Social media campaigning and voter behavior–evidence for the German federal election 2021 Abeer Ibtisam Aziz *, Ivo Bischoff Institute of Economics, University of Kassel, Nora-Platiel-Straße 4, 34109, Kassel, Germany A R T I C L E I N F O JEL classification: D72 D83 Keywords: Online political campaigns Voter behavior Facebook Twitter Multinomial logit model with alternative- specific constant Germany A B S T R A C T We analyze the relationship between party campaigning on social media and the voting intentions voiced by respondents in a large representative survey in the German federal election campaign in 2021. We argue that online campaigns spread through the online and offline networks of the initial recipients – thereby influencing substantial parts of the electorate. We exploit inter- temporal, inter-regional, and inter-party differences in the intensity of campaigning by parties and candidates on Facebook and Twitter. In addition, we control for the respondents’ choice in the previous federal election and a number of other personal characteristics. Using a multinomial logit model with alternative-specific constants, we find the probability of a respondent’s intention to vote for a party to increase in the state-specific campaigning activities on social media of this party the days before. While the literature suggests that especially populist right-wing parties will benefit from campaigning on social media, we find the marginal impact to be significantly higher for Christian democrats, Social Democrats, and Greens than for the right-wing “Alternative für Deutschland”. 1. Introduction Social media networks have become a heavily used arena for political campaigning (e.g., Stier et al., 2018; Bright et al., 2020; Zhuravskaya et al., 2020; Campante et al., 2018). The costs of campaigning are much lower than for other forms of campaigning (Broockman and Green, 2014; Leung and Yildirim, 2020). Campaigning via social media benefits from the fact that information spreads through the social networks of the citizens (Bond et al., 2012). Therefore, candidates’ posts not only reach their direct fol- lowers but also the friends and acquaintances of these direct followers (e.g., Bene, 2018) and subsequently also the followers and acquaintances of the latter (Shmargad and Sanchez, 2022). This is similar to the spillover impact of traditional media, where the information read or heard from a certain source is passed on to one’s family, friends, and acquaintances. Due to the characteristics of social media, however, information spreads much faster and potentially much further than through offline networks (e.g., Coppock et al., 2016; Cox et al., 2024; Aridor et al., 2024). Moreover, unlike in campaign ads on TV or radio, spillovers on social media imply that the original content is passed on through the network. These features make campaigns on social media especially far-reaching (e. g., Shmargad and Sanchez, 2022).1 As the formation of social networks is driven by homophily, the content on social media spreads primarily along the lines of like- minded citizens. This results in so-called echo chambers (e.g., Jamieson and Cappella, 2008). Through these, the flow of politically * Corresponding author. E-mail addresses: Aziz@uni-kassel.de (A.I. Aziz), Bischoff@wirtschaft.uni-kassel.de (I. Bischoff). 1 Spillovers may be partly driven by politically engaged internet and social media users (e.g. Norris and Curtice, 2008). Contents lists available at ScienceDirect European Journal of Political Economy journal homepage: www.elsevier.com/locate/ejpe https://doi.org/10.1016/j.ejpoleco.2025.102685 Received 24 January 2025; Received in revised form 24 April 2025; Accepted 4 May 2025 European Journal of Political Economy 88 (2025) 102685 Available online 8 May 2025 0176-2680/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). mailto:Aziz@uni-kassel.de mailto:Bischoff@wirtschaft.uni-kassel.de www.sciencedirect.com/science/journal/01762680 https://www.elsevier.com/locate/ejpe https://doi.org/10.1016/j.ejpoleco.2025.102685 https://doi.org/10.1016/j.ejpoleco.2025.102685 http://crossmark.crossref.org/dialog/?doi=10.1016/j.ejpoleco.2025.102685&domain=pdf http://creativecommons.org/licenses/by/4.0/ relevant information is fragmented and citizens predominantly encounter and engage with content that supports their own view (e.g., Garz et al., 2020; Aridor et al., 2024). The algorithms on Facebook and Twitter are found to support this fragmentation (Cinelli et al., 2021; Halberstam and Knight, 2016; Aridor et al., 2024). Micro-targeting potentially adds to this effect (e.g. Hoferer et al., 2020; Grossman and Helpman, 2023). It describes the fact that, based on the big data on social media users, parties can tailor their political campaigns to very specific groups of voters (Papakyriakopoulos et al., 2018; Liberini et al., 2025). The more fragmented the electorate is, the more effective is micro-targeting (Grossman and Helpman, 2023; Aridor et al., 2024). These effects are difficult to achieve through traditional media campaigns and thus make political campaigns through social media particularly promising. So far, the number of empirical studies that link social media campaigning to election outcomes is limited (e.g., Kelm et al., 2023). Thus, we know little about the true effectiveness of political campaigns on social media. This is where our paper comes in. We use data on the political campaigning activities on social media of the six major German parties in the weeks prior to the German federal election in 2021. The 2021 federal election campaign is highly suitable to estimate the effect of social media campaigns because the COVID-19 restrictions largely prevented face-to-face campaigning – especially canvassing and large-scale campaign events (e.g., Kelm et al., 2023). Our data contains all Facebook posts and tweets on Twitter of the candidates of the six major parties running for a seat in the German Bundestag in the 16 German states as well as all Facebook posts of their party chapters at the state level. We relate the inter- temporal, inter-regional, and inter-party differences in these social media campaigning to citizens’ voting intentions voiced in a representative survey in the weeks preceding the German federal election in 2021. The relevant data stems from the well-known survey run by the FORSA institute covering more than 14.000 German citizens in all 16 states in the weeks prior to the elections. In the survey, each respondent is asked to state her voting decision if the federal election was taking place next Sunday. They can choose between all major parties running in state s plus the options “I don’t know” or “I will not vote”. Small parties are captured in the category “other party”. We model the choice of respondents using a multinomial logit model. This model controls for personal characteristics – among them the self-reported voting decision in the federal election 2017. Our primary variable of interest is the intensity of social media campaigning by the major political parties in a certain state s in the days before respondents in this state participate in the survey. This campaigning intensity is modeled as attributes of the alternatives – i.e. the political parties – the respondent can choose from. We hypothesize that intense political campaigning by party p in state s on the days t-1 and t-2 increases the probability that a respondent from the same state will choose party p on day t. This empirical strategy is based on the assumption that the specific social media post has a stronger impact in the state where the sending account is located than in other states. For example, we assume that social media posts by a candidate from the Christian Democratic Party (CDU) in the German state of Hesse have a stronger impact on the inclination to vote for CDU among Hessian respondents than among respondents in other states. We exploit the inter-party, inter-day, and inter- regional variation in the intensity of social media campaigning to minimize endogeneity concerns. In addition, we control for the general political debate about the major parties and their candidates on Twitter as well as for the corresponding coverage in major state-specific TV news formats on public television. As large-scale experimental studies are not permissible in this field of research from an ethical perspective, studies on the impact of social media campaigns have to rely on happenstance data – even if this comes at the price that not all endogeneity concerns can be ruled out fully. We find social media campaigning by candidates on Twitter and party chapters on Facebook to have a positive impact on voting intentions voiced in the FORSA survey. The marginal effect is significantly larger for the CDU and Social Democrats (SPD) and Greens than it is for “Alternative für Deutschland” (AfD) – the right-wing populist party. This result contradicts the notion that especially populist parties benefit from campaigning on social media (e.g., Mutascu et al., 2025). The paper proceeds as follows: Section 2 reviews the relevant literature. Section 3 derives our hypotheses. Section 4 describes the data and our empirical strategy. The results are presented in section 5. Section 6 discusses our results and section 7 concludes. 2. Review of literature A large portion of the literature on political communication on social media restricts the analysis to the online sphere itself - analyzing the content of online traffic and its diffusion across time, regions, or socio-economic and political cleavages. The concepts of filter bubbles and echo chambers received significant attention in this literature (Conover et al., 2011; Halberstam and Knight, 2016; Levy and Ro’ee, 2021). Studies on social media communication in the US show a substantial level of segmentation and polarization in communication networks (e.g., Conover et al., 2011; Halberstam and Knight, 2016). This segmentation may be partly driven by the algorithms of the platform Levy and Ro’ee (2021).2 Polarization is found to be particularly strong on Twitter and Facebook (Cinelli et al., 2021). Some scholars argue that political segregation exists offline – albeit more difficult to detect (Barberá, 2014). Therefore, echo chambers and polarization observed on social media do not prove that social media leads to an increase in polarization among the 2 In an online field experiment, Levy and Ro’ee (2021) asks Facebook users to subscribe to news outlets and then observes the participants’ subsequent activity on Facebook, i.e, whether they read, clicked or shared the posts that they were exposed to. He analyses the impact of the news exposure and consumption on political opinions with a baseline and an end line survey. He finds that users consume news from opposing ideology outlets when they appear on their news feed. This exposure to the opposing ideology is found to reduce the probability that respondents hold a negative opinion or attitudes of the party they are opposing. However, participants are not exposed to the counter-attitudinal news as often as compared to the pro-attitudinal ones even if they subscribed to them according to their treatment group. