Does Media Coverage of Immigration Lead to Anti-Immigrant Behavior?

Home , 2019 Oval Office address

Beyond Agenda Setting: Does Media Coverage of Immigration Lead to Anti-Immigrant Behavior?

Masha Krupenkin, Shawndra Hill, David Rothschild

June 22, 2020 Abstract

This paper studies the influence of news coverage of immigrants on anti-immigrant beliefs and behaviors. Using news transcripts, Google Trends data, and a novel dataset of Bing web searches for immigration-related topics, we examine the role of media coverage of immigration topics during the 2016 election and Trump presidency on searches for information about immigrants related to crime and welfare dependency as well as how to report immigrants to Immigration and Customs Enforcement (ICE). We catalogue significant and sustained increases in news segments on crime by immigrants and their use of public services after Trump’s inauguration and find that searches for crime and welfare related to immigrants are correlated with the daily volume of news on immigrant crime and welfare. We find a similar sustained increase in searches for how to report undocumented immigrants that is strongly correlated with the daily volume of immigrant crime news. Using timestamped searches for immigration during broadcasts of Trump’s and Obama’s speeches, we confirm the causal effect of anti- immigrant media coverage on searches for immigration and crime, welfare, and how to report immigrants to ICE. The findings indicate that media’s choices regarding the coverage of immigrants can have a strong impact not only on attitudes but also behaviors. Introduction

The 2016 election ushered in a sea change regarding political discourse about immigration. Donald Trump made immigration a signature issue, kicking off his campaign with the pro- nouncement that “When Mexico sends its people, they’re not sending their best... They’re sending people that have lots of problems, and they’re bringing those problems with us.” Jour- nalists have long debated the media’s role in actively preventing the spread of misinformation,1 especially when that misinformation may have negative consequences, and prior research has shown that media narratives significantly influence viewers’ support for policy issues (Chong and Druckman 2007), including immigration (Knoll et al. 2011; Lahav and Courtemanche 2012). These debates found new relevance in 2019, as media outlets deliberated over whether to broadcast live an address by President Trump on immigration due to concerns of “malice and misinformation” regarding immigrants.2 We further examine the potential influence of media coverage on attitudes and behaviors. Using web search data, we find that anti-immigrant media broadcasts significantly affect not only anti-immigrant beliefs but also anti-immigrant behaviors.

This media and political scenario raises several questions. Does media coverage of issues related to immigrants strengthen and reinforce (in this case, negative) attitudes towards immigrants? Does it encourage Americans to engage in anti-immigrant behaviors such as reporting undocumented immigrants to ICE? To address these questions, we combine a novel dataset of Bing and Google searches for crime committed by immigrants, immigrants’ use of welfare, and how to report immigrants to ICE with automated text analysis of TV transcripts from CNN, MSNBC, and Fox News from 2014 to 2019. This methodology allows us to measure accurately Americans’ attitudes and behaviors regarding immigrants, instead of relying on survey data, which can be plagued by non-attitudes and social desirability bias. Furthermore,

1https://archives.cjr.org/feature/rethinking_objectivity 2https: www.nbcnews.com/news/all/air-or-not-air-networks-face-pressure-over-broadcasting-trump-n955846

1 because of the temporal granularity of our data, we can map when a user searched for an anti- immigrant term and thus determine of media events have a causal effect on searches.

We illustrate large shifts in news coverage of immigration after the 2016 election, especially on Fox News, and we identify a strong correlation between daily coverage of immigration and anti-immigrant searches—not just attitudinal searches about immigrant crime and welfare but also behavioral searches about how to report immigrants to ICE. Using this timestamped search data, we ﬁnd that anti-immigrant searches dramatically increased during broadcasts of President Trump’s speeches about immigration, suggesting a causal effect of anti- immigrant media broadcasts. We do not ﬁnd similar effects for Obama’s televised addresses.

We show that immigration news coverage can lead to both the reinforcement of anti-immigrant beliefs and the impetus to act on those beliefs. Frequently, immigration news coverage includes misinformation distributed by the media during speeches or press conferences by an established politician proliferating unvetted information or in segments covering such events in the following hours, days, and weeks. The downstream effects of this news coverage include the impacts of consuming not only the news coverage of the event but also the corresponding information available on the internet. Media coverage can lead to a spiral of reinforcing disin- formation, as consumers are motivated to engage in information-seeking behavior online after consuming a news report. Politicians disseminating and discussing anti-immigrant rhetoric can and do change the attitudes and actions of their supporters, but the breadth and depth of the impacts are dependent on the media’s choice to amplify the issue and frame the discussion, with serious real-world consequences to those choices.

The rest of the article is organized as follows. The subsequent section describes the theoretical debates that motivate this study. The third section presents our data sources and de- lineates our analysis strategy. The fourth section explores changes in news coverage of immigration after the 2016 election. The ﬁfth section demonstrates the increase in anti-immigrant searches after the 2016 election. The sixth section shows the strong association between news

2 segments about immigration and the volume of anti-immigrant searches. The seventh section uses hourly search data to estimate the causal effects of Donald Trump’s televised speeches on anti-immigrant searches and compares these to effects from speeches by Barack Obama. The ﬁnal section describes the implications of our ﬁndings.

Theory

Scholars studying the media have identified several effects of the media on public opinion (Strömberg 2015). Among the most important is the role of the mass media in activating public opinion. When a person’s opinion on an issue is activated, their opinion about this issue is both “salient in the mind and impels [them] to political action” (Lee 2002). Through activating public opinion, mass media can compel the public to engage in a variety of behaviors that can influence the political process. These behaviors include expressive behaviors such as tweeting about an issue (King et al. 2017) or communicating with elected officials (Lee 2002) and information-seeking behaviors with the aim of learning more about the issue of interest. Finally, some members of the public attempt to influence politics by modifying other behaviors such as changing their family planning (Bhavnani and Nellis 2016) or engaging in violence against ethnic outgroups (Yanagizawa-Drott 2014).

Activated public opinion has played a key role in shaping the politics of race and ethnicity in the United States. Lee (2002) described the role of civil rights activists in activating North- ern whites’ public opinion, which he argues led to the passage of civil rights legislation in the 1960s. More recently, due to the election of the nation’s ﬁrst black president, as well as the concurrent rise in non-white immigration, the issue of race has been extremely salient and “chronically accessible” (Tesler and Sears 2010). Some whites perceive a threat from the rise of non-white immigration, which has encouraged backlash (Abrajano and Hajnal 2017; Jardina 2019). The election of Donald Trump has further increased the salience of race, with Ameri-

3 cans’ worries about race relations at a record high.3

We argue that at this stage of chronic accessibility, the overwhelming saliency of race keeps the issue bubbling just below the surface for many members of the public, so that an event as small as an anti-immigrant media broadcast can activate anti-immigrant opinion, resulting in behaviors that have tangible negative impacts on immigrant communities, such as reporting (presumed) undocumented immigrants to ICE. Deportation carries obvious negative consequences for undocumented immigrants, including the separation of families, loss of liveli- hood, and risk of violence when deported to their countries of origin. Fear of deportation has extremely negative effects on both mental and physical health for undocumented immigrants, their families, and their communities (Allen et al. 2015; Gemmill et al. 2019; Krupenkin et al. 2019; Torres et al. 2018). ICE arrests can carry negative consequences for undocumented immigrants as well as citizens and legal residents. Numerous U.S. citizens have been arrested by ICE and held for long periods before they could prove their citizenship.4

Media coverage is an extremely important propagator of presidential cues. Americans tend to follow cues from co-partisan political ﬁgures when forming their policy preferences (Lenz 2013; Achen and Bartels 2017; Layman and Carsey 2002; Levendusky 2009). While presidents communicate with their supporters through mailings and, more recently, social media, the press plays an extremely important role in mediating these cues (Dalton et al. 1998), amplifying some while diminishing others. This role is especially salient in the era of Trump, with media organizations vigorously debating their role in fact-checking inaccurate claims.5

The media debate over amplifying Trump’s cues was notable in media discussions over whether to broadcast his 2019 Oval Ofﬁce address on immigration live on the air. Some journalists argued that the speech could contain an announcement of new policy while others be- lieved that amplifying the anti-immigrant message of the speech was inappropriate.6 Ulti-

3https://news.gallup.com/poll/1687/race-relations.aspx 4https://www.latimes.com/local/lanow/la-me-citizens-ice-20180427-htmlstory.html 5https://www.theguardian.com/media/2020/jan/06/media-leaders-agonize-amplifying-donald-trump-lies-2020-election 6https://www.forbes.com/sites/markjoyella/2019/01/08/trump-got-his-oval-office-address-in-prime-time-the-networks-got-played

4 mately, the speech contained no new policy content and gave Trump a signiﬁcant fundraising boost, with the amount of money raised surpassing the amount raised after his 2018 State of the Union address.7 The Washington Post referred to the speech as a “misleading and bleak picture of the situation at the U.S.-Mexico border.” 8 Ultimately, while Trump issues cues through other media such as Twitter, by broadcasting his speeches, the mass media serves as a crucial propagator of anti-immigrant messages.

Mass media has long engaged in negative portrayals of immigrants. Between 2000 and 2010, media regularly showed immigrant arrests and detentions, implying criminality (Far- ris and Silber Mohamed 2018). Immigrants have been disproportionately portrayed as male and Latino (Silber Mohamed and Farris 2019), invoking the “Latino threat” narrative, which emphasizes criminality and a lack of assimilation as common traits of immigrants from Latin America (Abrajano et al. 2017). These portrayals have not been without consequence. Brader et al. (2008) showed that anti-Latino cues trigger opposition to immigration, and Abrajano et al. (2017) found that negative framing of immigration is associated with macropartisan shifts toward the Republican party. In laboratory studies, fear cues about immigrants tended to increase information-seeking behavior (Gadarian and Albertson 2014). Our contribution to this literature is to highlight the causal relationship between anti-immigrant media coverage, information- seeking behavior, and immigrant reporting behavior.

Web search data is uniquely well-suited to measuring activated public opinion on immigration, especially compared to survey data. Survey data has several notable ﬂaws when measuring activated public opinion. First, survey data does a poor job of distinguishing between genuinely held beliefs and question-induced responses (non-attitudes), much less activated public opinion. Survey data cannot tell us whether a respondent would have given any thought to immigration without being prompted by the survey to do so. In contrast, web searches represent individually motivated behavior that is free from the inﬂuence of survey questions.

7https://www.thedailybeast.com/trumps-oval-office-speech-was-a-dud-but-it-could-be-a-fundraising-success 8https://www.washingtonpost.com/politics/2019/01/09/fact-checking-president-trumps-oval-office-address-immigration/

5 Furthermore, surveys provide little insight into political actions associated with a particular survey response. If there is an increase in the percentage of Americans who believe that immigration is a “good thing”,9 does this mean that these individuals are now more likely to donate money to pro-immigrant organizations? More likely to socialize with and/or hire immigrants? More likely to protest anti-immigrant legislation? It is difﬁcult to determine whether and how pro-immigrant survey answers correspond with pro-immigrant behaviors, and surveys that speciﬁcally ask about behaviors tend to be a poor measure of those actual behaviors(Ansolabehere and Hersh 2012). On the other hand, web searches are themselves an information- seeking behavior (signalling genuine interest in and engagement with an issue) and often point to explicit real-world behaviors.

