trump supporters 0.39** 0.45***
(0.17) (0.14)
N 978 2426
points omitted). Respondents supported Hillary Clinton or Donald Trump in the 2016 election (reference
category for Trump support is thus Clinton support). Outcome variable ranges from “Extremely biased
conservative” (-3) to “Extremely biased liberal” (3).
C-10
Table C11: Message exposure effects on factual interpretations

Trump supporter 0.71*** 0.43***
(0.10) (0.07)
Uncorrected statement 0.00 0.01
(0.10) (0.05)
Uncorrected ⇥ Trump support -0.15 0.01
(0.15) (0.10)
Fact-check 0.07 -0.04
(0.10) (0.05)
Fact-check ⇥ Trump support -0.31** 0.00
(0.14) (0.10)
Fact-check denial 0.27*** -0.00
(0.10) (0.05)
Denial ⇥ Trump support -0.36** -0.09
(0.14) (0.10)
Denial/source derogation 0.17* -0.08
(0.10) (0.05)
Derogation ⇥ Trump support -0.33** 0.05
(0.15) (0.09)
Constant -0.24*** -0.27***
(0.07) (0.03)
(0.10) (0.05)
Trump supporters -0.08* -0.05
(0.11) (0.09)
Clinton supporters 0.27*** -0.01
(0.10) (0.05)
(0.11) (0.09)
Clinton supporters 0.16* -0.08
(0.10) (0.05)
(0.12) (0.08)
R2 0.11 0.08
N 990 2427
supported Hillary Clinton or Donald Trump in the 2016 election (reference category for Trump support is
thus Clinton support). Outcome variable is coded -1 for liberal interpretations of reported changes in factual
beliefs, 1 for conservative interpretations, and 0 for other interpretations/don’t know or for respondents who
said crime had not changed.
C-11
Table C12: Message exposure effects on factual interpretations (ordered probit)