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 2 electorate (Zhuravskaya et al., 2020) .3 Moreover, information of the voters can also increase or decrease polarization. For instance, politically interested citizens may opt for a variety of media sources to keep themselves informed, thus reducing the chances of being in echo chambers (e.g., Dubois and Blank, 2018). On the other hand, informed voters might still contribute to political polarization, as political candidates take on polarized policy stances to win over the voters (Liu, 2024). However, the insights from studies restricted to the online political communication are limited.4 A separate strand of literature focusses on the link between political activities on social media and offline political behavior.5 Bond et al. (2012) run a field experiment involving 61 million Facebook users in the US congressional election in 2010 to test the impact of messages intended to mobilize voters. The messages were sent to adult Facebook users on the day of the election. They encourage them to vote in the current election, inform them about how many Facebook users have already reported to vote, and provide them with links helping them to find the nearest polling station. The authors show that informing users about the fact that other Facebook users have already voted caused a significant rise in political information-seeking, self-expression, and self-reported voter turnout. The effect of the messages is found to spread beyond the direct recipients of the message to their close friends while the effect on “ordinary Facebook friends” is limited (see also Cox et al., 2024). Broockman and Green (2014) study the impact of online political advertisements on voting intentions and candidates’ image in two field experiments. They use random assignment of the political ads on Facebook to assess whether exposure to the ads influenced the recognition, impression, or voting intention for the given candidate in the ads. In follow-up telephone surveys, they do not find a significant impact of the ads on any of the aforementioned variables. A set of studies exploits inter-regional differences in social media adoption by the general public and relates it to differences in election outcomes. Some studies find Twitter adoption to favor Republicans (e.g., Rotesi, 2019) while others find Democrats to benefit (e.g., Fujiwara et al. (2023). A study on the relationship between social media adoption and financial campaign contributions finds the contributions to candidates present on Twitter to be higher in states with higher Twitter adoption rates. This effect is especially prominent for the new candidates (Petrova et al., 2021). Thus, social media may increase political competition by making it easier for new parties and candidates to enter the political arena. Another set of studies relates political communication about candidates and/or parties to the latter’s performance in elections (e.g., Tumasjan et al., 2011; Jungherr, 2013; Vepsäläinen et al., 2017; Shmargad and Sanchez, 2022). Public Tweets related to party politics and mentions of the German parties in the public discourse on Twitter around the 2009 German Federal elections are found to be a good predictor of election outcomes (Tumasjan et al., 2011; Jungherr, 2013). In the Hungarian general election campaign 2014, post shares on Facebook, unlike other online metrics such as likes and comments on this platform, are associated with higher votes for the corresponding candidates (Bene, 2018). The study most closely related to our paper is the one by Kelm et al. (2023). They analyze the impact of social media usage of constituency candidates in the German federal election 2021 on their vote shares using linear regression models. They find candidates with a social media profile on Facebook, Twitter, or Instagram to win more votes than those without such profiles. They also find pronounced inter-party differences – with candidates form the Social Democratic Party and Green Party to benefit more. Their study differs from ours in several important ways. First, while they focus exclusively on constituency candidates, we include all the can- didates with publicly accessible accounts that were in the running for the Federal elections. Second, instead of relying on the final vote shares, we use the weekly polls before the elections to capture the temporal variation over the campaign period. This allows us to establish a much more direct link between social media campaigning and voting intentions. Third, whereas their main analysis evaluates the effect of the adoption of social media, we examine the impact of social media campaigning by employing a more rigorous approach using a multinomial logit model. The above literature can be summarized as follows: There are numerous studies within the online sphere supporting the notion that political campaigning on social media has an effect on voter behavior offline. Studies that explicitly link these two spheres are limited. The existing studies often find a positive correlation between voting outcomes and the frequency with which parties or candidates are mentioned in the public debate on Twitter and/or the intensity to which citizens respond to candidates’ or parties’ political campaigns on Facebook or Twitter in the form of likes, shares or retweets. Both indicators are clearly endogenous because they capture the re- action of the public and thus voters to campaigning rather than the campaigning impulse itself. Thus, the corresponding results cannot 3 One additional amplification of polarization may result from the activities of bots. Numerous studies point out that social bots are actively interfering in political debates – including the German federal elections (Neudert et al., 2017; Keller and Klinger, 2019; Aziz, 2023). They spread misinformation very fast (Shao et al., 2018), create unfounded hype (Howard et al., 2018), manipulate the flow of information (Gorodnichenko et al., 2021), and influence public opinion (Azzimonti and Fernandes, 2018). Studying the influence of social bots on political information diffusion and public opinions, Gorodnichenko et al. (2021) find support for online echo chambers. 4 A number of studies find a positive relationship between online political activities like following political candidates or reading/writing political blogs) to political knowledge and offline political behavior like participating in demonstrations or protests as well as voter turnout (e.g., Dimitrova et al., 2014; Boulianne and Theocharis, 2020; Potrafke and Roesel, 2025). 5 There has been considerable research on the impact of internet diffusion on voter knowledge and voter behavior in the field of economics (e.g., Falck et al., 2014; Gavazza et al., 2019; Poy and Schüller, 2020). Using data on Germany, Falck et al. (2014) find a negative effect of internet availability on voter turnout. They argue that while fast internet improves access to political information, it also presents a new source of enter- tainment and distraction. This crowding-out effect of the internet is also observed by Gavazza et al. (2019), supporting the idea that a decrease in the political information held by the electorate can dwindle political participation. Poy and Schüller (2020) argue that access to fast internet in the age of social media shifts the patterns of political information acquisition and positively impacts voter turnout. That is, with the advent of social media, the use and the role of the internet in information flows has changed considerably. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 3 be interpreted as causal evidence for the impact of parties’ social media campaigns on offline voting behavior. The current study uses parties’ and candidates’ initial campaigns as explanatory variables and exploits inter-temporal, inter- regional, and inter-party differences in the intensity of political campaigns to test for their impact on voter behavior. This mitigates the above-mentioned endogeneity problem. The dependent variable stems from a large representative survey that also contains infor- mation on which party the subjects voted for in previous elections. This enables us to capture a substantial degree of unobserved heterogeneity at the individual respondents’ level. 3. Hypotheses The primary aim of this paper is to test whether political campaigning on social media has an impact on voter behavior. Our focus rests on the time close to the election when policy platforms have already been fixed and candidates have been nominated. In this period, changes in voter behavior must be related to changes in their stock of politically relevant information and/or the interpretation of it (e.g., Schroeder and Stone, 2015; Kelm et al., 2023). Given that the single voter is very unlikely to be pivotal (e.g. Tyran, 2004), voters cannot be expected to spend significant effort and resources on collecting politically relevant information. Thus, the information underlying this decision is largely collected “en passant”. In times close to the election date, voters are exposed to large amounts of such information from a variety of sources. These include classical media like radio, TV, or newspapers, social media, and campaign activities and material provided by candidates and parties. Additional information is provided in everyday communication with colleagues, neighbors, family members, and friends. In sum, voters have incomplete and potentially biased knowledge about policy platforms and the characteristics of the political candidates. Following recent developments in the theory of voter behavior (e.g., Krasa and Polborn, 2010), we assume that voters evaluate policy platforms based on weighted issue preferences. Expression (1) describes the utility that a certain voter i expects from party p winning office: Ui ( Platformp ) = ∑n j=1 cjp i ⋅ vj i ⃒ ⃒ ⃒Blisspointj i − Platformj p ⃒ ⃒ ⃒ (1) This utility is equal to the sum of weighed differences between her most favored policies and the party p’s policy platform. Following the theory of campaigning (e.g. Denter, 2020), the weights consist of two components: cjp i represents party p’s perceived competence in solving policy field j while vj i represents the perceived valence of policy field j. Party campaigns aim at changing these two components in order to increase the expected utility of voters and thereby the expected number of votes. Each component is targeted by a specific campaigning strategy. Priming refers to measures that make certain issues more salient (e.g., Denter, 2020) – thus influencing vj i. In the context of the German federal election 2021, a priming campaign by the Green party aims at convincing the voter that environmental issues are of