Survey data can also be subject to signiﬁcant social desirability bias, a problem elided by the use of search data (Stephens-Davidowitz and Pabon 2017). Surveys’ social desirability bias is especially salient regarding immigration survey responses after the 2016 election. Trump ran on an anti-political correctness platform, arguing that Americans ought to feel more com- fortable expressing “politically incorrect” attitudes about race, ethnicity, and gender. Donald Trump’s victory in the 2016 election shifted Americans’ perceptions on the social acceptability of racially prejudiced statements.10

Given the high levels of partisan polarization among Americans (Iyengar et al. 2019), it is likely that Trump’s campaign had differential effects on the willingness of Democrats and Republicans to express controversial opinions about immigration. Survey data from Repub- licans, who were encouraged by Trump’s campaign, may be subject to a weaker social desirability effect around questions of immigration compared with pre-2016 survey data, while survey data from Democrats, who largely dislike Trump, may be subject to a stronger effect. Given these potential differences in social desirability effects, it is possible that polls show- ing Americans’ increased support for immigrants are an artifact of increased social desirability

9https://news.gallup.com/poll/235793/record-high-americans-say-immigration-good-thing. aspx 10https:\www.pewsocialtrends.org/2019/04/09/race-in-america-2019/

6 constraints brought on by anti-Trump sentiment rather than a genuine shift in opinion.

In this paper, we study the effect of anti-immigrant media coverage on two types of web searches—searches seeking information about anti-immigrant narratives and searches seeking information about how to report immigrants to ICE. We ﬁrst look at the role of media coverage in inducing information-seeking behavior concerning two dimensions of immigration— crime rates among immigrants and immigrants’ use of government beneﬁts. We chose these two dimensions due to both their prevalence in current rhetoric about immigration as well as their overall salience in modern discussions of race and ethnicity as components of the “immigrant threat” narrative (Abrajano and Hajnal 2017).

In many cases, information-seeking searches can lead to the reinforcement of incorrectly held beliefs (White 2013). Partisans engage in selective exposure to online news (Peterson et al. 2019), and search engine results may be especially vulnerable to selective exposure and conﬁrmation bias due to the simultaneous presentation of an array of information sources (Jonas et al. 2001). Indeed, Knobloch-Westerwick et al. (2014) found that web searchers exhibit strong conﬁrmation bias, preferring attitude-consistent results over attitude-inconsistent results. The misinformation received by people who consume inaccurate anti-immigrant messages via media broadcasts is likely to be reinforced if they engage in web search on related topics. Mere repetition of inaccurate news information increases its credibility among viewers (Pennycook et al. 2018), and the repetition of messages on particular issues increases their salience and impact on vote choice (Claibourn 2008).

We move beyond information-seeking behavior about immigrants to look at searches for information about reporting suspected immigration violators to ICE, a concrete anti-immigrant action that can have signiﬁcant negative consequences for individuals who are reported. Web searches have been shown to accurately predict consumption behavior (Goel et al. 2010), the volume of stock sales (Bordino et al. 2012), unemployment rates (Ettredge et al. 2005; DâA- muri and Marcucci 2017), cancer rates (Ofran et al. 2012), home and auto sales (Choi and Var-

7 ian 2012), tourism (Gawlik et al. 2011), voter registration (Street et al. 2015), voter turnout (Stephens-Davidowitz and Pabon 2017), and numerous other real-world behaviors. By studying searches for information on reporting immigrants to ICE, we can accurately measure a behavior that, while rare, has clear negative and signiﬁcant consequences for immigrants. A link between anti-immigrant media coverage and reporting behavior provides signiﬁcant evidence of anti-immigrant behavior driven by negative media coverage.

In this work, we test the effects of media coverage on two dimensions of activated public opinion: information-seeking behavior and political action. In the case of information-seeking behavior, we ﬁnd a strong association between anti-immigrant media coverage and the daily volume of anti-immigrant searches. This association is the strongest for topics explicitly covered by media broadcasts. An increase in media coverage of immigrant crime is strongly associated with an increase in searches for the same topic. With respect to political action, negative portrayals of immigrants are associated with a higher volume of searches on how to report immigrants. Finally, these relationships are causal as both information-seeking and immigrant- reporting behavior are found to spike during a widely-watched anti-immigrant media broadcast.

Empirical Approach

This paper uses a wide variety of data sources to study the effects of media messaging on anti- immigrant behavior. The main outcome variable of interest is web search behavior, measured by Bing and Google searches, at the daily and hourly levels.11 Table 1 summarizes the data sources and their availability.

We use two approaches to measure the effect of media on immigration searches. The ﬁrst approach looks at daily Google and Bing searches for anti-immigrant keywords and compares the volume of these searches with the amount of media coverage of anti-immigrant topics.

11The use of Bing data was reviewed and cleared by the Microsoft IRB-594.

8 Table 1: Data Summary Dataset Measures Availability Long Term Trends: Bing data Daily Bing searches 03/16-08/19 Google Trends data Daily Google searches 01/04-08/19

Media Effects: News transcript dataa News coverage 01/14-08/19 Hourly Bing search data Effects of televised Trump speeches 01/18-08/19 Hourly Google Trends data Effects of televised presidential speeches 01/15-08/19

Validation: Geographic Bing data Validate immigration search terms 03/16-08/19 CCES immigration question Validate immigration search terms

aNews transcript data available only for 01/15 - 08/19 for MSNBC

This approach tests whether there is an association between anti-immigrant media coverage and anti-immigrant search volume. Speciﬁcally, on days when the media spends more time discussing immigrant crime, are searches for reporting immigrants to ICE more common? The second approach looks at whether changes in anti-immigrant search patterns can be directly attributed to anti-immigrant media coverage. Here, we compare timestamped searches during Trump’s and Obama’s televised speeches to determine whether there is a spike in anti- immigrant web searches during a widely-watched anti-immigrant media broadcast.

Search Data Measures of Anti-Immigrant Beliefs and Behaviors In Real

Time

To measure anti-immigrant beliefs and behavior, we turn to web search data. Speciﬁcally, we look at Google and Bing for searches about immigration and crime, immigration and welfare, and information about reporting immigrants to ICE. We also use Bing and Google searches that include the term “weather” as a placebo, as we do not expect a relationship between anti-

9 Table 2: Top 10 Bing Searches For Each Immigrant or Immigration Topic

Crime Welfare Report Criminal Immigrants do illegal immigrants get welfare how to report illegal immigrants

fbi illegal immigration crime statistics can illegal immigrants get welfare report illegal immigrants

illegal immigrant crime statistics illegal immigrants on welfare how to report illegal immigrants anonymously

illegal immigrants crime statistics do illegal immigrants get beneﬁts how to report someone to immigration

illegal immigrant population statistics crime cost of illegal immigration how to report an illegal immigrant

illegal alien crime do illegals get government beneﬁts report illegal immigrants anonymously

illegal alien crime report how much does illegal immigrants cost america how to report immigration fraud anonymously

americans killed by illegal aliens how much do illegal immigrants cost taxpayers ice report illegal immigrants anonymously

illegal immigrant crime cost of illegal immigration 2018 report immigration fraud

Illegal Aliens Kill A Woman In Washing- cost of illegal immigrants to taxpayers How to Report Someone to Immigration ton, State

Notes: The table represents the top 10 searches for each search category in the United States. Searches are overwhelmingly negative in tone towards immigrants and are overwhelmingly focus on unauthorized immigrants.

10 immigrant media coverage and searches for weather.

For Bing searches,12 we define an immigration search as one that contains one of the following three strings: “immigr” (which includes words like “immigrant”,“immigration”), “illegals”, and “illegal alien”, which would also capture any searches for “undocumented immigr*”. A crime search is defined as a search that contains one of the listed immigration terms, plus one or more of the following: “crime”,“criminal”, “kill”, “murder”. A welfare search contains one of the immigration terms plus one of the following: “welfare”, “benefits”, “cost”. These lists are not exhaustive, but they are intended to be precise and to capture a substantial proportion of searches related to immigration/crime and immigration/welfare. Finally, we measure “reporting” searches on Bing using a similar measure as the crime and welfare measures that comprises searches containing the word “report” and one of the immigration terms or a string “to ice” (e.g., “how to report someone to ice”). Table 2 shows the top 10 Bing searches for each topic. These searches clearly express anti-immigrant sentiment.

To measure Google search volume, we downloaded results from Google Trends using the pytrends Python library. Pytrends provides a distinct advantage as it allows us to pull historical hourly search data, which is unavailable through the browser-based Google Trends interface. A downside of the Google Trends data, in contrast with the Bing data, is that it limits the number of terms that can be used and does not allow the manual removal of terms that, while match- ing the keywords, are not related to immigration (e.g., “news immigration report”).13 Despite these limitations, we ﬁnd that the Google Trends results consistently replicate the Bing results.

To measure immigration and crime searches on Google Trends, we pulled searches with the following string: “immigrant crime+immigrant criminal+immigrant murder+immigrant kill”. Immigration and welfare searches were represented by “immigrant welfare+immigrant cost+immigrant beneﬁts”, and immigration reporting searches were represented by “report im-

12We pull Bing searches from the en-US browser-based search market. 13The appendix includes a list of all terms removed from Bing search results as irrelevant or entered by an automated bot.

11 Table 3: Immigration Terminology by Source

Category % "Illegal" of % "Illegal" of % Unauthorized Unauthorized Total of Total Crime searches 97.2% 71.5% 73.5% Welfare searches 94.6% 72.9% 77.0% Reporting searches 97.5% 82.1% 84.2%

Fox News 95.2% 43.9% 46.0% MSNBC 56.1% 10.1% 17.9% CNN 50.7% 12.6% 24.8%

Notes: This table demonstrates the variation in sources’ language when describing immigrants. Quotes from government ofﬁcials are included in news sources. The ﬁrst column is the percentage of all mentions of unauthorized immigrants that include the terms “illegal immigrant”, “illegals”, or “illegal alien” (as opposed to “undocumented immigrant” or “unauthorized immigrant”). The second column is the percentage of all mentions of immigrants that include the term “"illegal”. Finally, the third column is the percentage of all mentions of unauthorized immigrants. Among the three news channels, Fox News was more likely to talk about unauthorized immigrants and more likely to use the term “illegal”. Searches for crime, welfare, and reporting tend to include language describing immigrants that has a higher level of similarity to Fox News than to MSNBC or CNN. migrant+report immigration+report illegals+report illegal alien+report to ice”.

Do Crime and Welfare Searches Express Anti-Immigrant Sentiment?

To determine whether crime and welfare search terms are an expression anti-immigrant sentiment, we use two measures: the language used to describe immigrants within searches and correlation with survey indicators of anti-immigrant sentiment.