Trump supporter 1.11*** 0.71***
(0.18) (0.11)
Uncorrected statement 0.01 0.01
(0.15) (0.08)
(0.25) (0.16)
Fact-check 0.11 -0.07
(0.15) (0.08)
Fact-check ⇥ Trump support -0.52** 0.00
(0.23) (0.16)
Fact-check denial 0.40*** -0.00
(0.15) (0.08)
Denial ⇥ Trump support -0.58** -0.15
(0.23) (0.16)
Denial/source derogation 0.25* -0.13
(0.15) (0.09)
Derogation ⇥ Trump support -0.54** 0.10
(0.24) (0.16)
(0.15) (0.09)
(0.17) (0.14)
Clinton supporters 0.40*** -0.01
(0.15) (0.08)
(0.18) (0.14)
Clinton supporters 0.25* -0.15
(0.15) (0.09)
(0.19) (0.14)
N 990 2427
points omitted). Respondents supported Hillary Clinton or Donald Trump in the 2016 election (reference
category for Trump support is thus Clinton support). Outcome variable is coded -1 for liberal interpretations
of reported changes in factual beliefs, 1 for conservative interpretations, and 0 for other interpretations/don’t
know or for respondents who said crime had not changed.
C-12
Table C13: Testing for Morning Consult/Mechanical Turk differences
Crime change Article accurate Article fair Statistics accurate Article bias Crime interpretation
Trump support 0.52*** -0.15 -0.07 -0.18** -0.29** 0.71***
(0.16) (0.13) (0.09) (0.08) (0.15) (0.10)
Mechanical Turk -0.63*** -0.43*** -0.07 -0.05 0.11 -0.03
(0.14) (0.10) (0.07) (0.06) (0.11) (0.08)
Trump support ⇥ Turk 0.30 0.15 0.04 0.10 0.10 -0.27**
(0.20) (0.16) (0.12) (0.10) (0.17) (0.12)
Uncorrected statement -0.34* -0.17 -0.37*** -0.08 0.81*** 0.00
(0.18) (0.12) (0.10) (0.08) (0.18) (0.10)
Uncorrected ⇥ Turk 0.28 0.22 -0.01 0.10 -0.05 0.01
(0.20) (0.14) (0.11) (0.09) (0.20) (0.11)
Uncorrected ⇥ Trump support 0.35 0.48*** 0.41*** 0.20 -0.36 -0.15
(0.23) (0.16) (0.13) (0.12) (0.25) (0.15)
Uncorrected ⇥ Trump support ⇥ Turk -0.19 0.02 0.08 -0.25* 0.00 0.16
(0.28) (0.20) (0.16) (0.14) (0.28) (0.18)
Fact-check -0.94*** 0.23** -0.10 -0.04 0.46*** 0.07
(0.19) (0.11) (0.09) (0.08) (0.16) (0.10)
Fact-check ⇥ Turk -0.03 0.48*** 0.29*** 0.14 -0.48*** -0.11
(0.21) (0.13) (0.11) (0.09) (0.18) (0.11)
Fact-check ⇥ Trump support 0.09 -0.08 -0.09 -0.18 -0.53** -0.31**
(0.26) (0.16) (0.14) (0.13) (0.25) (0.14)
Fact-check ⇥ Trump support ⇥ Turk 0.02 -0.28 -0.28 -0.07 -0.02 0.31*
(0.32) (0.20) (0.17) (0.15) (0.28) (0.17)
C-13
Fact-check denial -0.74*** 0.26** -0.09 0.04 0.20 0.27***
(0.19) (0.11) (0.10) (0.08) (0.17) (0.10)
Fact-check denial ⇥ Turk -0.19 0.49*** 0.18* 0.12 -0.07 -0.27**
(0.21) (0.13) (0.11) (0.09) (0.18) (0.11)
Fact-check denial ⇥ Trump support 0.21 -0.13 -0.24 -0.38*** -0.81*** -0.36**
(0.27) (0.17) (0.15) (0.14) (0.27) (0.14)
Fact-check denial ⇥ Trump support ⇥ Turk 0.17 -0.12 -0.00 -0.02 0.42 0.26
(0.32) (0.21) (0.18) (0.16) (0.30) (0.17)
Denial/source derogation -1.12*** 0.30*** -0.12 0.06 0.35** 0.17*
(0.19) (0.11) (0.09) (0.08) (0.17) (0.10)
Denial/source derogation ⇥ Turk 0.34 0.43*** 0.16 0.06 -0.06 -0.24**
(0.21) (0.13) (0.11) (0.09) (0.19) (0.11)
Denial/source derogation ⇥ Trump support 0.61** -0.15 -0.06 -0.29** -0.38 -0.33**
(0.27) (0.16) (0.14) (0.13) (0.26) (0.15)
Denial/source derogation ⇥ Trump support ⇥ Turk -0.31 -0.13 -0.09 -0.03 0.10 0.38**
(0.32) (0.20) (0.17) (0.15) (0.30) (0.18)
Constant 3.64*** 2.68*** 3.06*** 3.05*** 3.95*** -0.24***
(0.12) (0.09) (0.06) (0.06) (0.10) (0.07)
R2 0.20 0.12 0.05 0.07 0.11 0.11
N 3420 3413 3411 3416 3404 3417
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided); OLS models with robust standard errors. Respondents are Morning Consult respondents and Amazon Mechanical Turk workers who supported
Hillary Clinton or Donald Trump in the 2016 election. As such, the reference category for the Mechanical Turk indicator is Morning Consult and the reference category for Trump support is Clinton
support.
Table C14: Testing for Morning Consult/Mechanical Turk differences (ordered probit)
Crime change Article accurate Article fair Statistics accurate Article bias Crime interpretation
Trump support 0.45*** -0.26 -0.11 -0.36** -0.30** 1.17***
(0.14) (0.18) (0.14) (0.17) (0.15) (0.19)
Mechanical Turk -0.47*** -0.63*** -0.09 -0.12 0.12 -0.04
(0.11) (0.15) (0.11) (0.13) (0.11) (0.13)
Trump support ⇥ Turk 0.19 0.26 0.06 0.20 0.09 -0.49**
(0.17) (0.22) (0.18) (0.20) (0.16) (0.21)
Uncorrected statement -0.25* -0.28* -0.54*** -0.17 0.75*** 0.01
(0.14) (0.17) (0.14) (0.17) (0.17) (0.17)
Uncorrected ⇥ Turk 0.22 0.34* -0.04 0.22 -0.03 0.01
(0.16) (0.19) (0.17) (0.18) (0.18) (0.18)
Uncorrected ⇥ Trump suppo 0.26 0.69*** 0.59*** 0.39 -0.28 -0.28
(0.21) (0.23) (0.20) (0.24) (0.24) (0.26)
Uncorrected ⇥ Trump support ⇥ Turk -0.17 -0.03 0.14 -0.49* -0.02 0.29
(0.24) (0.28) (0.25) (0.29) (0.26) (0.31)
Fact-check -0.76*** 0.30* -0.14 -0.08 0.45*** 0.12
(0.15) (0.17) (0.14) (0.17) (0.16) (0.16)
Fact-check ⇥ Turk -0.08 0.67*** 0.44*** 0.28 -0.48*** -0.19
(0.17) (0.19) (0.17) (0.19) (0.17) (0.18)
Fact-check ⇥ Trump support 0.06 -0.11 -0.15 -0.30 -0.49** -0.54**
(0.22) (0.23) (0.21) (0.25) (0.23) (0.24)
C-14
Fact-check ⇥ Trump support ⇥ Turk 0.06 -0.41 -0.44* -0.18 0.01 0.55*
(0.27) (0.28) (0.26) (0.29) (0.27) (0.29)
Fact-check denial -0.59*** 0.34** -0.10 0.08 0.20 0.43***
(0.16) (0.17) (0.15) (0.16) (0.16) (0.16)
Fact-check denial ⇥ Turk -0.19 0.72*** 0.25 0.25 -0.08 -0.43**
(0.18) (0.19) (0.17) (0.18) (0.18) (0.18)
Fact-check denial ⇥ Trump support 0.14 -0.16 -0.37 -0.66*** -0.70*** -0.61**
(0.23) (0.24) (0.23) (0.26) (0.25) (0.25)
Fact-check denial ⇥ Trump support ⇥ Turk 0.17 -0.25 -0.02 -0.13 0.35 0.46
(0.27) (0.29) (0.27) (0.30) (0.28) (0.29)
Denial/source derogation -0.88*** 0.41** -0.18 0.13 0.33** 0.27*
(0.16) (0.17) (0.14) (0.16) (0.16) (0.16)
Denial/source derogation ⇥ Turk 0.26 0.64*** 0.24 0.15 -0.06 -0.40**
(0.17) (0.19) (0.17) (0.18) (0.18) (0.18)
Denial/source derogation ⇥ Trump support 0.44* -0.23 -0.10 -0.55** -0.32 -0.56**
(0.23) (0.24) (0.21) (0.24) (0.25) (0.26)
Denial/source derogation ⇥ Trump support ⇥ Turk -0.21 -0.24 -0.14 -0.10 0.07 0.66**
(0.27) (0.29) (0.27) (0.29) (0.29) (0.30)
N 3420 3413 3411 3416 3404 3417
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided); ordered probit models with robust standard errors (cutpoints omitted). Respondents are Morning Consult respondents and Amazon Mechanical
Turk workers who supported Hillary Clinton or Donald Trump in the 2016 election. As such, the reference category for the Mechanical Turk indicator is Morning Consult and the reference category
for Trump support is Clinton support.
Table C15: Message exposure effects on issue importance and policy preferences
Crime importance Treatment of criminals Law enforcement $