overwhelming importance while Christian democrats prime voters to believe that public safety or the economy is the most important issue. At the same time, persuasion targets the voters’ views on the parties or candidates. It aims at increasing their competence cjp i as perceived by the voters and thereby increase the party’s valence (e.g., Schroeder and Stone, 2015; Denter, 2020). This implies that German parties present their most prominent politicians (e. g. Armin Laschet, Olaf Scholz, Robert Habeck) in a positive way – e.g. by pointing out their achievements in previous or current terms as ministers or prime ministers at the state level. When combined with priming, persuasion is particularly effective if it stresses the politicians’ competence in the field that is also targeted by priming. A substantial amount of politically relevant information reaches citizens via social media. Empirical studies support the notion private networks play an important role in spreading political information (e.g., Schmitt-Beck and Mackenrodt, 2010; Coppock et al., 2016; Cox et al., 2024) – especially among young citizens (e.g. Aldrich et al., 2016). Thus, campaigns spread through the social media networks of the person who first encountered it and spill over to an audience that is multiple times larger than the group of original recipients. Additionally, information spills over from one particular platform to other social media platforms and to offline spheres of communication. Increasingly, traditional media covers content originating on social media (Aridor et al., 2024) – especially on Twitter (Kelm et al., 2023). Through these spillovers, social media communication indirectly affects a large number of voters even if many of them are not active on social media. These spillovers are likely to be substantially larger than those of campaigns through traditional media because they are cheaper and faster to spread. Social media campaigns add to the classical tools of face-to-face and mass campaigning. While the meta-study by Kalla and Broockman (2018) concludes that mass campaigning has little if any effect, face-to-face campaigns – especially canvassing – are shown to have an impact (e.g., Cantoni and Pons, 2021). Social media campaigns combine features of both types of campaigns. Like mass campaigns, they are designed to reach large numbers of voters – many of whom were not actively seeking politically relevant in- formation. When it comes to the initial recipients, social media campaigns are similar to face-to-face campaigning events in the sense that recipients actively decide to participate in them by following a certain politician or party. Beyond the initial recipient, campaign messages that spread through personal networks are closer to personal messages than messages sent through mass communication and thus receive more attention (e.g., Schmitt-Beck and Mackenrodt, 2010; Coppock et al., 2016). Micro-targeting makes social media campaigns similar to face-to-face campaigns because the message can be tailored to the specific audience. This effect is likely to hold beyond the initial recipients because voters’ private networks are dominated by people with similar characteristics and attitudes (see section 2). A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 4 We argue that the range of spillovers is strongly dependent on physical proximity because the network on social media centers around real-life acquaintances like family members, close friends, neighbors, and colleagues. Moreover, physical proximity certainly matters for all spillovers that develop offline in personal discussions with these people. Thus, politically relevant information spread by a certain social media account has a large impact in the place where the person or institution owning the account is located. This is particularly obvious for the impact of social media campaigning by local candidates and party chapters. For instance, voters in the state of Hesse are primarily exposed to social media communication of candidates and party chapters from Hesse. Thus, we arrive at the following hypothesis. Hypothesis 1. The intensity of social media campaigning of party p and its candidates in state s prior to day t increases the prob- ability that an individual i in this state votes for party p on day t. Our second hypothesis relates to the question of whether the impact of social media campaigns differs across parties. The literature generally supports the notion that especially populist right-wing parties benefit from campaigning on social media (e.g., Zhuravskaya et al., 2020). Keller and Klinger (2019) study the Twitter followers of the top six German parties during the campaign and election period before the 2017 Parliamentary elections. They find most of the parties share followers, except for the right-wing populist Alternative für Deutschland (AfD). In 2017, AfD had the largest number of single-party followers and also the most active followers overall. Thus, its campaigning on social media may be particularly effective. First, their active followers are likely to cause higher spillovers than those of other parties. Second, the recipients of these messages are less likely to be exposed to social media campaigns by other parties that have the potential to counteract the AfD campaigns.6 Thus, we arrive at our second hypothesis. Hypothesis 2. The increase of the probability that an individual i votes for the AfD is higher for a given intensity of social media campaigning by AfD and its candidates than the analogous increase for other parties. 4. Data and methodology 4.1. Data collection and processing 4.1.1. Offline data: opinion polls Close to federal elections, the well-known German polling institution FORSA publishes opinion polls every Sunday. The weekly polls are based on a representative sample of more than 2000 citizens being surveyed during the week. New samples of subjects are drawn for every poll. Fig. 1 provides a weekly overview of the voting intentions of the survey respondents every week for the duration of our study. FORSA provides individual-level data from their polls (forsa, 2022). We make use of this data for the six weeks before the election. Each participant has to answer a number of general socio-demographic questions and questions on her voting intentions for the upcoming federal election. The data includes the famous “Sunday question”: “Imagine that the federal election would be held this upcoming Sunday, how would you vote?”. The results are presented in Table 1. Compared to the actual election outcomes (see column 3), the vote share of AfD and the share of respondents who declare not to vote are substantially lower in the survey. We account for this fact as well as for the misrepresentation of certain socio-demographic groups by using weights provided by FORSA. 4.1.2. Online data: social media activity To capture the political campaigning on social media, we observe the tweets on Twitter and posts on Facebook of all candidates running for a direct election to the German Bundestag for one of the six major parties. Twitter and Facebook are used actively for political campaigns and for general public debates and they are accessible for automated data collection tools.7 A total of 6211 candidates from 47 parties ran for seats in the German Bundestag 2021. We focus on the candidates from the 6 major parties, namely CDU,8 SPD, Greens (Grüne), Leftwing party (Linke), Liberals (FDP), and AfD. Sältzer et al. (2021) published the Twitter accounts of the candidates for these 6 major parties (n = 2558). The dataset includes the basic information about the state, party, incumbency, etc. for these candidates. It provides Twitter ids for 60 percent of them. We used this dataset to scrape the Twitter timelines of the candidates from the Twitter search API for our study period.9 We collected 111,519 Tweets for 1062 candidates, where 6 This interpretation is supported by the findings of Schaub and Morisi (2020). They analyze the diffusion of the Internet in German municipalities in 2017 and find the support of AfD to be higher in municipalities with higher rates of broadband availability. On the other hand, Kelm et al. (2023) do not find a higher productivity of social media campaigning for AfD. 7 The latter does not hold for Instagram or WhatsApp – two widely used platforms in Germany. Stier et al. (2018) study the German federal election campaign of 2013, where they observe that campaigning on Facebook resembles the usual mass campaigning while Twitter is used for discourse about political events. Furthermore, they also find that Facebook is used for local appeal, as opposed to employing Twitter for the national audience.The use of social media for campaigning varies across the different platforms, depending on their digital architecture, differences in their audience, and outreach (Bossetta, 2018; Stier et al., 2018; Quinlan et al., 2018). For example, Facebook employs a relevance-based algorithm where the posts most relevant to the user will be visible first to them. Whereas Twitter follows a chronological order of posts, where relevance comes into play for the extra posts by users other than those on one’s follow list. The relevance-based algorithm can impact the reach of the non-paid ad- vertisements or messages unless boosted by the party or candidate’s online supporters (Bossetta, 2018). 8 In the state of Bavaria, the sister-party Christian Social Union (CSU) runs instead of the CDU. 9 Some of these candidates either deleted their accounts or did not post during our study period, thus not appearing in our dataset. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 5 22.6 percent of these tweets are original while 77.4 percent are referenced tweets (retweets, quotes, and replies). The most tweets (32 percent) were recorded for the Grüne while AfD had the smallest share (8 percent). In addition, we identified 1085 Facebook pages of these candidates and downloaded the data on the posts and number of followers for the six weeks prior to the election10. These amount to around 62,733 posts, where the candidates from CDU have been the most active, followed by AfD and FDP. There are considerable differences in party activity on Facebook and Twitter. The parties CDU and SPD have the highest number of candidates on Facebook (244 and 242 candidates respectively) while the Leftwing party has the lowest number (109 candidates). The average number of followers on Facebook is 10,643. Again, there are considerable differences across parties. Candidates from AfD have the highest number of average followers, around 26,156 while CDU has the lowest with only 5528 followers on average. Certain prominent candidates stand out as outliers in terms of their number of followers and/or posting intensity. For example, Sahra Wagenknecht from the Leftwing party has the highest number of fans on Facebook, followed by Gregor Gysi (same party) and Alice Weidel (AfD). Sahra Wagenknecht is among the most followed politicians on Twitter, with her follower count surpassed only by Karl Lauterbach, a prominent member of the SPD. Similarly, a number of candidates post noticeably more often than others. For example, Raimond Scheirich from AfD posted almost double (at 1114) that of Sina Beckmann of the Green party on Facebook, where the average number of posts is 418. Finally, we collected the relevant data on the Facebook pages of the party chapters of the six major parties for all 16 states in the relevant period of time. While not all candidates entertain a Facebook page, we find such pages for all six parties in all states. In the six weeks prior to the election, we collected data on 6489 posts. The party chapters’ activities on Facebook are similar to those of the respective parties’ candidates on FB, with the highest posts from CDU, followed by AfD and then FDP. Based on the data described above, we calculate the cumulated product of posts by (candidates of) party p and the number of their followers (fans) in state s on day t. Normalizing by the total population in state s yields our measure for primary outreach of party p in state s on day t: Fig. 1. Note: This figure displays the weekly aggregated voting intentions of survey respondents as recorded by Forsa. Respondents were asked: “Which party would you vote for if the federal election were held next Sunday?” Weeks 33–38 correspond to the six weeks leading up to the 2021 federal election, spanning from August 16 to September 26. Table 1 Predicted voting outcomes according to survey responses and vote shares in the election. Party Pr(party | selected) Actual Vote Shares Mean Standard Deviation AfD 0.0415433 0.1211204 0.079 CDU 0.2126096 0.2142361 0.185 FDP 0.0911809 0.1204675 0.081 Grüne 0.1984444 0.2121326 0.113 Linke 0.0515184 0.1157282 0.079 SPD 0.2333933 0.1889009 0.197 Don’t know 0.145256 0.0861365 – No vote 0.0260542 0.0552564 0.234 Note: The table shows the predicted probabilities of voting for the six major parties + two responses (don’t know and no vote) calculated used on the responses of all respondents included in the regressions later. The weights provided by FORSA are not used here. The last column shows the actual voting outcomes for the six parties and the share of non-voters. 10 We used FanpageKarma– a commercial social media analytics tool that offers access to public pages and information of various social media users. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 6 Outreachpst =100 ∑n 1 postspst ⋅ fanspst populations (4) This indicator is calculated separately for each of the three channels of social media campaigns, namely Twitter candidates, Facebook candidates, and Facebook party chapters. Each measure captures the maximum possible share of the population the (can- didates of) party p in state s reach directly with their posts on the particular channel on day t. The primary outreach differs massively across parties, days, and channels (see Table 2). Moreover, there are sizeable differences in the average outreach of parties across the 16 German states (see Fig. 2a–c). Our empirical strategy rests precisely on this variation in campaigning intensity across time, state, platform, and party. 4.2. Empirical strategy In the analysis below, we test whether the voting intention voiced by an individual i in the opinion poll by FORSA in state s on day t depends on the intensity of campaigning of this party’s chapter and candidates in state s on social media in the days before. Using the weighted utility function (expression 1 above), the utility of respondent i in state s from voting for party p on day t is given by the following expression11 U(v = p)ist =Xpst ⋅ β+(ziA) + εipst (5) with p = sub-index of party. i = sub-index for the individual t = sub-index of time s = sub-index of state. The expected utility voters assign to voting for party p on day t is not observable. However, we can observe the discrete choice in favor of one party as stated in the FORSA survey. According to expression (5), the choice for party p is driven by party p’s campaign outreach – relative to the outreach of the competing parties. In addition, the respondents’ personal characteristics drive their re- sponses. The explanatory variables captured Xpst and zi are presented in Table 3. The matrix Xpst contains the alternative-specific explanatory variables – capturing the characteristics of party campaigns and further party-specific controls (see below). The coefficient vector β informs us about the impact of these variables. In our specific context, it is important to note that the effect of party campaigns changes neither their policy platforms nor the true competence of the candidates. However, they impact the perceived valence of certain issues and the perceived competence of party candidates. We assume that party campaigns are intended to change these perceptions in a way that increases the voters’ perceived utility from voting for a certain party and thus the probability of voting for this party. We test whether the intended effect emerges. In elections, all parties engage in social media campaigns. Consequently, the impact of a certain party p’s campaigning on voter i’s perceived utility depends on party p’s campaigning effort relative to the campaigning effort of all other parties. The larger the dif- ferences in exposure of voter i to campaigns of party p relative to other parties is, the larger the impact on the latter’s expected utility and thus probability to vote for party p. We use differences in primary outreach as a proxy for differences in voters’ exposure to political campaigning. More precisely, we assume the following: The probability that a randomly chosen voter i in state s is exposed to a campaigning message of party p in state s increases in the number of voters who directly receive this message. This is assumed to hold also across time. This implies equally strong spillovers across parties, states, and time on average. In this case, the primary outreach in expression (4) is an unbiased proxy for inter-party, inter-state, and inter-temporal differences in voters’ exposure to social media campaigns. In the baseline specification, we use a cumulated two-day lag of primary campaigning outreach – thereby focusing on the impact of the most recent posts or tweets. We restrict the analysis to respondents that state to vote for the six large parties plus the categories “I don’t know” and “I will not vote”. Other parties and independent candidates were not considered. This restriction is necessary because independent candidates cannot be grouped and compared across states. The same holds for most of the small parties. Next to differences in party campaigning, we control for important characteristics of the respondents (captured in matrix zi). Next to socio-demographic characteristics, it contains information about which party the respondent reports to have voted for in the previous election. The matrix A contains party-specific coefficients capturing the effects of these characteristics on the probability of voting for each party separately. Finally, we control for other news channels that are likely to influence citizens’ voting intentions. To this end, we collect data on the state-specific news programs offered by the public broadcaster (ARD) for all 16 German states in the early evening hours. These programs cover news on state-specific topics on a daily basis and have a considerable range. We include the content relating to the 2021 federal election as a potential co-founder in our regression. To this end, we construct two dummy variables that equal 1 if a certain party p is mentioned in the aforementioned news program of state s on day t. One variable captures all cases in which a major 11 For further details on this mixed model, see Cameron and Trivedi (2005): section 15.4 and Long and Freese (2014): 460–464. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 7 focus rests on the party or its candidates and thus the party appears in the heading of a story/clip. The other variable captures whether a party is mentioned in the keywords characterizing one clip/story. This data captures the inter-temporal, inter-regional differences in media coverage of the six major parties’ campaigns on state TV.12 It is part of the matrix Xpst. Table 4 shows that federal elections only play a minor role in the state-specific news format. On average, the six major parties were mentioned no more than once every two weeks. The main political discourse takes place at the federal level – in TV formats and newspapers with nationwide outreach. We control for this by using clustered standard errors. Moreover, we introduce two robustness checks that explicitly account for day- specific differences in the public discourse of the six major parties. We need a regression model that accounts for differences across choices – i.e. differences in the campaigning outreach across parties – as well as for differences across respondents. We chose a multinomial logit approach with alternative-specific constants. We prefer this model over the alternative mixed logit model for two main reasons: First, the major advantage of mixed logit models – the absence of the IIA assumption – is irrelevant here because the choice-set chosen model is not artificially reduced as in many stated choice experiments but lists all possible options. Second, the likelihood-ratio test for the baseline model indicates that modeling the cam- paigning variables as random variables is not preferred over modeling them as fixed while both models yield coefficient estimators for all relevant variables that are very similar. 5. Results Table 5 reports the results of our regressions in the form of odds ratios. It reports the coefficient estimators for β covering the campaigning intensity for the three above-mentioned channels and the party-specific media coverage on public news broadcasting. In the baseline model, these variables capture the activities of parties respectively their media coverage in the two days preceding the survey. All respondents’ personal characteristics shown in Table 3 are controlled for. The coefficient matrix A is reported in the ap- pendix. Standard errors are clustered at the state x date level. Next to this baseline specification, Table 5 reports a number of other models. In the second model, we account for the fact that the frequency with which a certain party is mentioned in the public political discourse on Twitter is a good proxy for its popularity at a certain time during the election campaign (see section 2). This effect may partly be driven by the mere exposure effect (e.g., Pfister et al., 2023). We introduce additional controls for the political communication of the public using the dataset collected by Aziz (2023). It comprises around 8.2 million Tweets that largely originate from non-candidate accounts. It provides an overview of the public conversation around the given election.13 Around 4.7 million of these tweets can be characterized as pro, neutral, or against at least one of the parties. The highest percentage of classified tweets are neutral (64 percent), while Tweets with a negative inclination to- wards parties (33 percent) dominate the pro-party Tweets (5 percent). Overall, CDU has the highest number of Tweets to its name (2.2 Table 2 Daily outreach of candidates and party chapters by platform and party. Party Mean Std. Dev. Min Max Twitter (candidates) CDU/CSU 12.539 36.305 0 327.596 SPD 26.296 68.232 0 550.784 FDP 5.126 8.389 0 65.955 Grüne 17.962 26.87 0 182.025 Linke 10.629 14.076 0 101.052 AfD 3.292 5.539 0 30.41 Facebook (candidates) CDU/CSU 1.835 1.65 0 9.757 SPD 1.53 1.898 0 23.698 FDP 0.959 1.465 0 6.527 Grüne 0.837 1.521 0 12.361 Linke 1.848 4.228 0 30.657 AfD 2.748 3.359 0 19.316 Facebook (party chapters) CDU/CSU 1.653 3.858 0 29.361 SPD 0.43 0.569 0 4.242 FDP 0.376 0.386 0 2.565 Grüne 0.268 0.39 0 3.42 Linke 0.462 0.544 0 5.533 AfD 2.608 2.499 0 16.604 Note: The table represents summary statistics of the daily primary outreach of each party across the different platforms. This outreach is of the campaigning carried out either by the candidates of the parties or their party chapters. 