First, we look at differences in the use of the terms “illegal immigrant”, “illegals”, and ”illegal alien” versus the term “undocumented immigrant” to describe unauthorized immigration. The “illegal” framing of unauthorized immigration is considered an expression of anti-immigrant sentiment.14 Table 3 describes the percentage of references to unauthorized

14New York Times journalists are explicitly discouraged from using the “illegal” framing in their reporting

12 Table 4: Immigration Searches and Anti-Immigrant Sentiment

Dependent variable: Crime Welfare County % foreign born 4.907∗∗∗ (1.495) −0.738 (0.618) 3.962∗∗∗ (0.968) −1.363∗∗∗ (0.470)

Oppose immigration −0.369∗ (0.207) 0.183 (0.156) Oppose immigration x % foreign born 6.274∗∗∗ (2.201) 6.102∗∗∗ (1.502)

2016 Trump vote −0.242 (0.153) 0.178 (0.109) 2016 Trump vote x % foreign born 4.926∗∗∗ (1.481) 4.330∗∗∗ (1.040) Constant −13.027∗∗∗ (0.105) −12.727∗∗∗ (0.100) −12.659∗∗∗ (0.076) −12.866∗∗∗ (0.069)

Dependent variable: Report County % foreign born 7.112∗∗∗ (1.869) 1.017 (0.884)

Oppose immigration −0.777∗∗∗ (0.282) Oppose immigration x % foreign born 6.697∗∗ (2.798)

2016 Trump vote −0.597∗∗∗ (0.217) 2016 Trump vote x % foreign born 5.395∗∗∗ (1.911) Constant −15.335∗∗∗ (0.141) −14.638∗∗∗ (0.142)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: Searches for immigration crime, welfare, and reporting are the highest in areas with high anti-immigrant sentiment and a large proportion of immigrants. Using the percentage of voters who voted for Trump in 2016 instead of anti-immigrant sentiment reveals a similar pattern. Regression is a binomial logit, with standard errors clustered by county. The dependent variable is the proportion of searches from a particular county that included an anti-immigrant search term, with each anti-immigrant search counted as 1, and all other searches counted as 0.

13 immigrants15 that use the “illegal” terminology, as well as the percentage of all immigrant searches that use the “illegal” framing in news and web searches. Bing searches for immigrant crime, welfare, and reporting overwhelmingly use the “illegal” framing rather than the “undocumented” framing, and a large proportion of searches about these topics explicitly mention “illegal” immigrants rather than immigrants in general.

Second, if our search terms are a valid measure of anti-immigrant sentiment and reporting behavior, they should correlate with anti-immigrant sentiment. To better understand the relationship between anti-immigrant sentiment and the search terms, we estimated a county- level measure of anti-immigrant sentiment using a Cooperative Congressional Election Sur- vey (CCES) question about opposition to legal immigration. We ran the raw results through a multilevel regression with post-stratification. The model is: OUTCOME (1 | State) + (1 | State:County) + Urbanicity + MarStat + Party + Age + Gender + Race + Education + Race:Education, and the post-stratification space is derived from a mixture of voter files and Census data out- lined in (Konitzer et al. 2017). This allows us to compare the geographic frequency of search terms to geographic dispersion of anti-immigrant sentiment. Table 4 shows the relationship between anti-immigrant sentiment and immigration searches. Searches are the highest in areas with both high anti-immigrant sentiment and a large proportion of foreign-born residents, suggesting that these searches are tied to anti-immigrant sentiment.

Structural Topic Model Estimates of Content of Immigrant News Coverage

To measure immigration news coverage, we rely on MSNBC, Fox News, and CNN transcripts from 2014 to 2019. Time-stamped transcripts were downloaded from the Internet Archive. While Fox and CNN transcripts were available for 2014-2019, MSNBC was available only

(https:\www.nytimes.com/2017/03/10/insider/illegal-undocumented-unauthorized-the-terms-of-immigration-reporting. html) 15Here we are speciﬁcally measuring mentions of “illegal immigrant”, “illegals”, and “illegal alien”, as opposed to mentions of “illegal immigration” to measure negative attitudes towards immigrants themselves.

14 for 2015-2019.

To identify news segments on immigration, we used the following metric: any news segment with two or more mentions of our key strings (“immigr”, “illegals”, or “illegal alien”) within 60 seconds of each other was considered an immigration segment. The segment began at the ﬁrst mention of the keywords and ended at the last mention. Consecutive immigration segments16 were combined into a single longer immigration segment. This procedure yielded 61,229 immigration segments. This procedure generates segments with high precision, and possibly lower recall as we rely on a fairly strict set of keywords to identify the segments. Fur- thermore, counting only text bounded by keyword mentions inevitably truncates the segments, making them as short as possible since we do not count conversation before the ﬁrst and after the last keyword, thus underestimating the total length of the segment. However, since the focus of the paper is changes in media coverage over time, this truncation is not of major con- cern.

To measure changes in the media coverage of immigration before and after the 2016 election, we use a structural topic model (Wang et al. 2011) to estimate the prevalence of various immigration-related topics in CNN, MSNBC, and Fox News transcripts. Structural topic models are a form of semi-automated text analysis that allow for the inclusion of document covari- ates when estimating topical prevalence or content (Roberts et al. 2014). Topic models assign each document a topic proportion for each topic, which means that a document can (and often does) belong to multiple topics (e.g., 0.1 topic A, 0.2 topic B, 0 topic C).

To estimate topic proportions, we use a 30-topic structural topic model (STM) for the 61,229 immigration segments with an indicator variable for the time period relative to the 2016 election. The timeline is divided into three distinct periods: pre-campaign (before Trump an- nounced his candidacy), campaign (between Trump’s candidacy and his inauguration), and post-inauguration. We also include a variable for the channel (CNN, MSNBC, or Fox) and

16For example, three mentions of keywords within 120 seconds, each no more than 60 seconds after the previous mention.

15 interact it with the time period variable. In addition to the standard English stop words, the model also uses the words/phrases “immigrant”, “immigrants”, “immigrate”, “immigrated”, “immigrating”, “immigration”, “illegals”, “illegal alien”, and “illegal aliens” as stop words to prevent changes in language referring to immigrants from inﬂuencing the model.17

Table 5 shows the full results of the topic model. The model generates a wide variety of topics relating to immigration. We focus on three topics from the model: Topic 1 and Topic 3, which are classified as crime topics, and Topic 13, which is classified as the welfare topic. Figure 1 shows the documents most representative of each topic. These documents clearly represent an anti-immigrant point of view. Furthermore, as further demonstrated in the results section, these topics are significantly more likely to be covered by Fox News than by either MSNBC or CNN, further suggesting that these news segments depict immigrants in a negative light.

It is important to note the news transcripts are used as a proxy for media coverage of spe- ciﬁc aspects of the immigration issue. When we see an increase in coverage of immigration and crime in the transcripts, this suggests that multiple print, online, and TV news outlets increased their coverage of this issue. The results in this paper do not necessarily reﬂect TV alone, but news media coverage collectively, of which TV is by far the dominant mode of public consumption (Allen et al. 2020).

Measuring Media Effects Using Presidential Speeches

To determine whether changes in anti-immigrant searches can be directly attributed to anti- immigrant coverage, we look at hourly searches during presidential speeches. We use the same

17Ndulue et al. (2019) found an increase in references to unauthorized immigrants as “illegals” or “illegal aliens” after Trump was elected. In theory, these terms are likely to have a positive association with anti-immigrant coverage, so an increase in these terms could spuriously suggest an increase in anti-immigrant crime coverage, even when the only change is in the way immigrants are referenced. In practice, including these terms in stop word lists has little to no substantive effect on output of the model.

16 Table 5: Media Coverage of Immigration Topics Determined by the Structural Topic Model

Topic 1 Highest Prob: crime, crimin, illeg, commit, murder, convict, time Topic 16 Highest Prob: job, worker, work, american, economi, econom, number FREX: steinl, kate, rape, tibbett, pier, assault, crime FREX: wage, unemploy, worker, labor, farmer, growth, green Score: crime, murder, crimin, commit, kate, convict, steinl Score: worker, wage, job, skill, economi, labor, econom

Topic 2 Highest Prob: democrat, open, left, california, want, liber, pelosi Topic 17 Highest Prob: presid, hous, white, trump, report, meet, tweet FREX: oakland, nanci, liber, jerri, pelosi, pawn, warren FREX: miller, lindsey, graham, white, stephen, kushner, jare Score: democrat, pelosi, california, nanci, open, oakland, left Score: white, presid, hous, miller, graham, meet, stephen

Topic 3 Highest Prob: illeg, two, kill, year, man, ofﬁc, charg Topic 18 Highest Prob: question, ask, whether, tough, answer, get, can FREX: twice, singh, crash, man, drunk, suspect, dominican FREX: question, census, answer, tough, whether, ask, ramo Score: kill, man, polic, son, -year-old, suspect, illeg Score: question, ask, answer, tough, whether, census, count

Topic 4 Highest Prob: law, polici, administr, trump, chang, enforc, general Highest Prob: need, system, reform, comprehens, ﬁx, want, end FREX: session, jeff, polici, administr, general, attorney, book Topic 19 FREX: lotteri, chain, system, comprehens, merit, migrat, broken Score: law, polici, administr, session, attorney, enforc, jeff Score: system, reform, comprehens, need, ﬁx, lotteri, chain

Topic 5 Highest Prob: anti-, trump, rhetor, fear, peopl, group, racist Highest Prob: famili, children, separ, parent, kid, child, one FREX: brexit, anti-, manifesto, franc, nationalist, hatr, merkel Topic 20 FREX: reunit, parent, children, separ, kid, mother, zero Score: anti-, europ, racist, european, rhetor, germani, fear Score: children, famili, separ, parent, kid, child, mother

Topic 6 Highest Prob: border, wall, mexico, secur, come, crisi, stop Highest Prob: citi, feder, law, sanctuari, communiti, enforc, state FREX: southern, crisi, fenc, caravan, border, tariff, barrier Topic 21 FREX: chicago, local, citi, sanctuari, counti, rahm, cooper Score: border, wall, mexico, crisi, southern, patrol, cross Score: citi, sanctuari, feder, mayor, law, local, counti

Topic 7 Highest Prob: militari, full, street, north, war, michael, take Highest Prob: secur, plan, includ, protect, homeland, new, propos FREX: maria, korea, russian, russia, hurrican, unusu, journal Topic 22 FREX: homeland, dhs, secretari, memo, draft, includ, plan Score: militari, maria, michael, north, russia, full, russian Score: homeland, secur, plan, depart, secretari, includ, propos

Topic 8 Highest Prob: illeg, peopl, countri, want, come, million, get Highest Prob: court, judg, case, asylum, process, rule, will FREX: jess, overstay, illeg, amnesti, million, greg, everybodi Topic 23 FREX: judg, court, suprem, asylum, seeker, circuit, lawyer Score: illeg, million, peopl, amnesti, countri, want, dont Score: court, judg, asylum, suprem, process, case, facil

Topic 9 Highest Prob: peopl, just, that, dont, know, theyr, right Highest Prob: money, busi, check, washington, emerg, guy, pete FREX: theyr, cant, your, isnt, true, problem, understand Topic 24 FREX: neil, charl, pete, machin, emerg, box, check Score: theyr, dont, that, peopl, problem, know, just Score: pete, emerg, busi, money, check, neil, obamacar

Topic 10 Highest Prob: legal, status, member, gang, came, countri, citizen Highest Prob: issu, republican, parti, think, elect, voter, vote FREX: ms-, anim, legal, gang, applaus, status, distinct Topic 25 FREX: cantor, poll, parti, voter, elect, won, win Score: legal, gang, status, ms-, member, anim, applaus Score: republican, parti, voter, issu, poll, vote, elect

Topic 11 Highest Prob: presid, obama, execut, order, action, congress, will Highest Prob: state, unit, nation, countri, refuge, ban, muslim FREX: execut, action, unilater, defer, obama, permit, barack Topic 26 FREX: ban, vet, terrorist, isi, syrian, refuge, terror Score: execut, action, obama, presid, order, congress, constitut Score: unit, muslim, nation, ban, state, refuge, vet

Topic 12 Highest Prob: talk, think, hes, said, thing, say, want Highest Prob: ice, deport, enforc, crimin, undocu, agent, arrest FREX: talk, hes, bit, convers, term, littl, think Topic 27 FREX: custom, raid, ice, abolish, remov, oper, agenc Score: talk, hes, think, thing, said, lot, know Score: ice, raid, custom, agent, deport, arrest, enforc

Topic 13 Highest Prob: care, health, state, beneﬁt, undocu, school, pay Highest Prob: bill, senat, republican, democrat, deal, will, get FREX: health, tuition, insur, healthcar, care, medicaid, licens Topic 28 FREX: mcconnel, mitch, compromis, senat, bill, ryan, negoti Score: care, health, licens, welfar, school, driver, healthcar Score: bill, republican, senat, democrat, daca, hous, reform

Topic 14 Highest Prob: thank, join, new, good, news, tonight, now Highest Prob: trump, donald, campaign, will, clinton, speech, undocu FREX: ainsley, fox, join, shannon, brian, thank, martha Topic 29 FREX: soften, cruz, hillari, clinton, ted, marco, donald Score: thank, join, news, fox, morn, tonight, ainsley Score: donald, trump, rubio, hillari, cruz, clinton, marco

Topic 15 Highest Prob: countri, america, peopl, american, come, tucker, world Highest Prob: see, tri, now, make, come, back, get FREX: assimil, tucker, slave, societi, liberti, generat, statu Topic 30 FREX: type, see, tri, put, continu, seen, make Score: tucker, america, countri, american, assimil, peopl, world Score: see, type, tri, make, come, back, now

Notes: Topics of interest for this paper are in bold (Topics 1, 3, and 13). Topics 1 and 3 were coded as immigrant crime topics, and topic 13 was coded as the immigrant welfare topic. Topics were generated using the R stm package.