MC MT MC MT MC MT
Trump support 0.18 0.22** 0.29*** 0.39*** 0.26 -0.16
(0.12) (0.09) (0.10) (0.06) (0.18) (0.67)
Uncorrected statement -0.25** -0.04 -0.01 -0.02 -0.10 -0.26
(0.12) (0.07) (0.10) (0.05) (0.17) (0.34)
Uncorrected ⇥ Trump support 0.13 0.09 0.05 0.17* -0.12 0.34
(0.17) (0.12) (0.14) (0.09) (0.27) (1.02)
Fact-check -0.26** -0.13** -0.04 -0.03 -0.28 -0.01
(0.12) (0.07) (0.10) (0.05) (0.18) (0.10)
Fact-check ⇥ Trump support 0.11 0.22* 0.24* 0.17* 0.12 0.81
(0.17) (0.12) (0.13) (0.09) (0.25) (0.68)
Fact-check denial -0.11 -0.11* -0.09 -0.05 -0.06 -0.06
(0.11) (0.07) (0.09) (0.05) (0.17) (0.10)
Denial ⇥ Trump support 0.09 0.22* 0.34** 0.19** 0.21 0.82
(0.17) (0.12) (0.14) (0.09) (0.26) (0.68)
Denial/source derogation -0.21* -0.07 -0.04 -0.08 0.05 -0.09
(0.12) (0.07) (0.10) (0.06) (0.16) (0.10)
Derogation ⇥ Trump support 0.16 0.11 0.22 0.20** 0.06 0.88
(0.17) (0.13) (0.14) (0.09) (0.25) (0.68)
Constant 3.95*** 3.35*** 2.15*** 1.95*** 5.41*** 4.50***
(0.08) (0.05) (0.07) (0.04) (0.12) (0.07)
Clinton supporters 0.00 -0.09 -0.03 -0.04 -0.18 0.24
(0.13) (0.07) (0.10) (0.15) (0.18) (0.34)
Trump supporters -0.02 0.04 0.16* 0.05 0.07 0.71
(0.13) (0.10) (0.09) (0.17) (0.20) (0.70)
Clinton supporters 0.15 -0.07 -0.08 -0.12 0.04 0.20
(0.11) (0.07) (0.10) (0.15) (0.17) (0.34)
Trump supporters 0.11 0.06 0.22** 0.37** 0.37* 0.68
(0.14) (0.09) (0.10) (0.17) (0.21) (0.71)
Clinton supporters 0.04 -0.03 -0.03 -0.05 0.16 0.17
(0.12) (0.07) (0.10) (0.16) (0.17) (0.34)
Trump supporter 0.07 0.00 0.14 0.55** 0.34 0.71
(0.13) (0.10) (0.10) (0.22) (0.21) (0.70)
R2 0.03 0.04 0.11 0.12 0.02 0.01
N 990 2430 987 2430 991 2430
C-15
Table C16: Message effects on issue importance/policy preferences (ordered probit)
Crime importance Treatment of criminals Law enforcement $