12 Furthermore, one may argue that the online campaigns may pick up content on other state-specific media. If this is the case, the coefficients do not necessarily capture the effects of campaigning via social media. We compare the content of these news programs to the content of the cam- paigning messages of parties on Facebook and Twitter. We find direct sharing of links to be extremely seldom. 13 Refer to Aziz (2023) for details about data collection strategy and procedure. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 8 million), followed by AfD (0.93 million). It is significant to note that the highest proportion of pro-party tweets refer to AfD while that of against-party tweets refer to CDU.14 To account for the public debate on Twitter, we include the lagged daily public traffic con- taining pro, neutral, or against-inclinations for the given parties. They are part of the matrix Xpst . In model 3, we replace the variables introduced in model 2 with day-fixed effects (as part of matrix zi) as an alternative way to account for inter-day differences in the public discourse. The fourth model excludes all respondents who were too young to have voted in the 2017 election and for whom we thus do not have information about their previous voting behavior. This reduces the effect of unobserved heterogeneity among respondents. In the fifth model, the variables in matrix Xpst – i.e. social media campaigning outreach and general media coverage – are calculated using activities of the directly preceding day only (not two days before). In the baseline specification, we find significant odds ratios above 1 for the candidates’ campaigning outreach on Twitter as well as Fig. 2. Average outreach by party, state, and platform Note: The maps show the average primary outreach of parties across the 16 German states. Outreach is calculated as the average number of posts multiplied by the average number of followers calculated by party and state (see also expression (4)) calculated over the entire period of obser- vation. Fig. 2a, b, and 2c refer to the outreach of candidates on Twitter, on Facebook, and the party chapters’ outreach on Facebook, respectively. 14 The Twitter public data was scraped on the Tweet level, that is any tweets that matched the collection criteria were recorded from the Twitter streaming API. Whereas, the Twitter candidates data was collected for the particular accounts of the candidates. Since the public data captured the political conversations, it is only credible that some of the candidates’ tweets appear in the public data as well. Most of the candidates have appeared in the public data, however, their tweets amount to only 0.56 percent of the public data. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 9 for the party chapters’ campaigning outreach on Facebook. This result remains stable for all other specifications reported in Table 5. Thus, we find strong support for our first hypothesis. The proxies for general media coverage and the general public debate (model 2) are not significant.15 The odds ratio for candidates’ campaigning on Twitter suggests that the effect size for this channel is very small. This is different for campaigning via the Facebook pages of the party chapters at the state level. Here, an increase in outreach by one percentage point by certain party p in state s implies that the probability that a respondent in state s chooses this party increases by 1.2 percent – other things equal. For the Christian Democrats with a baseline probability of roughly 30 percent, this translates into an increase of 0.36 percentage points. The fifth model suggests that the effect is even larger for an increase in posts on the day directly preceding the day of the survey. Hypothesis 2 states that the effect of social media campaigns is expected to be larger for the AfD than for other parties. Fig. 3 reports the party-specific marginal effects for the two significant campaigning variables. The hypothesis is clearly rejected. Instead for AfD, a higher marginal productivity of social media campaigns is estimated for CDU, SPD, and Green Party. Table 3 Explanatory variables. Individual characteristics Gender Dummy variable = = 1, if female, else 0 Age Ln(age) Income (stated in brackets) Dummy variable = = 1 for income > €4000, else 0 Level of education Dummy variable = = 1 for a university degree, else 0 Voting decision for the last federal election 2017 Dummy variables for • major six parties, • other parties • “I don’t remember”, • “I was too young to vote” • “I did not vote” Party specific attributes Lagged Twitter outreach of candidates of party p in state s in previous two days Outreach calculated according to expression (4) for days t-1 & t-2 Lagged Facebook outreach of candidates of party p in state s in previous two days Outreach calculated according to expression (4) for days t-1 & t-2 Lagged Facebook outreach of state chapters of party p in state s in previous two days Outreach calculated according to expression (4) for days t-1 & t-2 Public discourse on Twitter Lagged number of positive tweets on party p in state s in previous two days Number of positive tweets for days t-1 & t-2 Lagged number of negative tweets on party p in state s in previous two days Number of negative tweets for days t-1 & t-2 Lagged number of neutral tweets on party p in state s in previous two days Number of neutral tweets for days t-1 & t-2 Note: The table describes the explanatory variables used in the subsequent regression models. They are taken from the FORSA survey (upper part) or calculated using data collected online (lower part). Table 4 Daily media coverage of the parties. Party Mean Std. Dev. Min Max Discussed with major focus CDU/CSU 0.038 0.191 0 1 SPD 0.044 0.205 0 1 FDP 0.048 0.214 0 1 Grüne 0.065 0.247 0 1 Linke 0.04 0.196 0 1 AfD 0.042 0.201 0 1 Mentioned without major focus CDU/CSU 0.055 0.227 0 1 SPD 0.059 0.235 0 1 FDP 0.046 0.21 0 1 Grüne 0.078 0.268 0 1 Linke 0.05 0.219 0 1 AfD 0.048 0.214 0 1 Note: The table reports the summary statistics of the daily media coverage of the six major parties in state-specific news programs. A party was classified as being “in focus” if it appeared in the headlines of a news story, while mentions only in subsequent keywords were categorized as coverage “without major focus". 15 We run a number of additional robustness tests. For instance, we use viewing figures for the coverage of parties in the daily state-specific TV news formats to weigh the coverage just like we weigh the campaigning on social media by number of fans. The results do not change. We ran the baseline specification with any two of our three campaigning variables to check whether the effect of candidates’ campaigns on Twitter or the effect of party chapter campaigning on Facebook are independent of the other social media channels. The results of all these models are qualitatively identical to those reported above. Results are available upon request. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 10 Finally, we test whether the impact of social media campaigns may differ by the socio-demographic characteristics of the re- spondents. To this end, we calculate the marginal effects for the two significant campaigning variables by gender, age groups, and level of education (and parties). The marginal effects do not show any significant differences (Fig. 3). 5.1. Robustness check: the role of sentiment The arguments in section 3 refer primarily to the instrumental utility of voting – i.e. the expected utility resulting from the prospect of influencing future policy choices to one’s individual advantage (see expression 1). According to the theory of expressive voting, however, rational voters know that the instrumental utility of voting (at all and for a certain party) is negligible in most elections because they are very unlikely to be pivotal (e.g., Tyran, 2004; Hamlin and Jennings, 2019). Thus, citizens’ choice of candidates must be explained by the immediate utility from the act of voting. This so-called expressive utility results from the pleasure of expressing one’s view while it is not related to the consequences that the individual vote has on election outcomes and thus future policies (e.g. Hamlin and Jennings, 2019). The pleasure is closely related to sentiments and thus sentimental messages are more likely to trigger expressive motives than neutral messages. This implies that political campaigns are not restricted to neutral messages adhering to instrumental utility but specifically target expressive motives. For instance, priming for instrumental motives implies that the Green Party emphasizes the long-term damage from global warming by providing facts. Priming for expressive motives implies emotional messages - e.g. pointing at the extinction of species like the polar bear. For right-wing parties, priming implies emotional messages about crime – especially child abuse and rape. Like other messages, priming tailored towards expressive motives will spill from one particular platform to other social platforms as well as to offline spheres of communication – thus potentially affecting a large number of voters who themselves are not active on social media. In fact, emotionally intense content is likely to spread even farther and faster than other media (e.g., Gorodnichenko et al., 2021). Again, the fact that the original content can be spread directly contributes to the particular potential of social media in spreading emotionally loaded campaigns. Thus, we arrive at a sentiment-related version of Hypothesis 1. Accordingly, the probability that an individual i in this state votes for party p on day t increases in the sentimental intensity of campaign content sent by party p and its