17 Figure 1: Immigration Crime and Welfare Topics Crime (topic 1):

Crime (topic 3):

Welfare (topic 13):

Notes: These documents have the highest association with Topics 1, 3, and 13.a

aThese docuemnts were selected using the FindThoughts() func- tion of the stm package, which uses the posterior probability of a topic given a document to return the documents most representative of a topic. The pictured documents are the top documents returned for each topic.

18 search terms described in the search data section, but we analyze the data at the hourly rather than the daily level.

To determine whether anti-immigrant searches can be attributed to anti-immigrant rhetoric, we compare searches during a speech to searches at the same time exactly one week before and one week after the speech. Searches tend to have signiﬁcant day-of-week effects. For example, people tend to search differently on weekends than on weekdays. Searches one week before or one week after are a better control group than searches one day before or one day after, as people tend to search in very different ways on Sundays than they do on Mondays.

Televised presidential addresses are an ideal test of the effects of anti-immigrant broadcasts on anti-immigrant search behavior. Each address is a major media event with an audience of tens of millions of Americans watching the speech live. Unlike nightly news broadcasts, there are no competing political media broadcasts of similar magnitude during a president’s speech, so spikes in anti-immigrant search terms during a presidential address can be attributed to the address. Finally, political media events of the same magnitude as televised presidential addresses are rare, so comparisons to search patterns at the same time one week before or after the speech are unlikely to be confounded by a different political media event.

Using the Bing data, we examine search patterns for three of Donald Trump’s speeches: his 2018 SOTU, his 2019 televised Oval Ofﬁce Address on immigration, and his 2019 SOTU. Using the hourly Google Trends data18, which is available for the prior ﬁve years, we examine searches during these three speeches, in addition to Trump’s 2017 SOTU and Obama’s 2016 and 2015 SOTUs.

The comparison with Obama’s speeches is useful because both presidents mentioned immigration in their SOTU speeches, but from very different perspectives. Trump’s comments on immigration during his speeches were much more negative, and focused on immigrant crim-

18Google Trends data was pulled using the pytrends package for the GMT timezone, which we then converted to EST.

19 inality and welfare dependence. On the other hand, Obama’s mentions of immigrants tended to be positive. This allows us to verify that anti-immigrant search spikes can be attributed to anti-immigrant messages, rather than positive messages about immigrants.

Results: Immigration Coverage Changed After 2016 Election

After the 2016 presidential election, media coverage of immigration increased and became more anti-immigrant. Figure 2 plots the total daily duration of immigration news segments across three periods and three cable news channels. While there was little change in the daily duration of immigration coverage between the pre-campaign and campaign periods, there was a signiﬁcant increase after the inauguration, especially on Fox News. While there was a modest post-inauguration increase in immigration coverage on CNN and MSNBC, Fox News’ average monthly number of immigration segments nearly doubled after the inauguration, from 303.5 to 545.2.

In addition to increases in the volume of coverage , the content of immigration news coverage also shifted.19 Fox News, which had the highest proportion of immigrant crime coverage of the three channels in our dataset, both increased its volume of immigration coverage and devoted more of its immigration coverage to crime. Figure 3 presents the proportion of daily coverage devoted to immigration and crime, and immigration and welfare for each of the three channels.20

In sum, cable news coverage became markedly more focused on immigrant crime and welfare during the 2016 campaign and especially after Trump’s inauguration. We treat cable news coverage as a proxy for media coverage more broadly, and the nature of this shift suggests that

19For results of the structural topic model, which shows changes in topic proportion over time, please see the appendix. 20 Daily crime coverage calculated as DailyCoverage jk = ∑i CrimeProportioni jk over all news segments i on channel j on date k

20 Figure 2: Immigration News Segments

Notes: There was a small increase in the monthly number of immigration segments aired by CNN and MSNBC after Trump’s inauguration while Fox News nearly doubled its’ immigration coverage.

21 Figure 3: Immigration News Segments by Topic Crime: Welfare:

Notes: The overall amount of immigration crime and immigration welfare coverage increased substantially in the post-inauguration period, especially on Fox News. Coverage of a topic is deﬁned as the sum of the crime/welfare topic proportion across all documents per month for that channel. the public had much greater exposure to messages about immigrant crime and welfare dependency after the 2016 campaign. In the next section, we examine web search patterns around immigration during the post-inauguration period.

Anti-Immigrant Searches Increased After the 2016 Election

There is a clear discontinuity in anti-immigrant search rates after the 2017 inauguration of Donald Trump. Figure 4 plots the proportion of Bing and Google searches for the three sets of immigration terms during the Trump, Obama, and Bush administrations.21 For each set of terms, there was a substantial increase in searches in the days after Trump’s inauguration. Fur- thermore, in addition to the immediate post-inauguration spike in searches, the baseline level

21Since Figures 2 and 3 show little change in media coverage of immigration campaign period relative to pre- campaign period and because we want to avoid the inclusion of a large number of periods in Figure 4, we let the campaign period be associated with the outgoing administration

22 Figure 4: Immigration Searches by Administration

Bing Google Crime Welfare Report

Notes: For all three sets of anti-immigrant terms, searches increased after Trump’s inauguration on both Bing and Google. There was also a spike in all three sets in the months immediately after the inauguration. For the Bing, data, the Y-axis is daily searches for the immigration terms as a percentage of daily Bing searches on a log10 scale, and each data point represents one day. For the Google data, the Y-axis is Google Trends’ “Interest over Time” measure, and each data point represents one month.

23 Table 6: Immigration Searches by Presidential Administration

Dependent variable: Crime Welfare Bing Google Bing Google Date 0.0004∗∗∗ (0.0001) 0.001 (0.001) 0.0004∗∗∗ (0.0001) −0.0002 (0.001) Bush Admin 8.280∗∗ (4.187) 7.819∗∗ (3.731) Obama Admin - - - - Trump Admin 0.304∗∗∗ (0.087) 19.903∗∗∗ (4.027) 0.265∗∗∗ (0.082) 18.560∗∗∗ (3.588) Constant −20.531∗∗∗ (1.965) −4.163 (23.241) −20.281∗∗∗ (1.855) 28.585 (20.710)

Report Bing Google Date −0.0004∗∗∗ (0.0001) −0.002∗ (0.001) Bush Admin 0.785 (2.899) Obama Admin - - Trump Admin 0.668∗∗∗ (0.051) 19.716∗∗∗ (2.788) Constant −7.649∗∗∗ (1.016) 72.574∗∗∗ (16.089)

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: After accounting for the linear time trend, there was still a clear increase in anti-immigrant searches after Trump’s inauguration. Trump admin is a dummy variable that is 1 when Date> 2017-01-20, and 0 otherwise. Bing regression is binomial logit on number of searches, with each search for an immigration term coded as 1, and a search for a non-immigration term coded as 0. Standard errors are clustered by date. Google regression is OLS on monthly search interest. of anti-immigrant searches either remained constant or increased over the course of the Trump administration. This difference is especially pronounced for reporting searches, where a sharp discontinuity at the inauguration was followed by relatively little change.

The increases in immigration searches after Trump’s inauguration were non-trivial and highly consistent across both the Google and Bing data. Compared with the Obama administration, searches for immigrant crime, welfare, and reporting during the Trump administration increased by a factor of 1.36 to 2.45. For example, the Google Trends data show that searches for the immigration and crime terms increased by a factor of 2.45 from the Obama administration to the Trump admin, while Bing shows that it increased by a factor of 1.8. Ta- ble 6 presents regression estimates for the effects of the administration change on immigration searches for both Google Trends and Bing and conﬁrms that the effects of the Trump adminis-

24 tration remain signiﬁcant even after introducing a linear time trend.

The results presented here establish the existence of a sizable and persistent increase in searches for immigration and crime, immigration and welfare, and methods for reporting immigrants to ICE after Trump’s inauguration. This difference manifests as a discontinuity immediately after the inauguration on January 20, 2017. Further, these results are highly consistent across both the Google and Bing search data. In the following section, we explore the link between immigration news coverage and anti-immigrant searches.

Anti-Immigrant Searches Closely Track Media Coverage

The level of anti-immigrant media coverage was found to be closely related to the level of anti-immigrant searches. Table 7 shows the relationship between media coverage and anti- immigrant searches. For all three types of anti-immigrant searches, the total daily number of immigration segments was strongly associated with the daily number of anti-immigrant searches.

The topic of the immigration coverage also influenced the number of searches. For immigrant crime, welfare, and reporting searches, immigration + crime coverage was significantly related to anti-immigrant search volume, even after controlling for the overall volume of immigration coverage. Immigration + welfare coverage was also significantly associated with increased immigration and welfare search volume on Bing but not on Google. On days when the news covered an anti-immigrant topic in detail, searches for that topic tended to be higher.

Furthermore, these associations remained signiﬁcant after controlling for the Trump administration. This suggests a genuine relationship between media coverage and anti-immigrant searches that is driven by speciﬁc events rather than by the overall higher level of immigration coverage during the Trump administration.

