Abstract
This study examines how mainstream political actors and other organizations use political targeted messages. For this purpose, a data set from ProPublica is used. The study examines 55,918 sponsored Facebook ads that were posted by 236 political actors (i.e., political elites and other organizations) in the United States. (1) Topic classification was used to identify policy issues, (2) network analysis to identify the main policy issues from the various political actors, and (3) Sankey diagrams to visualize microtargeted messages. Our findings indicate that actors focus on traditionally owned issues (i.e., the Democratic Party: environmental policy, social issues, and social welfare; the Republican Party: foreign affairs, law, and government finances). No clear evidence for a focus on wedge issues can be found, however, some first indications (e.g., a focus on reproductive rights, LGBTQ+) are present in a targeted media environment. All in all, the current study helps us to understand in what way political actors deploy targeted messages.
(Tar)getting You: The Use of Online Political Targeted Messages on Facebook
The use of online political targeted messages by political elites, but also other organizations that intervene in the political arena (Dobber and Borgesius, 2019), has gained popularity in recent years (Zuiderveen Borgesius et al., 2018; Strömbäck and Kiousis, 2014; Dommett, 2019; Dommett et al., 2020). Data-driven targeting, in particular online, involves the collection of vast amounts of personal data, the use of that data to identify and select specific (groups of) people that might be affected by the message, and subsequently, sending these people targeted messages (Zuiderveen Borgesius et al., 2018). Political actors, and other organizations, are increasingly relying on external intermediaries, such as Facebook, to assist with these practices (Dommett et al., 2020; Barrett, 2020).
Data-driven targeting practices also received a growing amount of journalistic attention (Baldwin-Philippi, 2017) and scholarly inquiry (Bodó et al., 2017). Research has focused on the use of data-driven techniques during elections (Kreiss and McGregor, 2018; Kreiss, 2016) from the perspective of political elites (Dobber et al., 2017), discussed the (possible) ways to regulate targeting (Dobber and Borgesius, 2019), and focused on the normative implications of targeting, specifically discussing the challenges that political targeting brings for democracy (Bayer, 2020; Zuiderveen Borgesius et al., 2018; Barocas, 2012). This latter strand of research argued that, on the basis of democratic theory, targeting might have negative implications for democracy as it could lead to voter manipulation, discrimination, and violations of privacy on a micro-level, and to more power to well-funded parties on a macro-level (Tufekci, 2014). Conversely, targeting might also be an effective tool to reach citizens that are difficult to reach via traditional media and by engaging those who are less interested in politics (Zuiderveen Borgesius et al., 2018).
However, while many claims about the democratic impact of targeting have been made, less attention has been devoted to the actual practice of targeted political messages in the online environment (Bennett and Gordon, 2020), most notably on social media such as Facebook. Previous work, for instance by Kim et al. (2018), has examined how outside anonymous groups run divisive issue campaigns and who they target. Endres (2016) studied the 2012 Republican campaign’s targeting efforts in three states. But the way in which mainstream, country-wide, political elites and organizations use Facebook ads to target specific audiences received little attention. Hence, the main research question of this paper is who (i.e., which political actor), sends what type of targeted message (i.e., the actual content) to whom (i.e., which targeted audience or individual; based on the ideas of Lasswell, 1948). The current research question answers a crucial, but yet to be explored question in communication science.
Thus, the current study aims to uncover how mainstream political actors and organizations use political targeted messages when targeting specific audiences, by giving insight into the content of such targeted messages, by building on the theory of issue ownership (Lefevere et al., 2015; Petrocik et al., 2003) and of the persuadable voter (Hillygus and Shields, 2008). It will examine the extent to which parties, candidates, and other organizations are associated with specific issues, but it will also focus on parties and candidates addressing issues which might be traditionally owned by their opponents. This is specified as “issue trespassing” or “issue convergence” (Walgrave et al., 2015, p. 779), and is expected to be specifically relevant when it comes to political targeting. We focus on sponsored Facebook ads posted in 2018 in the United Sates to address our aim. Investigating online targeted messages is challenging because of the opacity of the social media ecosystem (see e.g., Chester and Montgomery, 2017) and the limitations of ad archives (see e.g., Leerssen et al., 2019), therefore we use data from ProPublica. In addition, to empirically investigate vast amounts of online data, we analyzed 55,918 sponsored Facebook ads that were posted by 236 political actors.
Targeting in the political arena
Formally, targeting can be defined as the specific alignment and delivery of (commercial) messages at predefined audiences (Schlee, 2013). Most often, it refers to online advertising that is based on personal data the advertiser has about the recipient, such as (socio-) demographics, preferences and interests, social connections, browsing behavior, and location data (Schumann et al., 2014; Stiglbauer and Kovacs, 2018). While it is tested and employed in the commercial sector to a great extent, targeting found their entrance into political campaigning (Chester and Montgomery, 2017). This resulted in the widespread use of a wealth of personal data to mobilize and persuade voters, a practice which has been referred to as political microtargeting (Zuiderveen Borgesius et al., 2018). Much of these data-driven political campaigns are occurring in (and facilitated by) a highly commercialized social media marketplace. Social media platforms own the most valuable troves of personal data, which they subsequently offer to political actors to reach potential voters (Tufekci, 2014).