MC MT MC MT MC MT
Trump support 0.22 0.26** 0.47*** 0.62*** 0.23 0.39***
(0.15) (0.10) (0.16) (0.10) (0.15) (0.11)
Uncorrected statement -0.32** -0.05 -0.02 -0.03 -0.10 0.04
(0.15) (0.08) (0.16) (0.08) (0.14) (0.08)
Uncorrected ⇥ Trump support 0.16 0.11 0.07 0.30* -0.10 0.03
(0.22) (0.14) (0.23) (0.15) (0.22) (0.15)
Fact-check -0.31** -0.16** -0.06 -0.05 -0.21 -0.02
(0.15) (0.08) (0.15) (0.08) (0.14) (0.08)
Fact-check ⇥ Trump support 0.13 0.26* 0.43* 0.30** 0.05 0.14
(0.21) (0.15) (0.23) (0.15) (0.21) (0.15)
Fact-check denial -0.14 -0.13* -0.14 -0.07 -0.06 -0.05
(0.14) (0.08) (0.14) (0.08) (0.14) (0.08)
Denial ⇥ Trump support 0.13 0.27* 0.69*** 0.35** 0.20 0.16
(0.22) (0.14) (0.26) (0.16) (0.22) (0.15)
Denial/source derogation -0.27* -0.08 -0.06 -0.12 0.02 -0.08
(0.15) (0.08) (0.15) (0.09) (0.13) (0.08)
Derogation ⇥ Trump support 0.20 0.13 0.41* 0.33** 0.07 0.19
(0.22) (0.15) (0.25) (0.15) (0.22) (0.15)
Clinton supporters 0.00 -0.11 -0.04 -0.02 -0.11 -0.06
(0.15) (0.08) (0.15) (0.09) (0.14) (0.08)
Trump supporters -0.03 0.05 0.32* 0.27** 0.03 0.05
(0.16) (0.12) (0.17) (0.13) (0.16) (0.13)
Clinton supporters 0.18 -0.08 -0.12 -0.04 0.04 -0.09
(0.14) (0.08) (0.15) (0.09) (0.13) (0.08)
Trump supporters 0.14 0.08 0.50** 0.24* 0.33* 0.04
(0.18) (0.11) (0.22) (0.12) (0.18) (0.12)
Clinton supporters 0.05 -0.03 -0.05 -0.09 0.11 -0.12
(0.14) (0.09) (0.16) (0.09) (0.13) (0.08)
Trump supporter 0.09 -0.01 0.30 0.27** 0.28 0.04
(0.17) (0.13) (0.20) (0.13) (0.18) (0.12)
N 990 2430 987 2430 991 2430
C-16
Table C17: Message exposure effects on Trump favorability (ordered probit)