candidates in state s prior to day t. We rerun the models reported in Table 5 – albeit using sentiment-weighted measures of social media campaigns. They refer to the same campaigning messages of our three main outreach variables, but they weigh each post or tweet by the absolute value of its sentiment score (calculated by TextBlob (Loria, 2020)): Table 5 Regression results. VARIABLES 1 2 3 4 5 Lagged outreach of candidates on Facebook 0.999 0.999 0.999 0.996 ​ (0.00464) (0.00468) (0.00480) (0.00476) ​ Lagged outreach of candidates on Twitter 1.001** 1.001** 1.001** 1.001** ​ (0.000432) (0.000427) (0.000447) (0.000442) ​ Lagged outreach of party chapters on Facebook 1.011*** 1.012*** 1.012*** 1.012*** ​ (0.00301) (0.00312) (0.00203) (0.00312) ​ Lagged media coverage of parties (mention with focus) 1.004 1.004 1.004 0.998 ​ (0.0590) (0.0590) (0.0619) (0.0592) ​ Lagged media coverage of parties (any mention) 0.964 0.981 0.942 0.950 ​ (0.0475) (0.0476) (0.0449) (0.0474) ​ Lagged public traffic on parties (with pro inclination) on Twitter ​ 1.000*** ​ ​ ​ ​ (5.25e-07) ​ ​ ​ Lagged public traffic on parties (with against inclination) on Twitter ​ 1.000 ​ ​ ​ ​ (2.27e-08) ​ ​ ​ Lagged public traffic on parties (with neutral inclination) on Twitter ​ 1.000 ​ ​ ​ ​ (4.54e-08) ​ ​ ​ Single-day lagged outreach of candidates on Facebook ​ ​ ​ ​ 1.002 ​ ​ ​ ​ (0.00814) Single-day lagged outreach of candidates on Twitter ​ ​ ​ ​ 1.001 ​ ​ ​ ​ (0.000725) Single-day lagged outreach of party chapters on Facebook ​ ​ ​ ​ 1.021*** ​ ​ ​ ​ (0.00578) Single-day lagged media coverage of parties (mention with focus) ​ ​ ​ ​ 1.058 ​ ​ ​ ​ (0.0774) Single-day lagged media coverage of parties (any mention) ​ ​ ​ ​ 0.922 ​ ​ ​ ​ (0.0533) Observations 90,328 90,328 90,328 85,072 97,168 Note: We use a multinomial logit model with alternative-specific constants to predict the response of 13.035 subjects to the famous “Sunday- question”: “Imagine that the federal election would be held upcoming Sunday, how would you vote?” in the FORSA-survey. The regressions control for personal characteristics. Standard errors are clustered at the state level. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 11 Outreachsent pst =100 ∑n 1 postspst ⋅ fanspst ⋅ abs ( sentimentpst ) populations (6) The sentiment-weighted outreach is zero for neutral messages and positive for messages with positive or negative sentiment. It is intended to capture the degree to which campaigning refers to expressive motives. In all six models, we find a highly significant odds ratio above unity for party chapters’ outreach on Facebook only (see Table 6). Again, the estimates are larger than 1 while the odds ratios point to a much smaller effect size. An increase in sentiment-weighed outreach by one percentage point by certain party p in state s implies that the probability that a respondent in state s chooses this party increases by 0.1 percent – other things equal. For the Christian Democrats with a baseline probability of roughly 30 percent, this translates into an increase of 0.03 percentage points. 6. Discussion In sum, our results support our first hypothesis: Parties’ and their candidates’ campaigning on social media has an impact on the probability that respondents in the survey state to vote for this party. This holds in a number of different specifications and for the outreach with and without sentiment-weighing. The effect size for sentiment-weighted outreach is negligible while the outreach without sentiment-weighing yields a small but non-negligible effect size – especially in times with close elections and multiple party coalitions. Surprisingly, we find no support for our second hypothesis according to which the effect of campaigns by the right-wing party AfD is larger than for other parties. The main result is in line with the notion underlying the literature reviewed in section 2 that campaigns on social media increase support among voters. It suggests that social media campaigns are not just another tool of mass campaigning that has little if any effect (e.g., Kalla and Broockman, 2018) but may have properties similar to those of face-to-face campaigns that proved to be more effective. Like in face-to-face campaigns, the initial recipients of the social media campaigns actively chose to be subject to the campaign by becoming followers/friends of the party chapters pages. Microtargeting entails another similarity to face-to-face campaigning in that messages can be targeted more specifically to the individual recipient. Beyond the initial recipient, voters exposed to the campaign content learn about it through their personal networks and may therefore perceive it to come with a personal recommendation and thus be more salient than a campaign ad seen on the web, TV, or in a newspaper. Given the limited initial outreach of social media campaigns when measured by the initial recipients, our results support the notion that these campaigns must generate substantial spillovers. While our results are consistent with the existence of filter bubbles and echo chambers, they do not necessarily imply strong filter bubbles or echo chambers. Instead campaigning may be effective even if social networks are not highly fragmented and microtargeting only works for the initial recipients. In this case, the effect is driven by the fact that secondary recipients who receive a message not aligned with their political preferences may pay more attention to it if arrives through their personal networks than if takes the form of anonymous mass campaigning. Thus, our results cannot be taken as indirect evidence that fragmentation on social media is stronger than offline. As for Kelm et al. (2023), our results do not support the widespread notion that there is a causal link between the emergence of social media and the popularity of populist right-wing parties. On the contrary, they suggest that it is particularly the established parties with a less populistic approach that benefit from social media campaigns. The fact that we find the effect of the party chapters’ campaigns to be smaller for the right-wing extremist AfD than for other parties is not in line with the results or implications of previous studies. One possible explanation is that the spillovers for AfD campaigns are smaller than for other parties. This could be caused by the fact that many supporters of AfD are less likely to openly admit that they prefer this party than supporters of other parties because they fear social sanctions. This explanation is, however, ad hoc. Further research is needed to deepen our understanding of inter-party differences in the effectiveness of social media campaigning. Our study has a number of limitations. First, we assume that the primary outreach defined in expression (4) is a good proxy for the Fig. 3. Party-specific marginal effects of campaigning Note: The plots represent the marginal effects of campaigning on Twitter by the political candidates and on Facebook by the party chapters. They are computed for each political party to illustrate how the campaigning effect varies across the different parties. All the confidence intervals shown are at the 95 % level. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 12 (caption on next page) A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 13 effective outreach of parties’ campaigns. This assumption cannot be tested. However, the fact that the campaigning outreach yields significant marginal effects for all parties suggests that spillovers are sizeable for all parties. Second, we cannot differentiate between respondents who use social media actively and are thus more likely to be exposed to social media campaigns directly and those respondents whose exposure is restricted to offline spillovers because they are not active on social media themselves. The fact that we do not find any differences in effect sizes by demographic characteristics does not solve this issue. However, it suggests that offline spillovers are substantial. Third, we do not explicitly control for the role of the traditional press. The possible impact of nationwide newspapers is captured in the second specification where we introduce day-fixed-effects to capture any common shocks that have a nationwide impact. More- over, we test whether the impact of social media differs across respondents with different levels of education. This is a good proxy of the degree to which voters adhere to other sources of information – especially newspapers and magazines. We do not find a significant impact of respondents’ education on the impact of social media campaigns. We did not explicitly control for the more than 220 local newspapers in Germany. They are, however, very similar in the coverage of non-local news. This is so because most of them buy the so- called “Mantel” i.e. the pages coating the local part of the newspaper from news agencies or simply present the text provided by AP and the German equivalent dpa (Deutsche Presse-Agentur). The impact of this content is captured by day-fixed effects. Finally, we account for any regional differences in major news by including the lagged media coverage of parties in public television (see the fourth and fifth attributes included in all regression models). We believe that the major shocks to the popularity of parties or candidates are captured by these control variables. An additional concern relates to endogeneity issues (e.g., Sabatini, 2024): The treatment – large-scale campaigning on social media – is not randomly assigned but deliberately chosen by candidates and parties to achieve a certain aim. Rational parties are likely to focus their campaigns in places and times where the marginal productivity in terms of votes is high. This raises the concern that the significant β-coefficients just capture the effect that parties campaign more intensely when/where the electorate is more susceptive to campaigning. The concern would, of course, be valid if we used an approach that estimates the correlation between one party’s campaigning only. For two reasons, this concern is less severe for our multinomial choice architectures. First, the β-coefficients in this Fig. 4. Marginal effects by socio-demographic characteristics Note: Fig. 4a and b illustrate the marginal effects of socio-demographic characteristics (age, gender, and higher education) for each political party when campaigning is carried out on Twitter and on Facebook (chapters) respectively. The overlapping confidence intervals (calculated at 95 %) indicate no statistically significant differences in these characteristics across the parties. Table 6 Regression results with sentiment. VARIABLES 1 2 3 4 5 Lagged outreach of candidates on Facebook (sentiment-weighted) 1.000 1.000 1.000 1.000 ​ (0.000242) (0.000240) (0.000260) (0.000250) ​ Lagged outreach of candidates on Twitter (sentiment-weighted) 1.000* 1.000* 1.000** 1.000* ​ (2.00e-05) (1.98e-05) (1.71e-05) (2.05e-05) ​ Lagged outreach of party chapters on Facebook (sentiment-weighted) 1.001*** 1.001*** 1.001*** 1.001*** ​ (0.000195) (0.000199) (0.000141) (0.000200) ​ Lagged media coverage of parties (mention with focus) 1.003 1.003 1.007 0.997 ​ (0.0584) (0.0583) (0.0637) (0.0591) ​ Lagged media coverage of parties (any mention) 0.962 0.978 0.938 0.948 ​ (0.0485) (0.0484) (0.0454) (0.0484) ​ Lagged public traffic on parties (with pro inclination) on Twitter ​ 1.000*** ​ ​ ​ ​ (5.26e-07) ​ ​ ​ Lagged public traffic on parties (with against inclination) on Twitter ​ 1.000 ​ ​ ​ ​ (2.29e-08) ​ ​ ​ Lagged public traffic on parties (with neutral inclination) on Twitter ​ 1.000 ​ ​ ​ ​ (4.55e-08) ​ ​ ​ Single-day lagged outreach of candidates on Facebook (sentiment-weighted) ​ ​ ​ ​ 1.000 ​ ​ ​ ​ (0.000434) Single-day lagged outreach of candidates on Twitter (sentiment-weighted) ​ ​ ​ ​ 1.000 ​ ​ ​ ​ (2.94e-05) Single-day lagged outreach of party chapters on Facebook (sentiment-weighted) ​ ​ ​ ​ 1.001*** ​ ​ ​ ​ (0.000346) Single-day lagged media coverage of parties (mention with focus) ​ ​ ​ ​ 1.059 ​ ​ ​ ​ (0.0771) Single-day lagged media coverage of parties (any mention) ​ ​ ​ ​ 0.915 ​ ​ ​ ​ (0.0539) Observations 90,328 90,328 90,328 85,072 97,168 Note: This table reports the results of the multinomial logit model with alternative specific constants when the focal outreach variables are weighted by absolute sentiment scores. The regressions control for personal characteristics as in the standard case. Standard errors are clustered at the state level. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 14 model capture the effect of a certain party’s campaigning relative to the campaigning of all its major competitors. To see why this helps solve the endogeneity problem, consider the following case: Assume that a certain region is characterized by a high share of indecisive voters. In this case, all parties are likely to focus their campaigning effort in this region and the campaigns will neutralize each other – similar to a price battle where all firms try to attract additional customers by lowering prices. The fact that we still observe significant β-coefficients informs us that social media campaigning is productive at the margin: If a certain party has a higher campaigning outreach than others, the impact on the probability that a respondent voices the intention to vote for this party is positive. Second, we make use of inter-daily differences in campaigning. It is very unlikely that these are driven by inter-daily differences in respondents’ receptivity for campaigns that are observable to parties and their candidates. It is reasonable to assume that a large number of indecisive voters may attract extra campaigning efforts, especially among those parties whose candidates are part of a close race for the direct seat in parliament determined by the first vote (so-called “Erststimme”). This implies that large parties whose direct candidates have the best chances of winning are more responsive to perceive differences in the number of indecisive voters across districts. As we observe behavior at the state and not the district level, differences in the in- tensity of political competition for direct seats at the district level are unlikely to drive our results. This is especially true for the social media campaigns of state chapters – the form of campaigning for which we find a significant effect. Nevertheless, our empirical model does not fully rule out the endogeneity problem. However, we have to account for the fact that - from an ethical point of view – it is not permissible to run large-scale experimental interventions that – through their spillovers – have the potential to change election outcomes if social media campaigns are effective. Thus, studies on the topic must always rely on happenstance data and thus potentially suffer from the above-mentioned endogeneity problem. Given this unavoidable restriction, our study makes a significant contribution to the literature. 7. Conclusion Essential properties of social media nourish the notion that political campaigns on social media have the potential to change voter behavior in offline elections. First, social media campaigns spread through the networks of those citizens who initially receive them – both online and offline. Given the speed and the negligible costs, social media campaigns are likely to impact an audience multiple times larger than the group of original recipients. Second, the existence of filter bubbles and echo chambers plus the potential of micro- targeting allows parties to address relevant groups of voters much more sharply than this can be achieved with campaigning on traditional media. So far, however, the number of studies that explicitly test whether social media campaigns impact offline voter behavior is limited. We provide a study that helps to fill this gap. We address possible endogeneity issues by exploiting inter-party, inter-temporal, and inter-regional differences in the intensity of party campaigns during the federal election to the German Bundestag in 2021. We observe the campaigning activities of the six major parties and their candidates on Twitter and Facebook and relate them to the responses of more than 14,000 respondents in a large-scale representative survey. Next to voting intentions, this survey elicits socio-demographic characteristics and the respondents’ voting behavior in previous elections. This enables us to control for unobserved heterogeneity at the individual level. We use a multinomial logit model with alternative-specific constants to test whether the respondents’ voting intentions mentioned in the survey are driven by the intensity of the parties’ social media campaigning. We find evidence in favor of a positive effect of social media campaigns. While the existing literature suggests that the productivity of these campaigns is especially large for (right-wing) populist countries, we find the opposite to be true. Especially Christian Democrats, Social Democrats, and the Green party benefit from social media campaigns. Effect sizes are not particularly large, yet large enough to make social media campaigns an important tool in elections where the vote margins between competing parties are small and coalition governments often require more than two parties. In the future, more studies are needed to see whether our results hold in other contexts. Given the fact that field experiments are ethically not permissible in this field, these studies will – like our study – have to work with happenstance data. CRediT authorship contribution statement Abeer Ibtisam Aziz: Writing – review & editing, Writing – original draft, Visualization, Methodology, Formal analysis, Data curation, Conceptualization. Ivo Bischoff: Writing – review & editing, Writing – original draft, Visualization, Supervision, Method- ology, Investigation, Formal analysis, Data curation, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Appendix Matrix A: Effects of individual characteristics on the probability to vote for each party. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 15 VARIABLES AfD CDU FDP Grüne Linke Don’t Know No Vote Gender 0.744* 1.002 0.740*** 1.197** 0.942 1.468*** 0.914 (0.121) (0.0774) (0.0726) (0.0867) (0.133) (0.111) (0.127) Age 0.976*** 1.003 0.978*** 0.965*** 0.981*** 0.991*** 1.005 (0.00462) (0.00250) (0.00286) (0.00236) (0.00401) (0.00233) (0.00417) High income 0.757 1.177* 1.337*** 1.121 0.712** 0.732*** 0.943 (0.137) (0.0999) (0.143) (0.0989) (0.107) (0.0678) (0.182) High education 0.625** 1.010 1.210* 1.638*** 1.439*** 0.996 0.510*** (0.117) (0.0815) (0.125) (0.124) (0.182) (0.0815) (0.102) Voted for CDU in 2017 1.801** 4.665*** 1.761*** 0.775* 0.232*** 0.738** 0.495*** (0.481) (0.603) (0.258) (0.112) (0.0604) (0.0904) (0.130) Voted for SPD in 2017 0.277*** 0.0841*** 0.168*** 0.407*** 0.188*** 0.136*** 0.0663*** (0.0904) (0.0149) (0.0319) (0.0535) (0.0445) (0.0191) (0.0234) Voted for FDP in 2017 4.974*** 1.964*** 13.41*** 1.052 0.171*** 1.074 1.033 (1.731) (0.404) (2.637) (0.230) (0.0892) (0.210) (0.407) Voted for Grüne in 2017 0.156*** 0.279*** 0.397*** 4.890*** 0.806 0.256*** 0.229*** (0.103) (0.0542) (0.0839) (0.619) (0.178) (0.0378) (0.102) Voted for Linke in 2017 1.912 0.0197*** 0.411*** 1.419* 11.18*** 0.427*** 0.527 (0.810) (0.0200) (0.133) (0.288) (2.196) (0.0899) (0.217) Voted for AfD in 2017 137.1*** 1.180 2.547*** 0.121*** 1.424 0.674 1.414 (48.39) (0.392) (0.877) (0.0832) (0.577) (0.216) (0.741) Voted for another party in 2017 2.127e+07*** 0.629** 0.627 0.726 0.608*** 1.372e+06*** 0.644 (2.325e+07) (0.138) (0.183) (0.152) (0.104) (1.199e+06) (0.189) Did not vote in 2017 4.260*** 0.625 1.619 1.617* 1.624 1.368 0.873 (2.047) (0.227) (0.589) (0.456) (0.711) (0.346) (0.507) Too young to vote in 2017 5.237*** 1.578** 0.589** 0.603** 0.455*** 0.773 8.339*** (1.533) (0.313) (0.143) (0.130) (0.139) (0.141) (1.936) Observations 90,328 90,328 90,328 90,328 90,328 90,328 90,328 Note: The table reports the coefficients and their standard errors for the estimated effect of the individual characteristics of the respondents on their vote choice for each party. In addition to the socio-demographic characteristics of the respondents, their past voting behavior is also accounted for. Statistical significance is denoted by *, **, and *** for the 10 %, 5 %, and 1 % levels, respectively. Data availability Data will be made available on request. References Aldrich, J.H., Gibson, R.K., Cantijoch, M., Konitzer, T., 2016. Getting out the vote in the social media era. Party Polit. 22 (2), 165–178. https://doi.org/10.1177/ 1354068815605304. Aridor, G., Jiménez Durán, R., Levy, R., Ro’ee, Song, L., 2024. The economics of social media. SSRN Electron. J. https://doi.org/10.2139/ssrn.4708840. Advance online publication. Aziz, A.I., 2023. Social Bots’ Role in Online Political Communication: Evidence from German Federal Election 2021 (MAGKS Joint Discussion Paper Series in Economics 09-2023). Philipps-University Marburg, School of Business and Economics, Marburg. https://www.econstor.eu/handle/10419/278433. Azzimonti, M., Fernandes, M., 2018. Social media networks, fake news, and polarization. https://doi.org/10.3386/w24462. Barberá, P., 2014. How social media reduces mass political polarization. Evidence from Germany, Spain, and the US. http://pablobarbera.com/static/barbera_ polarization_apsa.pdf. Bene, M., 2018. Post shared, vote shared: Investigating the link between facebook performance and electoral success during the Hungarian general election campaign of 2014. Journal. Mass Commun. Q. 95 (2), 363–380. https://doi.org/10.1177/1077699018763309. Bond, R.M., Fariss, C.J., Jones, J.J., Kramer, A.D.I., Marlow, C., Settle, J.E., Fowler, J.H., 2012. A 61-million-person experiment in social influence and political mobilization. Nature 489 (7415), 295–298. https://doi.org/10.1038/nature11421. Bossetta, M., 2018. The digital architectures of social media: comparing political campaigning on facebook, twitter, Instagram, and snapchat in the 2016 U.S. Election. Journal. Mass Commun. Q. 95 (2), 471–496. https://doi.org/10.1177/1077699018763307. Boulianne, S., Theocharis, Y., 2020. Young people, digital media, and engagement: a meta-analysis of research. Soc. Sci. Comput. Rev. 38 (2), 111–127. https://doi. org/10.1177/0894439318814190. Bright, J., Hale, S., Ganesh, B., Bulovsky, A., Margetts, H., Howard, P [Phil, 2020. Does campaigning on social media make a difference? Evidence from candidate use of twitter during the 2015 and 2017 U.K. Elections. Commun. Res. 47 (7), 988–1009. https://doi.org/10.1177/0093650219872394. Broockman, D.E., Green, D.P., 2014. Do online advertisements increase political candidates’ name recognition or favorability? Evidence from randomized field experiments. Polit. Behav. 