25 Table 7: Media Coverage and Anti-Immigrant Searches

Dependent variable: Crime Welfare Bing Google Bing Google Immigration Segs 0.0003 (0.001) 0.081∗∗∗ (0.031) 0.003∗∗∗ (0.001) 0.190∗∗∗ (0.025) Immigr + Crime Coverage 0.126∗∗∗ (0.013) 4.542∗∗∗ (0.451) 0.022∗∗ (0.010) 0.265 (0.366) Immigr + Welfare Coverage 0.025 (0.028) −0.672 (1.169) 0.082∗∗ (0.035) 0.123 (0.950) Trump Admin −0.018 (0.086) 20.149∗∗∗ (3.045) 0.039 (0.060) 6.113∗∗ (2.474) Date 0.001∗∗∗ (0.0001) −0.007∗∗ (0.003) 0.0005∗∗∗ (0.0001) 0.007∗∗∗ (0.003) Day of Week FE X X X X Month FE X X X X Constant −23.421∗∗∗ (1.404) 142.177∗∗∗ (51.822) −20.998∗∗∗ (1.221) −89.987∗∗ (42.100) ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Dependent variable: Report Bing Google Immigration Segs 0.002∗∗∗ (0.0004) 0.055∗∗∗ (0.019) Immigr + Crime Coverage 0.019∗∗∗ (0.007) 0.685∗∗ (0.279) Immigr + Welfare Coverage −0.007 (0.014) 1.416∗ (0.723) Trump Admin 0.563∗∗∗ (0.043) 18.051∗∗∗ (1.883) Date −0.0004∗∗∗ (0.00005) −0.002 (0.002) Day of Week FE X X Month FE X X Constant −9.027∗∗∗ (0.842) 73.911∗∗ (32.050)

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: The volume of immigration crime coverage and overall immigration coverage was positively associated with higher rates of search for immigration crime, welfare, and reporting. The volume of immigration welfare coverage is positively associated with immigration and welfare searches. Bing Regression is binomial logit with standard errors clustered by date. OLS yields substantively similar results (see appendix). Google Regression is OLS on daily Google data.

26 Table 8: Immigration Coverage and Searches during the Obama Administration

Dependent variable: Crime Welfare Bing Google Bing Google Immigration Segs −0.005∗ (0.003) −0.120∗∗∗ (0.037) 0.001 (0.001) −0.002 (0.040) Immigr + Crime Coverage 0.175∗∗∗ (0.066) 5.039∗∗∗ (0.440) 0.021 (0.019) 0.395 (0.476) Immigr + Welfare Coverage 0.065 (0.092) −0.755 (1.705) 0.101∗ (0.052) 1.997 (1.846) Date 0.007 (0.005) 0.011∗∗∗ (0.003) −0.003∗ (0.002) 0.008∗∗ (0.004) Day of Week FE X X X X Month FE X X X X Constant −139.058 (87.117) −165.878∗∗∗ (56.005) 45.069 (33.080) −108.028∗ (60.622)

Dependent variable: Report Bing Google Immigration Segs −0.0005 (0.001) 0.010 (0.031) Immigr + Crime Coverage 0.017 (0.021) 0.354 (0.373) Immigr + Welfare Coverage 0.037 (0.060) 1.817 (1.444) Date −0.001 (0.002) 0.007∗∗ (0.003) Day of Week FE X X Month FE X X Constant 4.580 (40.323) −74.436 (47.441)

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: The relationship of anti-immigrant media coverage with anti-immigrant searches for welfare and crime was similar during both the Obama and Trump administrations. Coefﬁcients on reporting searches are not signiﬁcant but still in the expected direction. Regressions were run only for dates during the Obama administration (03/01/2016-01/19/2017 for Bing and 01/01/2015-01/19/2017 for Google). For an equivalent table with data for the Trump administration, please see appendix.

The strong and consistent relationship between crime coverage and reporting searches across both Bing and Google is worth underscoring. Linking immigration to crime has been a major messaging strategy of the Trump administration in justifying their restrictive immigration policies, and these results suggest that this message is effective at mobilizing anti- immigrant behavior.

While there was a signiﬁcant increase in negative coverage of immigrants after Trump’s inauguration, the Trump administration also engaged in substantial anti-immigrant messaging and policy initiatives. This raises the issue of whether media coverage of the Trump administration’s anti-immigrant messages sparked the increase in anti-immigrant searches. If this was

27 Table 9: Immigration Coverage and Immigration Searches during the Trump Administration

Dependent variable: Crime Welfare Bing Google Bing Google Immigration Segs 0.001 (0.002) −0.156∗∗ (0.066) 0.001 (0.001) 0.044 (0.053) Immigr + Trump Segs −0.006∗∗ (0.003) 0.142 (0.108) 0.0001 (0.002) −0.044 (0.088) Immigr + Crime Coverage 0.125∗∗∗ (0.013) 5.103∗∗∗ (0.455) 0.026∗∗ (0.010) 0.813∗∗ (0.368) Immigr + Welfare Coverage 0.022 (0.027) −0.998 (1.171) 0.081∗∗ (0.036) −0.479 (0.947) Trump Admin −0.192∗∗ (0.083) 13.625∗∗∗ (3.216) −0.046 (0.061) −0.717 (2.600) Date 0.001∗∗∗ (0.0001) −0.006∗ (0.003) 0.001∗∗∗ (0.0001) 0.009∗∗∗ (0.002) Trump Segs x Trump Admin 0.007∗∗ (0.003) 0.723∗∗∗ (0.115) 0.006∗∗∗ (0.002) 0.673∗∗∗ (0.093) Day of Week FE X X X X Month FE X X X X Constant −24.094∗∗∗ (1.476) 120.569∗∗ (51.437) −21.691∗∗∗ (1.193) −114.767∗∗∗ (41.580)

Dependent variable: Report Bing Google Immigration Segs −0.001 (0.001) −0.058 (0.041) Immigr + Trump Segs 0.001 (0.002) 0.094 (0.068) Immigr + Crime Coverage 0.023∗∗∗ (0.007) 0.913∗∗∗ (0.283) Immigr + Welfare Coverage −0.007 (0.013) 1.340∗ (0.729) Trump Admin 0.469∗∗∗ (0.044) 15.500∗∗∗ (2.002) Date −0.0003∗∗∗ (0.00005) −0.001 (0.002) Trump Segs x Trump Admin 0.008∗∗∗ (0.002) 0.300∗∗∗ (0.071) Day of Week FE X X Month FE X X Constant −9.760∗∗∗ (0.829) 65.889∗∗ (32.018)

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: Controlling for the amount of Trump-immigration coverage does not weaken the relationship between anti-immigrant coverage and anti-immigrant searches. The Trump News variable is not signiﬁcant, indicating that during the campaign, Trump’s statements on immigration have no relationship with anti-immigrant searches. The interaction between Trump News and Trump Admin, while signiﬁcant, does not weaken the relationship between the crime/welfare news coverage and anti-immigrant searches. Bing regression is binomial logit with standard errors clustered by day, and Google regression is OLS.

28 the case, the ﬁndings would be the result not of media content but of the anti-immigrant cues emanating directly from the Trump administration.

To assess this hypothesis, we use two tests. The first test examines the relationship between media content and anti-immigrant searches prior to Trump’s inauguration. One one hand, if the relationship between media coverage and anti-immigrant searches is the result of the Trump administration’s anti-immigrant cues, there should not be a relationship between coverage and searches. On the other hand, if media content, rather than presidential cues, is responsible for anti-immigrant searches, there should be a significant relationship between anti- immigrant coverage and searches, during both administrations. Table 8, which is the same regression from Table 7 but confined to the Obama administration, shows that, in line with the media content hypothesis, there was a significant relationship between anti-immigrant news coverage and anti-immigrant searches prior to Trump’s inauguration.

While the first test eliminates the possibility of search spikes as the result of actions of the Trump administration, the possibility remains that searchers were responding to candidate Trump’s anti-immigrant cues. To test this hypothesis, we implement a second test, which looks at the number of mentions of the word “Trump” in daily immigration coverage. When the Trump campaign or Trump administration emits anti-immigrant cues, there is often significant media coverage of the event and thus an increase in immigration segments containing the word “Trump”. If the relationship between immigration coverage and anti-immigrant web searches is the direct result of Trump’s words or actions, searches should be related to the amount of immigration coverage containing the word “Trump” rather than overall immigration coverage or coverage of a specific immigration topic. To capture differences in cue-taking during the campaign versus the Trump administration, we also include an interaction between number of segments of immigration-Trump coverage and Trump administration variable.

The results of these tests, shown in Table 9, demonstrate that the effects of news coverage on immigration searches are not the result of Trump’s cues on immigration. Even when

29 the Trump Coverage variable and its interaction with the Trump administration is included in the regression, the strong relationship between anti-immigrant searches and coverage of the immigrant crime/welfare topic does not decline in significance. In the case of Trump’s campaign cues, the regressions did not show a significant positive effect of Trump immigration coverage on anti-immigrant searches during the campaign and only a modest effect after Trump’s inauguration. The results of the tests indicate that the close relationship between anti- immigrant news coverage and anti-immigrant search behavior cannot be attributed to cues from Trump during the campaign or his administration. The relationship between anti-immigrant media coverage and anti-immigrant web searches was found to be strong during the Obama administration (as shown in Table 8), and there is no significant relationship between the volume of Trump mentions in immigration news segments and the volume of anti-immigrant web searches during the campaign period (as shown in Table 9). While there is a relationship between Trump immigration volume and anti-immigrant web searches during the Trump administration, this does not diminish the relationship between anti-immigrant media coverage and anti-immigrant searches.

In summary, when comparing immigration coverage with immigration searches, after controlling for the presidential administration, on days with more news coverage of immigration, there were more anti-immigrant Bing and Google searches. Furthermore, the topic of the news coverage had a signiﬁcant impact on searches; on days with more immigration and crime coverage, there were more anti-immigrant searches while days with more economic impact coverage had fewer. This relationship is not attributable to media coverage of Trump’s anti-immigrant cues. In the next section, we isolate the causal effect of anti-immigrant new segments by examining searches during Trump’s televised speeches about immigration.

30 Anti-Immigrant Searches Increase During Trump Broadcasts

The volume of anti-immigrant searches spiked during televised broadcasts of speeches by Trump. Figure 5 shows hourly Google searches for crime, welfare, and reporting during Trump’s 2017, 2018, and 2019 SOTUs and 2019 Oval Office address and Obama’s 2015 and 2016 SO- TUs. The plot compares these searches with searches for the same terms one week before the speech and one week after the speech. During Trump’s speeches, there is a clear spike for crime, welfare, and reporting, but no corresponding spike during Obama’s speeches. There is no increase in searches at the same time of day on the seventh day before and after the speech, which suggests that the increase in searches can be attributed to the speech itself rather than any time effects. Table 10 confirms that these spikes in searches are statistically significant.

The content of speeches affected which terms had the largest spikes in searches. While Obama mentioned immigrants in his 2015 and 2016 SOTU addresses, these mentions were positive or neutral, and did not imply that immigrants were more likely to commit crimes or rely disproportionately on the social safety net. On the other hand, in all four of Trump’s considered televised addresses, he provided examples of immigrant criminality as an argument for restrictive immigration policy. However, he talked about immigrants and the social safety net in only three of his four speeches. The speech in which he did not discuss immigrants’ use of welfare services was the only one for which there was no spike in immigration and welfare searches. Table 11 summarizes the speech content and search spikes for Obama’s and Trump’s speeches.

In summary, there was a clear and sharp increase in anti-immigrant web searches during Trump’s speeches but not during Obama’s speeches. This finding strongly supports the conclusion that Trump’s speeches had a causal effect on anti-immigrant searches, including reporting searches. These findings strongly support our finding that there is a relationship between anti- immigrant media coverage and anti-immigrant web searches.