(0.19) (0.11)
(0.20) (0.12)
(0.25) (0.16)
Fact-check 0.19 -0.23*
(0.19) (0.13)
Fact-check ⇥ Trump support -0.23 0.14
(0.24) (0.16)
Fact-check denial 0.13 -0.09
(0.19) (0.12)
Denial ⇥ Trump supportt -0.05 -0.03
(0.24) (0.15)
Denial/source derogation -0.17 -0.11
(0.20) (0.12)
Derogation ⇥ Trump support 0.27 0.06
(0.26) (0.16)
(0.19) (0.13)
Trump supporters -0.10 -0.19*
(0.14) (0.11)
(0.19) (0.12)
Trump supporters 0.01 -0.21*
(0.15) (0.11)
(0.19) (0.12)
(0.18) (0.12)
N 989 2430
C-17
Table C18: Message exposure effects on Trump favorability (all respondents)

(0.17) (0.13)
Clinton support -0.36** -0.35***
(0.17) (0.11)
Uncorrected statement 0.50** -0.06
(0.22) (0.13)
(0.26) (0.18)
Uncorrected ⇥ Clinton support -0.54** -0.01
(0.25) (0.14)
Fact-check 0.00 0.19
(0.19) (0.14)
Fact-check ⇥ Trump support -0.01 -0.34*
(0.24) (0.19)
Fact-check ⇥ Clinton support 0.10 -0.31**
(0.23) (0.16)
Fact-check denial 0.14 0.01
(0.21) (0.13)
Fact-check denial ⇥ Trump support -0.03 -0.14
(0.26) (0.18)
Fact-check denial ⇥ Clinton support -0.04 -0.06
(0.25) (0.15)
Denial/source derogation 0.13 0.13
(0.20) (0.14)
Denial/source derogation ⇥ Trump support -0.09 -0.24
(0.27) (0.19)
Denial/source derogation ⇥ Clinton support -0.26 -0.18
(0.23) (0.15)
Constant 1.79*** 1.70***
(0.14) (0.10)
(0.13) (0.06)
(0.14) (0.13)
(0.13) (0.06)
(0.16) (0.14)
(0.11) (0.06)
(0.18) (0.14)
R2 0.59 0.53
N 1198 2983
* p < 0.10, ** p < 0.05, *** p < .01; OLS models with robust standard errors. Respondents are Morning
Consult respondents (MC) and Amazon Mechanical Turk workers (MT).
C-18
Table C19: Message exposure effects on politician favorability
Obama favorability Clinton favorability