Changing campaign dynamics and normative implications of targeting
The practice of political microtargeting has not remained undisputed. In fact, it has recently become the subject of heightened academic and public debate (Baldwin-Philippi, 2017). A first set of concerns in this debate centers around the implications for the individual voter, since political microtargeting offers an innovative method that facilitates voter deception, manipulation, discrimination, and privacy intrusion (Zuiderveen Borgesius et al., 2018; Susser et al., 2019; Leerssen et al., 2019; Chester and Montgomery, 2017). Another set of concerns focus on the implications for our society as a whole, where political microtargeting could lead to a power transfer to wealthy and well-funded political elites (Tufekci, 2014).
Within this normative debate, a specific consequence - exposure to political targeted information about policy issues - receives significant attention. It is argued that in a more digitalized media environment, not only the way in which candidates communicate about political issues has changed, but also what they communicate (Hillygus and Shields, 2008). This can have negative consequences. Diversifying the issues of a political campaign to different voters might lead to a fragmented public sphere where certain groups of people are informed about political issues, while this information is being hidden from other groups (Zuiderveen Borgesius et al., 2018; Tufekci, 2014). This could lead to four types of negative consequences. First, following previous work, in a mass-communicated environment, voters are likely to be informed about a limited set of political standpoints that concern specific issues (Zuiderveen Borgesius et al., 2018). Yet, if voters are only interested in the issues they personally care most about, instead of focusing on the overarching issues, they might not engage in deliberative processes with others, such as public debates, about these core issues (Moeller et al., 2016, for a discussion on the common core). Second, by emphasizing different issues to different voters, one voter might assume that a specific issue is central in the political campaign or important to the political party, while another voter might assume that another issue is most important. As a consequence, when these individual voters get repeatedly exposed to targeted messages about these issues, this leads to a biased view of the “issue priorities” of political actors (Zuiderveen Borgesius et al., 2018, p. 89). This view is supported by Bennett and Gordon (2020). They suggest that if one message is directed to one group of voters, and another message to another (group of) voter(s), this might lead to “more ambiguous political mandates for elected representatives” (p. 3), see also Barocas (2012). Third, microtargeting - ‘dark posts’ - that contain false or disinformation, cannot be countered in real time by fact checkers or journalists. Repeated and exclusive exposure to these targeted messages, could lead to misinformed voters. As noted by others, “[i]n the open, false claims might be challenged; in secret, they may stand unchallenged” (Bennett and Gordon, 2020, p. 3). A last concern relates to the focus on wedge issues. Evidence suggests that targeted messages are more likely to focus on ‘wedge issues’ compared to mass-communicated campaign messages (Bennett and Gordon, 2020; Hillygus and Shields, 2008). Wedge issues, which are issues of divisive nature, such as reproductive rights, LGBTQ+, minimum wage, and immigration, might polarize the electorate especially if they are repeatedly and only targeted with these wedge issues (Hillygus and Shields, 2008; van de Wardt et al., 2014).
However, some scholars note that the use of microtargeting might also have a positive impact, as targeting could make election campaigns more diverse; targeted messages offer relevant information about political issues to voters (Zuiderveen Borgesius et al., 2018). In other words, it is argued that by using targeted messages, political actors can inform voters about policy issues that they are most concerned about. Zuiderveen Borgesius et al. (2018) emphasized that this is less likely in “an exclusive mass-communicated information environment” (p. 86), where there is room for a few issues that are more extensively discussed. Niche or other relevant issues might be ignored. Thus, by targeting specific issues to voters, political actors can emphasize and communicate more diverse issues to the electorate.
Much of these concerns about political microtargeting stem from the fact that the social media advertising ecosystem is difficult to monitor, which undermines efforts to identify, diagnose, and remedy potential harms (Leerssen et al., 2019; Chester and Montgomery, 2017). This opacity is mainly due to the personalized nature of political microtargeting, which makes targeted ads invisible to everyone except the specific users that are targeted by these ads (and thus, these ads are hidden from oversight by ‘critical outsiders’; Shaffer, 2019; Tufekci, 2014). In an attempt to increase transparency, big technology companies, such as Facebook and Twitter, have recently implemented publicly accessible ad archives that include political ads running on their platforms. If designed properly, these archives can be a powerful governance tool that enable monitoring of political advertising. However, recent analyses identified major limitations and weaknesses in these ad archive, and as a result, they have been criticized by researchers who find their contents to be incomplete and unreliable (Edelson et al., 2019, 2020; Leerssen et al., 2019; Mathias and Hounsel, 2018). This has highlighted the need for a more systematic observation and reporting of political ads on social media (Leerssen et al., 2019). This brings us to the main focus of this article, which will be theoretically underpinned in the next paragraph.