31 Figure 5: Google Immigration Searches During Presidential Speeches Crime: Welfare:

Report: Notes: There was a clear spike in anti- immigrant searches during Trump’s televised addresses but no similar spike for Obama. Each point represents Google Trends data for one hour. The dashed lines represent speech times. The Y-axes of the plots are comparable within, but not between, days. A rating of 100 represents the hour with the most searches on that speciﬁc date but is not comparable with searches for other dates. For Bing results, please see appendix

32 Table 10: Google Immigration Searches During Presidential Speeches

Dependent variable: Crime Welfare Obama Trump Obama Trump Speech date −2.920 (2.797) −35.689∗∗∗ (2.774) −0.705 (3.449) −0.992 (2.769) Speech hour 2.148 (5.594) 8.439 (5.547) 11.068 (6.898) 6.106 (5.537) Speech date x Speech hour 5.420 (9.689) 68.273∗∗∗ (9.608) −9.545 (11.948) 24.242∗∗ (9.591) Constant 16.602∗∗∗ (1.615) 43.811∗∗∗ (1.601) 19.682∗∗∗ (1.991) 37.477∗∗∗ (1.598) Observations 144 216 144 216

Dependent variable: Report Obama Trump Speech Date −1.875 (4.298) −11.848∗∗∗ (2.337) Speech Hour 1.920 (8.597) −1.447 (4.674) Speech Date x Speech Hour 7.875 (14.890) 22.932∗∗∗ (8.096) Constant 37.830∗∗∗ (2.482) 59.197∗∗∗ (1.349) Observations 144 216 ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: There was a signiﬁcant spike in Google searches for immigrant crime, welfare, and reporting during Trump’s speeches. Regression is OLS. The Speech date variable is 1 if the search was performed on the date of the speech and is 0 otherwise (i.e., performed on the seventh dat before or after the speech). The speech hour variable is 1 if the search was performed during the hours of 9 pm or 10pm EST (with the exception of the Oval Ofﬁce speech, which was only 9pm EST).

Table 11: Speech Content and Search Types

Speech Mentions Search Spike

Speech Viewers # Immigr Mentionsa Immigrant Crime Immigrant Welfare Crime Welfare Report

Obama 2015 SOTU 32M 4 N N Obama 2016 SOTU 31.3M 5 N Nb Trump 2017 SOTU 48M 11 Y Y X X Trump 2018 SOTU 45.6M 11 Y Nc X Trump 2019 Address 41M 8 Y Y X X X Trump 2019 SOTU 46M 9 Y Y X X X Notes: Variance in Trump’s speech content corresponds with variance in search spikes. Search spikes correspond with the anti-immigrant content in the speech. X denotes a statistically signiﬁcant search spike at p<0.05.

aNumber of mentions of “immigr”, “illegals, or “illegal alien” during the speech. bObama stated “Immigrants aren’t the principal reason wages haven’t gone up.” cTrump described a merit-based immigration program that would admit “people who are skilled, who want to work, who will contribute to our society” but stopped short of claiming that unauthorized immigrants are costly or a burden on public services, as he did in the 2017 SOTU, 2019 SOTU, and 2019 Oval Ofﬁce Address.

33 Conclusion

Does anti-immigrant media coverage lead to anti-immigrant behavior? Our evidence suggests that it does. Using Bing and Google search data, we measured information-seeking behavior around immigrant crime and welfare as well as searches seeking to report immigrants to ICE. We examined both daily media coverage of immigration over time as well as televised presidential speeches, which are major media events watched by tens of millions of Americans. We ﬁnd that both anti-immigrant media coverage and anti-immigrant web searches increased after Trump was inaugurated. The volume of coverage on immigrants and crime as well immigrants and welfare increased substantially, driven mostly by an increase in coverage by Fox News. Searches for immigrant crime, welfare, and reporting show a similar pattern, with a sharp discontinuity immediately after the inauguration.

Media coverage played a signiﬁcant role in the increase in searches. Anti-immigrant searches closely tracked anti-immigrant media coverage, even after controlling for immigration cues from Donald Trump. For both Bing and Google data, daily search volume for reporting immigrants was found to be highly associated with the daily volume of immigrant crime coverage. This is especially notable given the consistent use of the immigrant crime narrative as a rhetor- ical tool by the Trump administration.

We ﬁnd that searches for immigrant crime, welfare, and reporting spiked during Trump’s televised speeches, suggesting a causal relationship between anti-immigrant media coverage and anti-immigrant searches. There was no spike in anti-immigrant searches during Obama’s televised speeches, suggesting the increase in searches is conditional on media messages con- veying anti-immigrant sentiment. Furthermore, the content of the speeches matters. Trump talked about immigrants burdening the social safety net in three of his four televised speeches, and searches for immigrants and welfare spiked only during those three speeches.

These ﬁndings have serious implications for media coverage of immigration. First, anti-

34 immigrant media coverage, especially coverage of immigrants and crime, has tangible and serious negative impacts on immigrants’ well-being. A media story that repeats inaccurate claims about immigrants and crime could motivate a viewer to report a (presumed) undocumented immigrant to ICE, resulting in that person’s arrest and deportation. Immigrants have reported “living in fear”22 under the Trump administration, and anti-immigrant news coverage may make these fears come to fruition.

The second implication concerns news coverage of Trump’s speeches and other statements. While there was signiﬁcant media debate over airing Trump’s 2019 Oval Ofﬁce Ad- dress live, networks chose to air the address. After the broadcast, news outlets called the address a “dud”23 and “bewildering”24, concluding that, unless the address motivated Ameri- cans to call their congressmembers, “Trump’s speech changed nothing”.25 Even if Trump’s address did not have substantial political impacts or change the hearts and minds of the population at-large, it resulted in one of the largest ever Google search spikes from reporting immigrants, suggesting that at least some reports of immigrants to ICE were made as a result of the speech. Accordingly, airing anti-immigrant statements to a large audience potenially causes harm to immigrants.

References

Abrajano, M. and Z. L. Hajnal (2017). White backlash: Immigration, race, and American politics. Princeton University Press.

Abrajano, M. A., Z. Hajnal, and H. J. Hassell (2017). Media framing and partisan identity: The case of immigration coverage and white macropartisanship. Journal of Race, Ethnicity and Politics 2(1), 5–34.

22https://www.theguardian.com/us-news/2018/oct/18/immigration-ice-deportation-undocumented-trump 23https://www.thedailybeast.com/trumps-oval-ofﬁce-speech-was-a-dud-but-it-could-be-a-fundraising-success 24https://www.theatlantic.com/politics/archive/2019/01/president-trumps-speech-shutdown-and-wall/579844/ 25https://www.washingtonpost.com/outlook/2019/01/09/did-trumps-oval-ofﬁce-address-change-anything/

35 Achen, C. H. and L. M. Bartels (2017). Democracy for realists: Why elections do not produce responsive government, Volume 4. Princeton University Press.

Allen, B., E. M. Cisneros, and A. Tellez (2015). The children left behind: The impact of parental deportation on mental health. Journal of Child and Family Studies 24(2), 386–392.

Allen, J., B. Howland, M. Mobius, D. Rothschild, and D. J. Watts (2020). Evaluating the fake news problem at the scale of the information ecosystem. Science Advances 6(14), eaay3539.

Ansolabehere, S. and E. Hersh (2012). Validation: What big data reveal about survey misre- porting and the real electorate. Political Analysis 20(4), 437–459.

Bhavnani, R. R. and G. Nellis (2016). Can government-controlled media cause social change? television and fertility in india. University of Wisconsin.

Bordino, I., S. Battiston, G. Caldarelli, M. Cristelli, A. Ukkonen, and I. Weber (2012). Web search queries can predict stock market volumes. PloS one 7(7).

Brader, T., N. A. Valentino, and E. Suhay (2008). What triggers public opposition to immigration? anxiety, group cues, and immigration threat. American Journal of Political Sci- ence 52(4), 959–978.

Choi, H. and H. Varian (2012). Predicting the present with google trends. Economic record 88, 2–9.

Chong, D. and J. N. Druckman (2007). Framing public opinion in competitive democracies. American Political Science Review 101(4), 637–655.

Claibourn, M. P. (2008). Making a connection: Repetition and priming in presidential cam- paigns. The Journal of Politics 70(4), 1142–1159.

Dalton, R. J., P. A. Beck, and R. Huckfeldt (1998). Partisan cues and the media: Information ﬂows in the 1992 presidential election. American Political Science Review 92(1), 111–126.

36 DâAmuri, F. and J. Marcucci (2017). The predictive power of google searches in forecasting us unemployment. International Journal of Forecasting 33(4), 801–816.

Ettredge, M., J. Gerdes, and G. Karuga (2005). Using web-based search data to predict macroeconomic statistics. Communications of the ACM 48(11), 87–92.

Farris, E. M. and H. Silber Mohamed (2018). Picturing immigration: how the media criminal- izes immigrants. Politics, Groups, and Identities 6(4), 814–824.

Gadarian, S. K. and B. Albertson (2014). Anxiety, immigration, and the search for information. Political Psychology 35(2), 133–164.

Gawlik, E., H. Kabaria, and S. Kaur (2011). Predicting tourism trends with google insights. Accessed December 1, 2012.

Gemmill, A., R. Catalano, J. A. Casey, D. Karasek, H. E. Alcalá, H. Elser, and J. M. Torres (2019). Association of preterm births among us latina women with the 2016 presidential election. JAMA network open 2(7), e197084–e197084.

Goel, S., J. M. Hofman, S. Lahaie, D. M. Pennock, and D. J. Watts (2010). Predicting con- sumer behavior with web search. Proceedings of the National academy of sciences 107(41), 17486–17490.

Iyengar, S., Y. Lelkes, M. Levendusky, N. Malhotra, and S. J. Westwood (2019). The origins and consequences of affective polarization in the united states. Annual Review of Political Science 22, 129–146.

Jardina, A. (2019). White identity politics. Cambridge University Press.

Jonas, E., S. Schulz-Hardt, D. Frey, and N. Thelen (2001). Conﬁrmation bias in sequential information search after preliminary decisions: an expansion of dissonance theoretical research on selective exposure to information. Journal of personality and social psychology 80(4), 557.

37 King, G., B. Schneer, and A. White (2017). How the news media activate public expression and inﬂuence national agendas. Science 358(6364), 776–780.

Knobloch-Westerwick, S., B. K. Johnson, and A. Westerwick (2014). Conﬁrmation bias in online searches: Impacts of selective exposure before an election on political attitude strength and shifts. Journal of Computer-Mediated Communication 20(2), 171–187.

Knoll, B. R., D. P. Redlawsk, and H. Sanborn (2011). Framing labels and immigration policy attitudes in the iowa caucuses:âtrying to out-tancredo tancredoâ. Political Behavior 33(3), 433–454.

Konitzer, T. et al. (2017). Non-representative surveys: Modes, dynamics, party, and likely voter space. Technical report, Working Paper.

Krupenkin, M., D. Rothschild, S. Hill, and E. Yom-Tov (2019). President trump stress disor- der: partisanship, ethnicity, and expressive reporting of mental distress after the 2016 election. Sage open 9(1), 2158244019830865.

Lahav, G. and M. Courtemanche (2012). The ideological effects of framing threat on immigration and civil liberties. Political Behavior 34(3), 477–505.

Layman, G. C. and T. M. Carsey (2002). Party polarization and party structuring of policy attitudes: A comparison of three nes panel studies. Political Behavior 24(3), 199–236.

Lee, T. (2002). Mobilizing public opinion: Black insurgency and racial attitudes in the civil rights era. University of Chicago Press.

Lenz, G. S. (2013). Follow the leader?: how voters respond to politicians’ policies and perfor- mance. University of Chicago Press.

Levendusky, M. S. (2009). The microfoundations of mass polarization. Political Analy- sis 17(2), 162–176.