MC MT MC MT
Trump support -2.67*** -2.43*** -2.77*** -2.26***
(0.14) (0.10) (0.12) (0.09)
Uncorrected statement -0.00 -0.02 0.10 -0.01
(0.12) (0.08) (0.13) (0.08)
Uncorrected ⇥ Trump support 0.04 0.15 0.04 -0.04
(0.21) (0.15) (0.19) (0.13)
Fact-check -0.11 0.05 0.03 -0.09
(0.12) (0.08) (0.12) (0.09)
Fact-check ⇥ Trump support -0.07 0.13 -0.04 0.20
(0.19) (0.15) (0.17) (0.14)
Fact-check denial -0.08 0.02 -0.06 -0.12
(0.12) (0.07) (0.14) (0.08)
Denial ⇥ Trump support -0.05 -0.01 0.14 0.07
(0.20) (0.14) (0.19) (0.13)
Denial/source derogation -0.01 0.04 -0.03 -0.03
(0.12) (0.08) (0.13) (0.08)
Derogation ⇥ Trump support 0.06 0.19 0.14 0.13
(0.22) (0.15) (0.19) (0.14)
Constant 4.38*** 4.17*** 4.15*** 3.65***
(0.08) (0.05) (0.09) (0.06)
Clinton supporters -0.11 0.07 -0.07 -0.09
(0.12) (0.08) (0.12) (0.09)
Trump supporters -0.21 0.05 -0.14 0.15
(0.17) (0.14) (0.15) (0.10)
(0.12) (0.08) (0.13) (0.09)
Trump supporters -0.17 -0.13 -0.06 0.00
(0.18) (0.13) (0.16) (0.10)
Clinton supporters 0.00 0.06 -0.13 -0.03
(0.12) (0.08) (0.13) (0.09)
Trump supporters 0.02 0.10 -0.03 0.14
(0.20) (0.13) (0.17) (0.11)
R2 0.64 0.52 0.67 0.48
N 990 2430 990 2430
C-19
Table C20: Message exposure effects on politician favorability (ordered probit)
Obama favorability Clinton favorability

MC MT MC MT
Trump support -2.35*** -2.12*** -2.59*** -2.18***
(0.17) (0.12) (0.17) (0.13)
Uncorrected statement -0.04 -0.02 0.09 0.01
(0.15) (0.08) (0.14) (0.08)
Uncorrected ⇥ Trump support 0.06 0.16 0.10 -0.12
(0.25) (0.16) (0.26) (0.19)
Fact-check -0.14 0.09 -0.02 -0.07
(0.15) (0.09) (0.13) (0.08)
Fact-check ⇥ Trump support -0.06 0.11 -0.05 0.24
(0.23) (0.16) (0.24) (0.18)
Fact-check denial -0.13 0.01 -0.04 -0.08
(0.14) (0.08) (0.14) (0.08)
Denial ⇥ Trump support -0.02 0.02 0.19 -0.04
(0.24) (0.16) (0.25) (0.19)
Denial/source derogation 0.02 0.06 -0.05 -0.02
(0.15) (0.09) (0.14) (0.08)
Derogation ⇥ Trump support 0.03 0.20 0.22 0.14
(0.25) (0.16) (0.26) (0.19)
(0.15) (0.09) (0.14) (0.08)
Trump supporters -0.22 0.06 -0.27 0.28
(0.20) (0.13) (0.22) (0.17)
(0.14) (0.08) (0.15) (0.08)
Trump supporters -0.17 -0.11 -0.05 -0.01
(0.21) (0.13) (0.23) (0.18)
Clinton supporters 0.06 0.08 -0.14 -0.02
(0.15) (0.09) (0.14) (0.08)
Trump supporters 0.04 0.12 -0.02 0.24
(0.21) (0.13) (0.24) (0.18)
N 990 2430 990 2430
C-20
Table C21: Message exposure effects on unemployment beliefs (ordered probit)
(1) (2)
Trump support 0.35⇤⇤⇤ 0.36⇤⇤⇤
(0.10) (0.11)
Fact-check 0.31⇤⇤⇤ 0.31⇤⇤⇤
(0.09) (0.09)
(0.15) (0.15)
Condition: MSNBC (post-debate coverage) 0.05
(0.12)
Condition: MSNBC (no post-debate coverage) 0.03
(0.12)
Condition: Fox (post-debate coverage) 0.09
(0.12)
Condition: Fox (no post-debate coverage) 0.04
(0.12)
Constant 2.70⇤⇤⇤ 2.71⇤⇤⇤
(0.07) (0.12)
(0.09) (0.09)
Trump supporters -0.38** -0.38**
(0.12) (0.12)
N 825 825
points omitted). Respondents are Mechanical Turk workers who supported Hillary Clinton or Donald Trump
and were assigned to watch the first presidential debate in 2016 (reference category for Trump support is
thus Clinton support). The omitted category for the orthogonal media manipulation is C-SPAN with no
post-debate coverage.
C-21
Table C22: Message exposure effects on perceptions of Trump performance (probit)
Trump won debate Trump evaluation Trump vote