The content of microtargeting: issue ownership, wedge issues, and social media posts
To understand what political actors talk about online when using targeted messages (i.e., the policy issue they focus on), we turn to issue ownership theory (Petrocik et al., 2003; Walgrave et al., 2015). To give a short definition, “issue ownership means that some political parties are affiliated with specific issues, and considered best able to deal with them” (Walgrave et al., 2015, p. 778). For instance, Democrats are associated with issues like social welfare, education, health care, the environment, and abortion, while Republicans ‘own’ other issues, such as taxes, defense, crime, trade, and the economy (Petrocik et al., 2003; Banda, 2015; NBC/Wall Street Journal, 2018; NPR/PBS NewsHour/Marist, 2019)
Studies investigating the consequences of issue ownership found that issue ownership affects citizens’ vote choice, albeit dependent on party preferences and issue salience. For instance, Bélanger and Meguid (2008) found that issue ownership affects voting decisions, but only for people who also consider the issue to be important. As argued by the same authors, in general, when parties are perceived to be more competent on a specific issue, voters are more likely to support that party. Thus, issue ownership theory might explain, to some extent, why policy issues are important in explaining how voters make up their minds (Hillygus and Shields, 2008). It is therefore not surprising that previous research showed that parties have a preference to discuss owned issues, for instance, in television advertisement (Brazeal and Benoit, 2008). Moreover, more recent work on issue ownership investigated the extent to which parties address issues owned by other parties (Banda, 2016; Kaplan et al., 2006; Tresch et al., 2015); this is defined as “issue trespassing” or “issue convergence” (Walgrave et al., 2015, p. 779). Research found that over the course of campaigns, political candidates respond to each other by strategically addressing issues owned by their opponents (Banda, 2015). This effect is often found in competitive political environments. By focusing on television advertisements, Banda (2015) asserts that candidates mention their opponents’ owned issues more often over the course of time, showing evidence for issue convergence. Furthermore, whether or not it is possible to ‘steal’ an already owned issue from an opponent also depends on the party system. In two-party systems, issue convergence (thus the presence of overlapping issues) is more likely (Walgrave et al., 2015; Egan, 2013).
Issue convergence is particularly relevant when it comes to microtargeting. By deploying targeted messages, political actors are “[a]rmed with information on which individuals are the most likely to agree (or oppose) their position on any number of issues” (Endres, 2016, p.771). Using this personal information from voters, parties can design more issue-based persuasion strategies (Endres, 2016). Thus, by focusing on an issue that is owned by an opponent, political targeting might increase the likelihood that a political actor can ‘steal’ that issue. For instance, a Republican candidate might ‘send’ an ad about ‘air pollution solutions’ to undecided voters who care about the environment in order to change their vote choice (environment is an issue owned by the Democratic Party). Since issue ownership might affect citizens’ vote choice, this can be a strategic move to persuade voters.
If we turn to the empirical work that has been conducted on politicians’ (organic) social media posts, we find some interesting results. Borah (2016) observed that, in 2008 and 2012, most presidential candidates use Facebook posts to promote the campaign. Approximately one out of three posts were about policy issues. The study also shows that Republicans attack their opponent more often in their posts. Another study, focusing on the Justin Trudeau’s use of Instagram in 2015, showed that various issues were covered in the posts. The posts contained official announcements related to policy issues, such as employment, social development, economy, youth, health, and the environment. The study also shows that official statements were announced, and other posts covered specific events (Lalancette and Raynauld, 2019). Even more interesting, a study, focusing on Senate candidates’ use of Twitter and messaging via broadcast media (i.e., campaign advertising), showed that only a small minority of ads in broadcast media contained no topic, while on Twitter, more than half of the tweets had no issue content (Bode et al., 2016). This resonates with the idea that social media posts might be more often focused on the campaign, on attacking opponents, and they might include more often information about the politicians themselves, instead of focusing on policy issues (Yarchi and Samuel-Azran, 2018).
Taken together, based on these insights, it can be expected that targeted messages overlap, due to issue convergence. However, it might be safer for political actors to pay for messages that are focused on owned-issue to mobilize voters, instead of attracting new voters. Based on these considerations, we expect that:
(H1a): Political actors affiliated with the Republican Party are more likely to use Republican-owned issues than Democratic-owned issues in targeted messages.
(H1b): Political actors affiliated with the Democratic Party are more likely to use Democratic-owned issues than Republican-owned issues in targeted messages.
In addition, previous research indicates that targeted messages might focus more on ‘wedge issues’ compared to mass-communicated campaign messages (Bennett and Gordon, 2020; Hillygus and Shields, 2008). Hillygus and Shields (2008) argue that voters – who disagree on specific issues with their preferred party – are more likely to be persuaded, especially when it comes to issues that they care about. By emphasizing this issue to receptive voters (and thus highlighting internal conflict), political actors can attract these voters. Because the targeting technologies make it easier to identify receptive voters, Hillygus and Shields (2008) argue that the use of wedge issues is more likely. They support their argument by showing that direct mails (a form of targeted content) contain more wedge issues (such as abortion and gay marriage) compared to television advertising. Thus, when using mainstream media, political actors are more likely to use “consensual policy issues”, and when using targeted communication, they are more likely to focus on wedge issues (Hillygus and Shields, 2008, p. 169).