38 Ndulue, E., F. Bermejo, K. Ramos, S. E. Loewe, N. Hoffman, and E. Zuckerman (2019). The language of immigration reporting: Normalizing vs. watchdogging in a nativist age. Techni- cal report.

Ofran, Y., O. Paltiel, D. Pelleg, J. M. Rowe, and E. Yom-Tov (2012). Patterns of information- seeking for cancer on the internet: an analysis of real world data. PloS one 7(9).

Pennycook, G., T. D. Cannon, and D. G. Rand (2018). Prior exposure increases perceived accuracy of fake news. Journal of experimental psychology: general.

Peterson, E., S. Goel, and S. Iyengar (2019). Partisan selective exposure in online news consumption: evidence from the 2016 presidential campaign. Political Science Research and Methods, 1–17.

Roberts, M. E., B. M. Stewart, D. Tingley, C. Lucas, J. Leder-Luis, S. K. Gadarian, B. Al- bertson, and D. G. Rand (2014). Structural topic models for open-ended survey responses. American Journal of Political Science 58(4), 1064–1082.

Silber Mohamed, H. and E. M. Farris (2019). âbad hombresâ? an examination of identities in us media coverage of immigration. Journal of Ethnic and Migration Studies, 1–19.

Stephens-Davidowitz, S. and A. Pabon (2017). Everybody lies: Big data, new data, and what the internet can tell us about who we really are. HarperCollins New York.

Street, A., T. A. Murray, J. Blitzer, and R. S. Patel (2015). Estimating voter registration dead- line effects with web search data. Political Analysis 23(2), 225–241.

Strömberg, D. (2015). Media and politics. economics 7(1), 173–205.

Tesler, M. and D. O. Sears (2010). Obama’s race: The 2008 election and the dream of a post- racial America. University of Chicago Press.

Torres, J. M., J. Deardorff, R. B. Gunier, K. G. Harley, A. Alkon, K. Kogut, and B. Eskenazi (2018). Worry about deportation and cardiovascular disease risk factors among adult women:

39 The center for the health assessment of mothers and children of salinas study. Annals of Behavioral Medicine 52(2), 186–193.

Wang, H., D. Zhang, and C. Zhai (2011). Structural topic model for latent topical structure analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 1526–1535.

White, R. (2013). Beliefs and biases in web search. In Proceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval, pp. 3–12. ACM.

Yanagizawa-Drott, D. (2014). Propaganda and conﬂict: Evidence from the rwandan genocide. The Quarterly Journal of Economics 129(4), 1947–1994.

40 Appendix

Contents

A.1 Methods ...... 42

A.1.1 List of Removed Bing Terms ...... 42

A.2 Robustness to Model Speciﬁcation ...... 44

A.2.1 Results with OLS ...... 44

A.2.2 Media and Search During Trump Administration ...... 46

A.3 Additional Analysis ...... 47

A.3.1 STM Results Over Time ...... 47

A.3.2 HSI Search Results ...... 48

A.3.3 Bing Speech Results ...... 51

A.3.4 Weather Placebo Results ...... 53

41 A.1 Methods

Here, we provide some additional details about the search data.

Google Trends only displays searches at the daily (rather than weekly or monthly) level for a few months (it provides weekly data for ﬁve-year intervals). In order to generate a time series of daily Google searches, we used monthly and weekly data to normalize the searches to a single scale. The procedure was as follows: First, for each month, we took the mean weekly Google trends value (mean search volume across the four weeks of each month). We used this monthly search average to calculate an adjustment value for each month, ad j = mean weekly search . We then mul- tiplied each week by this adjustment value to get an adjusted weekly score for each week. We then repeated the process with the adjusted weekly score and the daily score to get a ﬁnal time series of adjusted daily scores.

This daily Google time series was moderately correlated with our daily Bing data from 2016 to 2019, with a correlation of 0.39 for crime, 0.51 for welfare, 0.32 for report, and 0.55 for weather. These correlations are especially notable given our limitations in generating Google Search strings, as compared to Bing. However, daily data in the early part of the Google daily time series (2004-2010) was extremely noisy, so we chose to use monthly data for our comparisons of searches across presidential administrations.

Increases in anti-immigrant searches were highly consistent across Google and Bing, as shown in Table A1.

A.1.1 List of Removed Bing Terms

These are searches that were identified as irrelevant or botted. Searches in italics were identified as "botted". A search is identified as botted if it had a one-day spike of at least 2 orders of magnitude above the previous day and there was no similar spike in closely related searches

42 Table A1: Increase in anti-immigrant searches Category Google Bing Crime 2.45x 1.80x Welfare 1.71x 1.72x Report 1.36x 1.51x Report (HSI) − 1.41x

Notes: Bing and Google data showed similar increases in immigration searches during the Trump administration. Table displays the proportionate increase in searches for each category from the Obama administration to the Trump admin.

(eg searches containing the same keywords, but in a different order). For example, "california immigrants health benefits" had a extreme spike, but "ca immigrant health benefits" and "immigrant health benefits california" did not. These searches were removed from the Bing data, including all searches that included the following strings (case-insensitive): Crime: skill, deer were killed by predators, sweden immigrants crime, ICE immigrant- crime hotline calls, criminal justice reform, ’theodore roosevelt on race, riots, immigration, and crime book’ Welfare: low cost immigration law services, immigration benefits, costa rica immigration, immigrant costumes, benefits of illegal immigration, process and cost for immigrants coming to usa, benefits of immigrants, low cost immigration law services orange county ca, benefits of immigration, California immigrants health benefits Report: immigration commission report, ’: Report’, us news and world report, crime report, the parents report, summarizing a report, IG Report, MSNBC Report, criminal illegal alien report, Report On, sample report, thai immigration 90 day report, report from the Gov- ernment Accounting Office, intelligence report, responds to immigration report, center for immigration studies report

43 Table A2: Table 6 OLS

Dependent variable:

Crime Welfare Report Date 0.00000∗∗∗ (0.00000) 0.00000∗∗∗ (0.00000) −0.00000∗∗∗ (0.000) Obama Admin − − − Trump Admin 0.0001∗∗ (0.00002) 0.00004∗∗ (0.00002) 0.00003∗∗∗ (0.00000) Constant −0.002∗∗∗ (0.001) −0.002∗∗∗ (0.0005) 0.0004∗∗∗ (0.0001)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: The results of this table replicate those of table 6, which show that Bing anti-immigrant searches were higher during the Trump administration.

Table A3: Table 7 OLS

Dependent variable:

Crime Welfare Report Immigration Segments −9e-09 (6e-09) 8.7e-09∗∗∗ (3.1e-09) 1.3e-09∗∗∗ (3e-10) Immigr + Crime Coverage 8.21e-07∗∗∗ (1.45e-07) 9.09e-08∗∗ (4.61e-08) 1.1e-08∗∗ (4.9e-09) Immigr + Welfare Coverage 1.78e-07 (1.39e-07) 3.32e-07∗ (1.73e-07) −3.9e-09 (7.5e-09) Trump Admin −3.7e-07∗ (1.92e-07) −6.3e-09 (1.33e-07) 2.5e-07∗∗∗ (2.1e-08) Date 1.7e-09∗∗∗ (3e-10) 1e-09∗∗∗ (2e-10) −2e-10∗∗∗ (0e+00) Day of Week FE X X X Month FE X X X Constant −2.64e-05∗∗∗ (4.31e-06) −1.48e-05∗∗∗ (3.13e-06) 3.32e-06∗∗∗ (4.55e-07)

Note: ∗p<0.1; ∗∗p<5e-02; ∗∗∗p<1e-02

Notes: The results of this table replicate those of table 7, which show that Bing anti-immigrant searches were higher on days with higher coverage of immigration + crime and/or immigration + welfare.

A.2 Robustness to Model Speciﬁcation

A.2.1 Results with OLS

In this section, we show that our Bing results from tables 6-9 are substantively identical whether we use OLS (this section) or logistic regression (body of the paper). In the case of OLS, the dependent variable is the proportion of all Bing searches for a particular set of terms (eg immigration and crime). In both cases, standard errors are clustered by date.

44 Table A4: Table 8 OLS

Dependent variable:

Crime Welfare Report Immigration Segs −0e+00 (0e+00) 0e+00 (0e+00) −0e+00 (0e+00) Immigr + Crime Coverage 4e-07∗∗ (2e-07) 1e-07∗ (3e-08) 0e+00 (0e+00) Immigr + Welfare Coverage 3e-07 (2e-07) 2e-07∗∗ (1e-07) 0e+00 (2e-08) Date 0e+00 (0e+00) −0e+00∗∗ (0e+00) −0e+00 (0e+00) Day of Week FE X X X Month FE X X X Constant −1.78e-04 (1.73e-04) 1.06e-04∗∗ (5.28e-05) 6.6e-06 (1.42e-05)

Note: ∗p<0.1; ∗∗p<5e-02; ∗∗∗p<1e-02

Notes:The results of this table replicate those of table 8, which show that the relationship between anti-immigrant searches and anti-immigrant coverage persisted throughout the Obama administration.

Table A5: Table 9 OLS

Dependent variable:

Crime Welfare Report Immigration Segs −0e+00 (0e+00) 0e+00 (0e+00) −0e+00 (0e+00) Immigr + Trump Segs −3e-08∗∗∗ (0e+00) −0e+00 (0e+00) 0e+00 (0e+00) Immigr + Crime Coverage 8e-07∗∗∗ (1e-07) 1e-07∗∗ (5e-08) 2e-08∗∗∗ (0e+00) Immigr + Welfare Coverage 1e-07 (1e-07) 3e-07∗ (2e-07) −0e+00 (0e+00) Trump Admin −9e-07∗∗∗ (3e-07) −3e-07∗∗ (1e-07) 2e-07∗∗∗ (2e-08) Date 0e+00∗∗∗ (0e+00) 0e+00∗∗∗ (0e+00) −0e+00∗∗∗ (0e+00) Trump Segs x Trump Admin 4e-08∗∗ (2e-08) 3e-08∗∗∗ (0e+00) 0e+00∗∗∗ (0e+00) Day of Week FE X X X Month FE X X X Constant −2.82e-05∗∗∗ (4.5e-06) −1.67e-05∗∗∗ (3e-06) 2.8e-06∗∗∗ (4e-07)

Note: ∗p<0.1; ∗∗p<5e-02; ∗∗∗p<1e-02

Notes: The results of this table replicate those of table 9, which show that the relationship between anti-immigrant searches and anti-immigrant coverage persisted even after controlling for news coverage of Trump’s statements and actions on immigration.

45 Table A6: Trump Admin Media and Search

Crime Welfare Bing Google Bing Google Immigration Segs 0.001 (0.001) 0.117∗∗∗ (0.044) 0.003∗∗∗ (0.001) 0.244∗∗∗ (0.033) Immigr + Crime Coverage 0.122∗∗∗ (0.012) 3.790∗∗∗ (0.712) 0.022∗∗ (0.011) 0.571 (0.545) Immigr + Welfare Coverage 0.013 (0.029) 0.186 (1.543) 0.077∗∗ (0.035) −0.811 (1.181) Date 0.001∗∗∗ (0.0001) −0.017∗∗∗ (0.004) 0.001∗∗∗ (0.0001) 0.007∗ (0.003) Day of Week FE X X X X Month FE X X X X Constant −24.874∗∗∗ (1.507) 354.439∗∗∗ (77.486) −21.828∗∗∗ (1.305) −75.740 (59.298)

Report Bing Google Immigration Segs 0.003∗∗∗ (0.0004) 0.050∗ (0.025) Immigr + Crime Coverage 0.021∗∗∗ (0.008) 0.824∗∗ (0.413) Immigr + Welfare Coverage −0.011 (0.014) 1.662∗ (0.896) Date −0.0004∗∗∗ (0.00005) −0.007∗∗∗ (0.003) Day of Week FE X X Month FE X X Constant −7.978∗∗∗ (0.868) 190.799∗∗∗ (45.006) Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: We see a strong relationship between anti-immigrant news coverage and anti-immigrant searches during the Trump administration.