Trump support (W1) 2.59⇤⇤⇤ 4.86⇤⇤⇤ 3.33⇤⇤⇤
(0.27) (0.11) (0.27)
Fact-check 0.28 0.23 0.01
(0.29) (0.16) (0.27)
Fact-check ⇥ Trump support (W1) 0.68⇤ 0.23 0.03
(0.35) (0.17) (0.39)
Constant 2.11⇤⇤⇤ 0.89⇤⇤⇤ 1.89⇤⇤⇤
(0.23) (0.11) (0.19)
Clinton supporters 0.28 0.02 -0.01
(0.29) (0.16) (0.27)
Trump supporters -0.39* 0.0001 0.04
(0.12) (0.12) (0.28)
N 527 527 527
* p < 0.10, ** p < 0.05, *** p < .01 (two-sided); probit models with robust standard errors. Respondents
are Mechanical Turk workers who supported Hillary Clinton or Donald Trump and were assigned to watch
the first presidential debate in 2016 (reference category for Trump support is thus Clinton support).
C-22

Trump Corrections

Uploaded by

Document Information

Original Title

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

Trump Corrections

Uploaded by

Copyright:

Available Formats

Taking Fact-checks Literally But Not Seriously?

The Effects of Journalistic Fact-checking on

Brendan Nyhan† Ethan Porter‡

Jason Reifler§ Thomas J. Wood¶

January 11, 2019

Theory and hypotheses

H1 (motivated resistance): Respondents will resist unwelcome facts about controversial

H2 (differential acceptance): Respondents will accept journalistic fact-checks that are

H3 (attitude-consistent interpretation): Respondents will accept journalistic fact-checks

Our fact-check for the same issue in Study 1 is very similar:

In fact, according to the Bureau of Labor Statistics, unemployment has fallen

Study 1: Crime perceptions

Experimental design and instrument

• Control: A birdwatching article.

• Rising crime message: A news article summarizing Trump’s claims

• Rising crime message + fact-check: A news article summarizing Trump’s claims

• Rising crime message + fact-check + denial + source derogation: A news article

Morning Consult Mechanical Turk

for Clinton and Trump supporters relative to the uncorrected condition.

Article accurate Article fairness Statistics accurate

Morning Consult Mechanical Turk

study differs slightly from the official transcript.

derogation conditions. The omitted category is an uncorrected statement.

orthogonal media manipulation.

Table 5: Fact-check effects at Wave 3

Trump won debate Trump favorability Trump vote

Coppock, Alexander, and Andrew Guess. N.d. “Does Counter-Attitudinal Information

Schleifer, Theodore. 2016. “Paul Manafort doubts FBI statistics af-

Mechanical Turk questionnaire introductory section

What is your age?

In which state do you currently reside? [Drop down menu]

What is your gender?

What is your annual household income?

[If “100 thousand or more dollars” is selected] Is that...

Generally speaking, do you think of yourself as...

[If “Democrat” is selected] Would you call yourself...

[If “Republican” is selected] Would you call yourself...

What is the last grade or class you completed in school?

Which term best describes your race or background?

Morning Consult questionnaire introductory section

What is your age?

What is your gender?

What is the last grade or class you completed in school?

Which term best describes your race or background?

Generally speaking, do you think of yourself as...

[If “Democrat” is selected] Would you call yourself...

[If “Republican” is selected] Would you call yourself...

[Omitted Morning Consult questions not present in the MT questionnaire]

In which state do you currently reside? [Drop down menu]

[Omitted Morning Consult questions not present in the MT questionnaire]

[Omitted Morning Consult questions not present in the MT questionnaire]

[If “100 thousand or more dollars” is selected] Is that...

[Omitted Morning Consult questions not present in the MT questionnaire]

[Omitted Morning Consult questions not present in the MT questionnaire]

How much do you personally worry about crime and violence?

How important is the issue of crime to you personally?

How important is the issue of health care to you personally?

How important is the issue of global warming/climate change to you personally?

How many U.S. senators are there from each state?

Who is the Prime Minister of the United Kingdom?

[Article condition 1: control]

[Article condition 2: Trump rising crime message]

[Article condition 3: Trump rising crime message + fact-check]