Based on the work of Hillygus and Shields (2008), we also focus on typical wedge issues to examine closely how these more contentious issues are used in targeted messages. Based on Gimpel (2019), we focus on the issue of immigration, as he suggested that since 2016, voters “began to align their policy views with the positions of their favored political parties” (p. 467) and that “[t]he association between party identification and immigration policy views reaches a local peak by 2018” (Gimpel, 2019, p. 473). Furthermore, we focus on divisive issues, based on previous work by Hillygus and Shields (2008) and on the issues which were in 2018 of importance in the public debate. These are reproductive rights, LGBTQ+, gun control, and environment. Based on the arguments of the ’persuadable voter’, we expect that political actors focus more on divisive issue compared to non-profit organizations. We pose the following hypothesis:
(H2): Political actors affiliated with the Republican and Democratic party are more likely to focus on divisive issues in targeted messages (i.e., immigration, reproductive rights, LGBTQ+, gun control, and environment) than non-profit organizations.
It is important to note that we direct our focus on mainstream politics, by including political parties and candidates. In addition, we also included political action committees (PACs) as they often used targeted ads. Lastly, non-profit organizations are included because they are the baseline in our second hypothesis.
Lastly, this study examines who political actors target (i.e., the audiences) using issue-based targeted messages. In other words, who do political actors reach with their issue choices. Insights in the reach of targeted ads received little attention. Kim et al. (2018) focused on anonymous divisive issue campaigns and found that these campaigns targeted specific demographic groups. People with a low income, received more targeted ads about immigration and racial conflict. People with an average income, were targeted with issues about nationalism. They also found that, compared to other ethnic groups, whites receive more targeted ads about immigration and nationalism. So, while this study offers important insights, it does not focus on mainstream political actors. Endres (2016) found that the Republican Party in 2012 was only partially successful in identifying potential wedge issues of specific voters. The Republican campaign’s predictions were “right much of the time, but are far from perfect” (p. 773). However, this study focuses on the 2012 election in three states (Florida, Virginia, and Colorado), and on only one political actor: the republican campaign. Since then, targeting capabilities have evolved and insights into the targeted audience of US-wide political elites are still left to be desired. Based on these considerations, and because a lack of research, the following research question is asked:
(RQ1) How does the audience that is targeted (in terms of gender, age, and political affiliation) vary depending on the divisive issues mentioned in the targeted messages of political actors affiliated with the Republican and Democratic Party?
Method
This study takes a multi-method approach to examine political targeted messages on Facebook 1 . First, topic classification was used to identify policy issues in sponsored political ads. Second, we used network analysis to identify the main policy issues from various political actors (i.e., political elites and other organizations). Finally, we visualize the targeted audience by using Sankey diagrams.
The ProPublica data set
In this work, we study the political ads data set collected by ProPublica, an American non-profit newsroom. The database includes ads that ran on Facebook in the period before the 2018 U.S. midterm elections. ProPublica asked users to install a browser extension that automatically collected advertisements on their Facebook pages and sent them to ProPublica’s servers. They used a machine learning classifier to select which ads were likely political. The data set includes, for example, the title, message, images, the advertiser, targeting information (e.g., related to age, gender, education, ethnicity, country, interests, language), as well as the number of users who have indicated that the ad is political. When examining the data, we found that more Facebook ads were coming from actors from the Democratic Party. We believe that this is caused by the people who installed the browser plugin. We want to point out that the biased sample is an important limitation in this study.
The selection of the accounts happened in two steps. In the first step, we selected 200 accounts with the most sponsored ads in the ProPublica political ads dataset. Four coders coded all 200 different accounts and have distinguished various types of actors as well as their political leaning. This was done in a deductive and inductive manner. We first defined the actors before the start, which resulted in a list of seven actors and one “other category”. However, we did not find actors in our list for each category, so after discussion among the coders, we slightly changed this in the list that can be found in the next paragraph. Furthermore, we used popular sources and online databases (such as www.opensecrets.org, websites, and newspapers) to code the actors. The first step led to a over-representation of democratic sources. To increase the number of conservative sources in our data set, in a second step, we coded an additional 800 accounts – and added 36 Republican accounts. All sources were double coded by another coder, which resulted in a score of .97, indicating conceptually valid and useful output. In case of disagreement, we discussed the discrepancy.