A.2.2 Media and Search During Trump Administration

Here, we replicate Table 8 using only data from the Trump administration, in order to demonstrate that the results presented in the body of the paper are not driven solely by the Obama administration results. Here, we continue to see a strong correlation between anti-immigrant media coverage and anti-immigrant searches.

46 A.3 Additional Analysis

A.3.1 STM Results Over Time

Here, we present some more information about our topic model. In the body of the paper (Fig 3), we show the overall amount of daily news coverage about immigrant crime and immigrant

welfare, which is the sum of the daily topic proportions, DailyCoverage jk = ∑i CrimeProportioni jk over all news segments i on channel j on date k. This measure captures the overall volume of ant-immigrant coverage. Here, we present regressions that estimate the effect of the different time periods on the daily mean of topic proportions, which covers the proportion of immigration coverage that is of a particular topic.

Table A7 presents the estimates for topic prevalence by time period and by channel. Dur- ing and after the campaign, there was a signiﬁcant increase in coverage of immigration and crime. On the other hand, the overall proportion of immigration coverage about welfare declined during this period, likely because anti-immigrant programs signiﬁcantly increased their messaging about crime.

47 Table A7: STM Results - Immigration Coverage by Topic

Dependent variable: Topic Proportion Crime (1) Crime (3) Welfare (13) CNN - - - Fox 0.008∗∗∗ (0.002) 0.010∗∗∗ (0.002) 0.014∗∗∗ (0.002) MSNBC −0.002 (0.003) −0.001 (0.003) 0.003∗∗∗ (0.002)

Campaign 0.026∗∗∗ (0.002) 0.005∗∗∗ (0.002) −0.006∗∗∗ (0.002) Campaign x Fox 0.008∗∗∗ (0.003) 0.003∗∗∗ (0.003) −0.004∗∗∗ (0.003) Campaign x MSNBC −0.010∗∗∗ (0.003) −0.0002∗∗∗ (0.003) −0.001∗∗∗ (0.003)

Post-Inaug 0.016∗∗∗ (0.002) −0.002∗∗∗ (0.002) −0.002∗∗ (0.002) Post-Inaug x Fox 0.008∗∗∗ (0.002) 0.005∗∗∗ (0.002) −0.003∗∗ (0.002) Post-Inaug x MSNBC −0.001 (0.003) −0.001 (0.003) −0.002 (0.003)

Constant 0.006∗∗∗ (0.002) 0.020∗∗∗ (0.002) 0.020∗∗∗ (0.002) Observations 61,213 61,213 61,213 ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: Immigration and crime coverage increased signiﬁcantly as a proportion of all immigration coverage during both the campaign and the post-inauguration period. Immigration and welfare coverage, on the other hand, declined as a proportion of overall immigration coverage.

A.3.2 HSI Search Results

We have two different measures of “reporting” searches on Bing. These two measures represent a precision versus recall tradeoff. The ﬁrst measure is described in the body of the paper. This measure very precisely measures intent, but it misses a very large portion of people who are interested in reporting immigrants.26

To understand which searches lead to immigrant reporting behavior, we looked at search queries that resulted in at least one click to the HSI tip form page.27 HSI terms were much more popular than direct “report” terms. On the average day, there were 1465% as many HSI

26Only 34% of clicks to the HSI tip form came from a reporting search 27There are two ways to report immigrants to ICE: the HSI phone tipline, and the HSI online tip form. Clicks directly to the online tip form from the search results are relatively rare, since the form is not clearly labelled in the search results as the form to report immigrants. Presumably, most people who report immigrants to ICE ﬁrst click on the ICE website, and then make their way to the tip form. Alternatively, they may use the phone number. Due to the rarity of the direct clicks to the tip form, we chose to use search queries that have led at least one user to directly click the tip form, with the rationale that those search queries represent intent to report immigrants.

48 searches as there were “report” searches. The most popular term that led to an HSI click was “ice”, followed by “hsi tip form”. While this measure has less precision than the previously described “report” measure, it does have higher recall. Together, these measures paint a clear picture of searches about reporting immigrants.

Unfortunately, due to Google Trends’ inability to pull searches for a speciﬁc ﬁxed string (Google Trends for ”ice” would give searches for both “ice tip line” and “ice cream”), we were unable to pull the equivalent results for the HSI searches.

Table A8 and Figure A1 are the HSI equivalents for table 2 and figure 4, respectively. Fig- ure 4 confirms that there was an increase in searches about reporting immigrants after Trump was inaugurated. Table A9 shows that while there was no significant association between immigrant crime coverage and HSI searches, there was still a strong association between total coverage of immigration and HSI searches.

The HSI search ﬁndings replicate our reporting immigrant results from the body of the paper, providing powerful evidence that searches for reporting immigrants did increase after Trump was inaugurated, and this increase was associated with higher media coverage of immigration.

49 Table A8: Top 10 terms that resulted in HSI tip form clicks

HSI ice hsi tip form www.ice.gov/tips report illegal immigrants anonymously ice tip line www.ice.gov/webform/hsi-tip-form ice report ice reporting illegal immigrants ice.gov immigration and customs enforcement

Notes: Table represents top 10 searches that result in clicks to the hsi tip form.

Figure A1: HSI Bing Searches by Time

Notes: Searches for reporting immigrants increased after Trump was inaugurated. This ﬁgure is the counterpart to ﬁgure 4 in the body of the paper.

50 Table A9: Immigration Coverage and HSI Searches

Dependent variable:

Immigration Segs 0.002∗∗∗ (0.0003) Immigr + Crime Coverage −0.00004 (0.004) Immigr + Welfare Coverage −0.027∗∗ (0.013) Trump Admin 0.299∗∗∗ (0.026) Date −0.00003 (0.00003) Day of Week FE X Month FE X Constant −11.598∗∗∗ (0.583)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: There was a strong correlation between daily immigration coverage and searches for how to report immigrants using the HSI measure. This table is the counterpart to table 7 in the body of the paper.

A.3.3 Bing Speech Results

While we were unable to compare Trump and Obama speeches using Bing due to data limitations (first Bing data was collected after Trump’s first State of the Union), here we present the data from the Bing analysis of Trump’s speeches in figure A2 and Table A10. The data clearly replicates the Google Trends data presented in the body of the paper, with clear increases in anti-immigrant searches during Trump’s speech broadcasts.

51 Figure A2: Bing Immigration Searches by During Trump Speeches Crime: Welfare:

Report: Report (HSI):

Notes: Searches for immigrant crime, immigrant welfare, and reporting immigrants (HSI) increased sharply during Trump’s televised speeches. Dataset includes 2018 State of the Union, 2019 State of the Union, and 2019 immigration Oval Office address. For crime and report (HSI), increases were consistent across all three speeches, whereas for welfare the increase occurred only during the Oval Office address. This figure is the Bing counterpart to figure 5 in the body of the paper.

52 Table A10: Bing Immigration Searches During Presidential Speeches

Dependent variable:

Crime Welfare Report Report (HSI) Speech date 0.240 (0.173) 0.141 (0.094) 0.138 (0.090) 0.026 (0.045) Speech hour 0.357 (0.233) 0.332∗∗∗ (0.084) 0.242 (0.185) 0.135 (0.113) Speech date x speech hour 1.433∗∗ (0.558) 0.077 (0.403) −0.048 (0.259) 1.318∗∗∗ (0.314) Constant −11.754∗∗∗ (0.041) −11.816∗∗∗ (0.040) −13.751∗∗∗ (0.052) −11.726∗∗∗ (0.029)

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: There was a signiﬁcant spike in Bing searches for crime and reporting immigrants (HSI) during Trump’s speeches. Regression is binomial logit with SEs clustered by date-hour. Speech Date variable was 1 if the search was performed on the Date of the speech; 0 otherwise (performed on the date one week before the speech or one week after). Speech hour variable was 1 if the search was performed during the hours of 9 or 10pm EST (with the exception of the Oval Ofﬁce speech, for which it was only 9pm EST). This table is the Bing counterpart to table 10 in the body of the paper.

A.3.4 Weather Placebo Results

In order to ensure that our immigration results were not due to larger shifts in web searching behavior, we replicated our analysis using searches for the word "weather". There should be no media and no Trump administration effect on weather searches.

Figure A3 repeats the analysis by administration (ﬁgure 4) for Bing and Google searches for a placebo search term, “weather”. Google trends showed an unusual overall pattern for this search, with searches for weather increasing at a steady rate throughout the time series. How- ever, there appeared to be no discontinuity at the Trump administration for the Google searches - instead there appeared to be a substantial bump in searches around mid-2018. There was also no discontinuity in Bing weather searches, although we did see a slight upward time trend similar to the Google data. These ﬁndings suggest that the observed shifts in immigration searches were not the result of an overall shift in search patterns.

Table A11 shows the results of the “weather” placebo test on immigration media coverage. Searches including the term "weather" should have no clear relationship to immigration

53 Table A11: Media Coverage and Weather Searches (Placebo)

Dependent variable: Bing Google Immigration Segs 0.0003 (0.0003) 0.018 (0.014) Immigr + Crime Coverage −0.011∗∗ (0.005) −0.254 (0.200) Immigr + Welfare Coverage −0.003 (0.007) −0.733 (0.518) Trump Admin −0.117∗∗∗ (0.032) 7.024∗∗∗ (1.350) Date 0.0003∗∗∗ (0.00003) 0.014∗∗∗ (0.001) Dow FE X X Month FE X X Constant −11.324∗∗∗ (0.582) −168.512∗∗∗ (22.980)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Notes: There was no consistent relationship between immigration coverage and weather searches for either Bing or Google. Bing regression is binomial logit with standard errors clustered by day; Google regression is OLS. coverage. As expected, there is no consistent relationship between immigration coverage and weather searches on either Google or Bing. This suggests that the observed relationship between immigration coverage and anti-immigrant searches was not the result of general patterns in search volume.

Finally, to check that the observed search spikes were not the result of overall shifts in search patterns, we used searches for “weather” during the same time periods as a placebo. Figure A4 shows that for both Google and Bing, there was no spike in searches for weather during speech times for either Obama or Trump. In sum, these three tests show that the observed change in immigration searches is not the result of broader shifts in search behavior.

54 Figure A3: Weather Placebo Searches and Presidential Administrations Bing: Google:

Notes: There was no clear Trump administration effect on weather searches. Bing searches showed no discontinuity between the Obama and Trump administrations. While Google weather searches did appear to be higher during the Trump administration, there was no clear discontinuity during the ﬁrst few months of the administration, and the clear upward trend of the Bush and Obama administrations in weather searches continued.

55 Figure A4: Weather Searches During Presidential Speeches Bing:

Google:

Notes: Neither Bing nor Google showed any spike in weather searches during any televised speeches. For Google searches, plot Y-axes are comparable within, but not between, days. A rating of 100 represents the hour with the most searches on that speciﬁc date, but is not comparable with searches for other dates.