In total, our data set includes 55,918 sponsored Facebook ads by 236 actors (see Table 1). We distinguished the following types of actors: Political parties and organizations (n = 7; Democrat = 2, Republican = 5; e.g., @democrats), Political candidates (n = 72, Democrat = 47, Republican = 22, other = 3; e.g., @betoorourke, @DonaldTrump), PACs (n = 44; Democrat = 24, Republican = 16, other = 3, e.g., @PlannedParenthoodAction, @WorkingFamilies), Public figures (n = 3; e.g., @shaunking), Non-profit organizations (n = 83; e.g., @amnestyusa, @worldwildlifefund), Businesses (n = 9; e.g., @4oceanBracelets), and Other (n = 19; e.g., @nytimes 2 ).
Sample size description.
Automated content analysis of targeted ads
We used TextRazor to obtain a better understanding of the content of the ads. TextRazor is a commercial service comprised of different modules for text extraction. In this work, we focus on the Topic Tagging module that leverages an ensemble of automated techniques to assign topics to the ads. TextRazor has a very large knowledge base (based on Wikipedia, DBPedia, and Wikidata) to automatically assign hundreds of thousands of different topics at different levels of abstraction to political targeted ads. The models used by TextRazor are trained on this knowledge base. Based on a score ranging from 0 to 1, TextRazor determines the relevance of this topic to the processed text. We set the threshold to .80, representing a somewhat high relevance of the topic to the processed text. Next, to examine whether the methods included in the commercially available software of TextRazor are optimal for our data set, and to prevent wrong or biased results (Grimmer and Stewart, 2013), we validated the output. Since the topics of interest are not defined a priori, it is not possible to use a so-called gold standard (Maier et al., 2018). Instead, two coders double coded the returned topics for 120 political ads—the number of correct identified topics set against the number of total identified topics for each ad–resulting in a score of .92, indicating conceptually valid and useful output.
In total, we identified 7,586 unique topics (at a .80 level) for 38,897 sponsored Facebook ads (see Table 1). TextRazor could not classify 17,018 sponsored ads (approximately 30 percent). This is in line with previous work, for example Bode et al. (2016) indicated that 5.9 percent ad airings (on television) have no issue content, whilst issue mentions are much less prevalent on social media (i.e., 54.2 percent of tweets have no issue content). In our data set, the identified topics vary from general topics, such as Politics (n = 20,543), Government (n = 17,635), Health (n = 3,342), Law (n = 7,836), Elections (n = 5,333), and Economy (n = 1,185), to more specific topics, such as Human rights (n = 2,950), Immigration (n = 1,408), Gender equality (n = 986), Climate change (n = 648), Tax (n = 440), Electoral fraud (n = 281), Gun violence (n = 193), and Brexit (n = 5).
Analysis
We use a network approach to obtain a better understanding in political targeted messages on Facebook. We use topics as nodes, whilst edges were created for topics that co-occurred in the same targeted ad. First, a Python script was developed to iterate through each of the 55,918 ads and create edges between topics within the same ad (e.g., Environment - Global warming). The edges between topics have been saved and imported in R. We used the package igraph to conduct network analyses, as it is capable of handling large graphs efficiently (Csardi and Nepusz, 2006).
The ties within these networks formed an undirected graph, as edges in the graph do not have an associated direction. We employ weighted edges to indicate the strength of a relationship between two topics. The value of the relationship (i.e., weight) is added as an edge attribute. We created different topic-networks for different types of actors.
Finally, we explore targeting information—an array of one or more of Facebook’s “Why am I seeing this?” disclosures provided to users of the plug-in. We use Sankey diagrams to visualize political targeted messages. We focus on three different types of personal data, namely age (i.e., <18, 18-35, 36-50, 51-65, >65), gender (i.e., women, men, unspecified), and political leaning (i.e., very conservative, conservative, moderate, liberal, very liberal).
Results
To obtain a better understanding in targeted messages on Facebook, we have created different topic-networks for both political elites as well as other organizations.
Who sends what type of message?
Political Organizations and Candidates
First, we focus on political organizations and candidates—representing 2,712 unique topics in 16,757 sponsored ads (see Table 1). There was some degree of overlap on topics (n = 459). Political funding (n = 391), Foreign electoral intervention (n = 313), and Russia intelligence operations (n = 311) are the most common topics for the Democratic Party that do not occur in sponsored ads by accounts affiliated with the Republican Party. The most common non-overlapping topics for the Republican Party are News media manipulation (n = 138), Criticism of journalism (n = 138), and Fake news (n = 78).
To obtain a better understanding in sponsored ads on Facebook, we first turn our attention to political actors affiliated with the Democratic Party. The top 50 topics for the Democratic Party can be found in Table 6, including topics such as Politics, Elections, and Political events. Next, we created a topic-network for sponsored ads created by political actors affiliated with the Democratic Party. In total, the topic-network consists of 2,429 nodes (i.e., topics), and 77,161 edges (i.e., topics that co-occurred in the same sponsored ad). To obtain a better understanding of the main issues from the various political organizations and candidates, we use a community detection approach. Networks are often characterized by clustering, which is hard to judge by eye. There are many existing community detection approaches (Luke, 2015). As we employ undirected and weighted networks, we rely on the Louvain community detection algorithm in igraph (Csardi and Nepusz, 2006). Modularity is fairly acceptable (.41), suggesting that the Louvain algorithm has done a sufficient job at detecting subgroup structure in our topic-network. “Modularity is a measure of the structure of the network, specifically the extent to which nodes exhibit clustering where there is greater density within the clusters and less density between them” (Luke, 2015, p.115). The membership function reveals that 44 different subgroups have been identified. We examined the six largest clusters in Table 2. Looking at these six clusters, we can see a clear divide between various issues. Political organizations and candidates affiliated with the Democratic Party mainly focus on Social issues, Social welfare, and Environmental issues.
The six largest clusters of sponsored ads by political parties and candidates.
Note. D = accounts affiliated with the Democratic Party; R = accounts affiliated with the Republican Party.
We conducted comparable analyses for politicians affiliated with the Republican Party. The top 50 topics for the Republican Party can be found in Table 7, including topics such as Government, Donald Trump, and Public opinion. The topic-network consists of 752 nodes, and 12,368 edges. Again, we identified multiple clusters within this topic-network (n = 29; modularity = .55). We examined the six largest clusters in Table 2. Looking at these six clusters, we can see a clear divide between various issues. In comparison with the Democratic Party, political organizations and candidates affiliated with the Republican Party increasingly focus on Foreign affairs, Law and Economy. Interestingly, two large clusters of political actors affiliated with the Republican Party relate to Environmental (e.g., Climate change policy, Air pollution) and Health care issues.
Political Action Committees - PACs
Second, we turn our attention to political action committees (PACs)—representing 2,541 unique topics in 9,809 sponsored ads (see Table 1). The top 50 topics for PACs affiliated with the Democratic Party can be found in Table 6 and for the Republican Party in Table 7. Again, there was some degree of overlap on topics (n = 427). Human reproduction (n = 529), Women’s rights (n = 519), and Brett Kavanaugh (n = 509) are the most common topics for the Democratic Party that do not occur in sponsored ads by accounts affiliated with the Republican Party. The most common non-overlapping topics for the Republican Party are Legal concepts (n = 32), Politics of Montana (n = 23), Montana Legislature (n = 23).
The network of the Democratic PACs consists of 2,290 nodes, and 67,859 edges. Again, we identified multiple clusters within this topic-network (n = 47; modularity = .46). We examined the six largest clusters (see Table 3). Looking at these six clusters, we can see a clear divide between various policy issues. PACs affiliated with the Democratic Party focus on Environmental issues, Law & Foreign affairs, Social issues, and Education.
The six largest clusters of sponsored ads by Political Action Committees.
Note. D = accounts affiliated with the Democratic Party; R = accounts affiliated with the Republican Party.
We conducted comparable analyses for Republican PACs, which consists of 676 nodes, and 13,991 edges. Again, we identified multiple clusters within this topic-network (n = 21; modularity = .59). We examined the six largest clusters in Table 3. In comparison with the Democratic Party, PACs affiliated with the Republican Party increasingly focus on Foreign affairs, Economy, Law, and Donald Trump.
Non-profit Organizations
Finally, we focus on sponsored ads from non-profit organizations (4,575 unique topics in 21,561 sponsored ads; see Table 1). We present the top 50 topics present in sponsored ads from non-profit organizations in Table 8. The network of non-profit organizations consists of 4,575 nodes, and 170,517 edges. We identified multiple clusters within this topic-network (n = 67; modularity = .57). We examined the six largest clusters in 4, including Environment, Social Issues, and Multiculturalism & Religion.
The six largest clusters of sponsored ads by non-profit organizations.
Coming back to our hypotheses (H1a en b), we find mixed evidence. It seems that political actors affiliated with the Democratic party focus on Democratic owned-issues (such as Environment, Social welfare, and Social issues), while they also focus on Republican owned-issues (such as Foreign affairs). Likewise, political actors affiliated with the Republican party focus on Republican owned-issues (such as Economy, Law, and Foreign affairs), while they also focus on Democratic owned-issues (such as Environment and Health care).
Policy issues
As a next step, we use the output of our topic-networks to examine five policy issues—as mentioned in our theoretical framework—namely: (1) Environment (N
Percentage of policy issues in sponsored Facebook ads.
1 = Environment, 2 = Immigration, 3 = Reproductive rights, 4 = LGBTQ+, 5 = Gun control.
Who sends what type of targeted message to whom?
To answer our research question, we link three targeting criteria, namely age, gender, and political leaning, to five policy issues. We use Sankey diagrams to visualize our findings. The arrows have a width proportional to the flow quantity presented (in percentages). The absolute values are visualized in Appendix C. Figure 1(a) presents the percentage of sponsored Facebook ads targeting age (N = 52,979). Categories are not mutually exclusive. Based on our data set, we found that political parties, candidates, PACs as well as non-profit organizations seem to rarely target Facebook users younger than 18 years. As shown in Figure 1(b), political candidates and PACs affiliated with the Republican Party mainly target ads related to Immigration to specific age groups. Political candidates affiliated with the Democratic Party mainly target ads related to LGBTQ+ to specific age groups. Finally, non-profit organization mainly target messages about Environment. Yet, no clear divide between age groups can be found.
Percentage of policy issues in sponsored Facebook ads targeting age. (a) Age; (b) Age & Policy issues
Next, Figure 2(a) presents the percentage of sponsored Facebook ads targeting gender (N = 2,961; men = 931, women = 2,030) in our data set. Political candidates affiliated with the Democratic Party primarily choose to target women, whereas political candidates affiliated with the Republican Party also target men. Next, as visualized in Figure 2(b), political parties, candidates, PACs, and non-profit organizations rarely target men about Reproductive rights, Gun control, and Environment. In this dataset, political candidates affiliated with the Democratic Party target both men and women about LGBTQ+ and Immigration.
Percentage of policy issues in sponsored Facebook ads targeting gender. (a) Gender; (b) Gender & Policy issues
Finally, Figure 3(a) presents the percentage of sponsored Facebook ads targeting political leaning (N = 2,885). Categories are not mutually exclusive. Based on our data, it seems that political candidates and PACs affiliated with the Democratic Party as well as non-profit organizations primarily choose to target (very) liberal Facebook users, whereas political candidates and PACs affiliated with the Republican Party often target moderate as well as (very) conservative Facebook users. Interestingly, as visualized in Figure 3(b), political candidates affiliated with the Republican Party target very liberal Facebook users with sponsored ads about Immigration. Political candidates PACs affiliated with the Democratic Party rarely target (very) conservative Facebook users about Reproductive rights, LGBTQ+, Immigration, Environment, and Gun control.
Percentage of policy issues in sponsored Facebook ads targeting political leaning. (a) Political leaning; (b) Political leaning & Policy issues
Conclusion & Discussion
The purpose of this study was to get more insights into how and which mainstream political actors use sponsored ads by investigating the issues mentioned in almost 56,000 political Facebook ads using a data base from ProPublica. Additionally, we examined the audience that was targeted with these specific messages. Studying the actual content of targeted messages is important, as it helps us to understand in what way political actors deploy this strategy during and after election campaigns, as the content may affect support for political actors and policies. Research showed that targeted ads can mobilize (younger) voters (Haenschen and Jennings, 2019), however, many worry about the detrimental effects of manipulation, loss of privacy, and demobilization among voters (Zuiderveen Borgesius et al., 2018). Yet, as we discussed, investigating online targeted messages is challenging because of the opacity of the social media ecosystem (see e.g., Chester and Montgomery, 2017) and the limitations of ad archives (see e.g., Leerssen et al., 2019). We believe that this is an important reason why the actual content of targeted ads is rarely examined on a large scale. The research has some interesting findings. First, based on the ProPublica data set, we showed that Democratic political candidates mainly focus on social issues, social welfare, and environmental issues. The PACs that are affiliated with the Democratic Party focus most often on environmental issues, law and foreign affairs, social issues, and education. Republican candidates focus more on foreign affairs, law and economy in their Facebook ads. PACs affiliated with the Republican Party increasingly focus on foreign affairs, economy, law, and Donald Trump. These findings are largely in line with issue ownership theory (Petrocik et al., 2003; Walgrave et al., 2015): it seems that the candidates focus more on traditionally owned issues. However, we also found – for the largest clusters in our data set – some evidence of issue convergence (Walgrave et al., 2015). For instance, political actors affiliated with the Republican Party focus also on environmental issues (e.g., climate change policy, air pollution). This indicates that, while often targeting voters with ‘owned issues’, in some instances, in 2018 and in our data set, political actors are also buying Facebook ads to target an issue owned by the other party (as it was the case for the Republican Party). It is thus likely that by focusing on these ‘stolen’ issues, targeted ads seem to be used for issue-based persuasion strategies. This is consistent with previous work by Endres (2016). It is also not surprising to find that non-profit organizations are focusing on issues, such as the environment, social issues, and multiculturalism and religion.
Second, we found in the data that non-profit organizations focus on the issues of Environment and Immigration. The Republican candidates and organizations focus slightly more on issues of Immigration, while the PACs affiliated with the Democratic Party focus more the issue of Reproductive rights and LGBTQ+. This shows, interestingly, some modest evidence of an increased focus on wedge issues (but for the issue of gun control – no pattern could be found). This might indicate that of an increased focus on wedge issues in a more targeted media environment, which is in line with Hillygus and Shields (2008), yet future work needs to investigate whether this finding holds, in particular as our data is not representative.
Third, we examined three targeting criteria, namely age, gender, and political leaning. The results gives some first indication how political actors and other organizations target audiences, in our data set for instance more women than men, particularly about wedge issues (e.g., LGBTQ+, Reproductive rights, Environment). Besides, political actors affiliated with the Republican Party seem to target both moderate as well as (very) conservative Facebook users, and seem to target very liberal users about Immigration. Political actors affiliated with the Democratic Party rarely target (very) conservative Facebook users about wedge issues. We could not find a clear divide between age groups. Again, future work needs to investigate these findings using more representative data.
These findings implicate, from a societal perspective, that different citizens receive different pieces of political information, which could advance more fragmentation. Indeed, Bennett and Pfetsch (2018) as well as Waisbord (2016) warned that fragmentation might lead to the disappearance of a shared common ground. In any case, a targeted focus on different and wegde issues reinforces the already existing inequalities of the US electoral system where persuadable voters receive much more political information than voters perceived to be less persuadable (Hillygus and Shields, 2008). From an individual citizen position, a focus on specific issues can on the one hand be beneficial because this strategy provides the citizen with insights on personally relevant issues. On the other hand, focusing on certain issues and ‘hiding’ other information might lead to false citizen’ perceptions about issue priorities. A citizen might wrongly assume that, say, abortion is a priority because that citizen sees many ads about abortion, while in reality only a relatively small group of citizens gets to see ads about abortion (see e.g., Zuiderveen Borgesius et al., 2018). From a candidate’s perspective, running a targeted campaign means communicating many different specific pieces of policy information to relatively small groups might raise questions about mandate interpretation (Hillygus and Shields, 2008).
Furthermore, an important contribution of this research is its adherence to open data principles. As an important ingredient of the open science paradigm (Dienlin et al., 2020), open data refers to the practice of opening up one’s data and code and uploading it to a public repository (Klein et al., 2018). The data is available on the website of ProPublica, which is updated on a daily basis, and contains political advertisements that ran on Facebook. We made our code publicly available on the Open Science Framework (see Footnote 1). With this open-source approach, we aim to contribute to research transparency and data re-usability, as well as ensure that our findings are reproducible and provide a solid basis for future collaborations (Van Atteveldt et al., 2019; Dienlin et al., 2020).
Finally, we address limitations of the study and discuss directions for future research. The first limitation relates to the breadth and depth of our data. Using automated content analysis, we are able to examine a broad span of Facebook ads. However, this also results in a large amount of topics under scrutiny. By zooming in on the six most important clusters, and thus the most often included policy issues in the Facebook ads, we gain more in-depth knowledge about these specific topics. However, this also means that we cannot focus on all other, smaller clusters (policy issues) in our data-set. This would also make less sense empirically, due to the long tail (a lot of smaller clusters of topics are only a few times included in our data-set). So, consequently, the breath of our approach comes at the expense of the depth. The approach we used, might be considered a ‘mass media approach’ instead of a ‘tailored media approach’. In other words, we did not merely focus on niche topics discussed in targeted Facebook ads, while those might be also very important when it comes to microtargeting. However, it would be less valid to focus on them as we were not sure whether or not they are actually niche topics (due to the fact the data is not representative). Yet, future work could focus on these more specific targeting practices.
The second limitation is related to the type of data. The data we used included more Facebook ads coming from actors from the Democratic Party. We believe that this is caused by the people who installed the browser plugin, which are probably democratic voters. While we do not make any claims about the representativeness of our data, we want to acknowledge this is an important limitation of our study. We still believe the data-set has advantages over the Facebook ad archive. But we encourage future work, that uses content analyses, to work with a representative sample of targeted ads.
The third limitation relates to the use of a commercial off-the-shelf natural language processing tool (i.e., TextRazor). This tool involves a proprietary algorithm to conduct topic modeling on large data-sets. This means that the exact formulation of the algorithm–the source code–is not publicly available for scrutiny. In such instance, the software is essentially a black box: it is difficult to understand its internal workings (Trilling and Jonkman, 2018). Although we took steps to validate the output of TextRazor (see our method section), we have to acknowledge the potential challenges that this commercial tool introduces with respect to providing full transparency of the current findings (Broussard, 2016; Busch, 2014).
To conclude, we have merely focused on analyzing textual content. We have not analyzed visuals (e.g., images and videos) belonging to targeted messages. Previous work has already indicated that images are not only easier to recall, but they can also communicate information in a much more efficient and emotional way. To obtain a better understanding of the role of visuals in political targeted messages, future work should focus on using computational methods to automatically study and extract key information from many images and videos.
The findings in this study on political targeted messages provide new insights into the continuous campaign efforts of political elites. We found that political actors affiliated with the Republican and Democratic Party are more likely to use both Republican-owned and Democratic-owned issues in targeted ads issues, indicating strategies that prioritize issues that already owned (issue ownership) and focusing on issues owned by the opponent (convergence). No clear evidence for a focus on wedge issues can be found, however, some first indications of this are present. We also found some first indications about the strategic targeting practices, for instance, political actors affiliated with the Republican Party target very liberal users about immigration. On the other hand, we found an absence of this in other instances. Political actors affiliated with the Democratic Party rarely target (very) conservative Facebook users about wedge issues. Whether these practices were actually part of a larger campaign strategy cannot be answered in this study, but our findings give a first notion about potential strategies.