36 (2), 263–289. https://doi.org/10.1007/s11109-013-9239-z. Cameron, A.C., Trivedi, P.K., 2005. Microeconometrics: Methods and Applications. Cambridge University Press. Campante, F., Durante, R., Sobbrio, F., 2018. Politics 2.0: the multifaceted effect of broadband internet on political participation. J. Eur. Econ. Assoc. 16 (4), 1094–1136. https://doi.org/10.1093/jeea/jvx044. Cantoni, E., Pons, V., 2021. Do interactions with candidates increase voter support and participation? Experimental evidence from Italy. Econ. Polit. 33 (2), 379–402. https://doi.org/10.1111/ecpo.12167. Cinelli, M., Francisci Morales, G. de, Galeazzi, A., Quattrociocchi, W., Starnini, M., 2021. The echo chamber effect on social media. Proc. Natl. Acad. Sci. U. S. A 118 (9), e2023301118. https://doi.org/10.1073/pnas.2023301118. Conover, M., Ratkiewicz, J., Francisco, M., Goncalves, B., Menczer, F., Flammini, A., 2011. Political polarization on twitter. Proceed. Int. AAAI Confer. Web Social Media 5 (1), 89–96. https://doi.org/10.1609/icwsm.v5i1.14126. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 16 https://doi.org/10.1177/1354068815605304 https://doi.org/10.1177/1354068815605304 https://doi.org/10.2139/ssrn.4708840 https://www.econstor.eu/handle/10419/278433 https://doi.org/10.3386/w24462 http://pablobarbera.com/static/barbera_polarization_apsa.pdf http://pablobarbera.com/static/barbera_polarization_apsa.pdf https://doi.org/10.1177/1077699018763309 https://doi.org/10.1038/nature11421 https://doi.org/10.1177/1077699018763307 https://doi.org/10.1177/0894439318814190 https://doi.org/10.1177/0894439318814190 https://doi.org/10.1177/0093650219872394 https://doi.org/10.1007/s11109-013-9239-z http://refhub.elsevier.com/S0176-2680(25)00045-X/sref14 https://doi.org/10.1093/jeea/jvx044 https://doi.org/10.1111/ecpo.12167 https://doi.org/10.1073/pnas.2023301118 https://doi.org/10.1609/icwsm.v5i1.14126 Coppock, A., Guess, A., Ternovski, J., 2016. When treatments are tweets: a network mobilization experiment over twitter. Polit. Behav. 38 (1), 105–128. https://doi. org/10.1007/s11109-015-9308-6. Cox, G.W., Fiva, J.H., King, M.-E.M., 2024. Bound by borders: voter mobilization through social networks. Br. J. Polit. Sci. 1–19. https://doi.org/10.1017/ S0007123424000164. Denter, P., 2020. Campaign contests. Eur. Econ. Rev. 127, 103459. https://doi.org/10.1016/j.euroecorev.2020.103459. Dimitrova, D.V., Shehata, A., Strömbäck, J., Nord, L.W., 2014. The effects of digital media on political knowledge and participation in election campaigns. Commun. Res. 41 (1), 95–118. https://doi.org/10.1177/0093650211426004. Dubois, E., Blank, G., 2018. The echo chamber is overstated: the moderating effect of political interest and diverse media. Inf. Commun. Soc. 21 (5), 729–745. https:// doi.org/10.1080/1369118X.2018.1428656. Falck, O., Gold, R., Heblich, S., 2014. E-Lections: voting behavior and the internet. Am. Econ. Rev. 104 (7), 2238–2265. https://doi.org/10.1257/aer.104.7.2238. forsa, B., 2022. Forsa-bus 2021. https://doi.org/10.4232/1.13957. Fujiwara, T., Müller, K., Schwarz, C., 2023. The effect of social media on elections: evidence from the United States. J. Eur. Econ. Assoc. https://doi.org/10.1093/ jeea/jvad058. Article jvad058. Advance online publication. Garz, M., Sörensen, J., Stone, D.F., 2020. Partisan selective engagement: evidence from facebook. J. Econ. Behav. Organ. 177, 91–108. https://doi.org/10.1016/j. jebo.2020.06.016. Gavazza, A., Nardotto, M., Valletti, T., 2019. Internet and politics: evidence from U.K. Local elections and local government policies. Rev. Econ. Stud. 86 (5), 2092–2135. https://doi.org/10.1093/restud/rdy028. Gorodnichenko, Y., Pham, T., Talavera, O., 2021. Social media, sentiment and public opinions: evidence from #Brexit and #USElection. Eur. Econ. Rev. 136, 103772. https://doi.org/10.1016/j.euroecorev.2021.103772. Grossman, G.M., Helpman, E., 2023. Electoral competition with fake news. Eur. J. Polit. Econ. 77, 102315. https://doi.org/10.1016/j.ejpoleco.2022.102315. Halberstam, Y., Knight, B., 2016. Homophily, group size, and the diffusion of political information in social networks: evidence from Twitter. J. Publ. Econ. 143, 73–88. https://doi.org/10.1016/j.jpubeco.2016.08.011. Hamlin, A., Jennings, C., 2019. Expressive voting. In: Congleton, R.D., Grofman, B., Voigt, S., Hamlin, A., Jennings, C. (Eds.), The Oxford Handbook of Public Choice, vol. 1. Oxford University Press, pp. 332–350. https://doi.org/10.1093/oxfordhb/9780190469733.013.16. Hoferer, M., Böttcher, L., Herrmann, H.J., Gersbach, H., 2020. The impact of technologies in political campaigns. Phys. Stat. Mech. Appl. 538, 122795. https://doi. org/10.1016/j.physa.2019.122795. Howard, P.N., Woolley, S., Calo, R., 2018. Algorithms, bots, and political communication in the US 2016 election: the challenge of automated political communication for election law and administration. J. Inf. Technol. Polit. 15 (2), 81–93. https://doi.org/10.1080/19331681.2018.1448735. Jamieson, K.H., Cappella, J.N., 2008. Echo Chamber: Rush Limbaugh and the Conservative Media Establishment. Oxford University Press. Jungherr, A., 2013. Tweets and votes, a special relationship. In: Proceedings of the 2nd Workshop on Politics, Elections and Data. ACM. https://doi.org/10.1145/ 2508436.2508437. Kalla, J.L., Broockman, D.E., 2018. The minimal persuasive effects of campaign contact in general elections: evidence from 49 field experiments. Am. Polit. Sci. Rev. 112 (1), 148–166. https://doi.org/10.1017/S0003055417000363. Keller, T.R., Klinger, U., 2019. Social bots in election campaigns: theoretical, empirical, and methodological implications. Polit. Commun. 36 (1), 171–189. https:// doi.org/10.1080/10584609.2018.1526238. Kelm, O., Angenendt, M., Poguntke, T., Rosar, U., 2023. Which candidates benefit from social media? An analysis of the 2021 German federal election. Elect. Stud. 86, 102701. https://doi.org/10.1016/j.electstud.2023.102701. Krasa, S., Polborn, M., 2010. The binary policy model. J. Econ. Theor. 145 (2), 661–688. https://doi.org/10.1016/j.jet.2009.11.005. Leung, B.T.K., Yildirim, P., 2020. Competition, politics, & social media. http://arxiv.org/pdf/2012.03327v1. Levy, R., Ro’ee, 2021. Social media, news consumption, and polarization: evidence from a field experiment. Am. Econ. Rev. 111 (3), 831–870. https://doi.org/ 10.1257/aer.20191777. Liberini, F., Redoano, M., Russo, A., Cuevas, A., Cuevas, R., 2025. Politics in the facebook era evidence from the 2016 us presidential elections. Eur. J. Polit. Econ., 102641 https://doi.org/10.1016/j.ejpoleco.2025.102641. Liu, Z., 2024. Information and polarization. J. Econ. Behav. Organ. 226, 106686. https://doi.org/10.1016/j.jebo.2024.106686. Long, J.S., Freese, J., 2014. Regression models for categorical dependent variables using Stata. In: J. Scott Long, Jeremy Freese, third ed. StataCorp LP. Loria, S., 2020. textblob Documentation. https://buildmedia.readthedocs.org/media/pdf/textblob/dev/textblob.pdf. Mutascu, M., Strango, C., Turcu, C., 2025. Online social media and populism in Europe. Eur. J. Polit. Econ. 86, 102619. https://doi.org/10.1016/j. ejpoleco.2024.102619. Neudert, L., Kollanyi, B., Howard, P.[P., 2017. Junk news and bots during the German parliamentary election: what are German voters sharing over twitter? Computational Propaganda Project. https://ora.ox.ac.uk/objects/uuid:33f7d85c-a25e-4796-aebf-a0e601fc84c1. Norris, P., Curtice, J., 2008. Getting the message out: a two-step model of the role of the internet in campaign communication flows during the 2005 British general election. J. Inf. Technol. Polit. 4 (4), 3–13. https://doi.org/10.1080/19331680801975359. Papakyriakopoulos, O., Hegelich, S., Shahrezaye, M., Serrano, J.C.M., 2018. Social media and microtargeting: political data processing and the consequences for Germany. Big Data & Society 5 (2), 205395171881184. https://doi.org/10.1177/2053951718811844. Petrova, M., Sen, A., Yildirim, P., 2021. Social media and political contributions: the impact of new technology on political competition. Manag. Sci. 67 (5), 2997–3021. https://doi.org/10.1287/mnsc.2020.3740. Pfister, R., Schwarz, K.A., Holzmann, P., Reis, M., Yogeeswaran, K., Kunde, W., 2023. Headlines win elections: mere exposure to fictitious news media alters voting behavior. PLoS One 18 (8), e0289341. https://doi.org/10.1371/journal.pone.0289341. Potrafke, N., Roesel, F., 2025. Online versus offline: which networks spur protests? J. Econ. Behav. Organ. 229, 106856. https://doi.org/10.1016/j.jebo.2024.106856. Poy, S., Schüller, S., 2020. Internet and voting in the social media era: evidence from a local broadband policy. Res. Pol. 49 (1), 103861. https://doi.org/10.1016/j. respol.2019.103861. Quinlan, S., Gummer, T., Roßmann, J., Wolf, C., 2018. ‘Show me the money and the party!’ – variation in Facebook and Twitter adoption by politicians. Inf. Commun. Soc. 21 (8), 1031–1049. https://doi.org/10.1080/1369118X.2017.1301521. Rotesi, T., 2019. The impact of twitter on political participation. Sabatini, F., 2024. The behavioral, economic, and political impact of the Internet and social media: empirical challenges and approaches. J. Econ. Surv. https://doi. org/10.1111/joes.12629. Article joes.12629. Advance online publication. Sältzer, M., Stier, S., Bäuerle, J., Blumenberg, M., Mechkova, V., Pemstein, D., Seim, B., Wilson, S., 2021. Twitter-Accounts der Kandidierenden zur Bundestagswahl 2021 (GLES). https://doi.org/10.4232/1.13790. Schaub, M., Morisi, D., 2020. Voter mobilisation in the echo chamber: broadband internet and the rise of populism in Europe. Eur. J. Polit. Res. 59 (4), 752–773. https://doi.org/10.1111/1475-6765.12373. Schmitt-Beck, R., Mackenrodt, C., 2010. Social networks and mass media as mobilizers and demobilizers: a study of turnout at a German local election. Elect. Stud. 29 (3), 392–404. https://doi.org/10.1016/j.electstud.2010.03.011. Schroeder, E., Stone, D.F., 2015. Fox News and political knowledge. J. Publ. Econ. 126, 52–63. https://doi.org/10.1016/j.jpubeco.2015.03.009. Shao, C., Ciampaglia, G.L., Varol, O., Yang, K.-C., Flammini, A., Menczer, F., 2018. The spread of low-credibility content by social bots. Nat. Commun. 9 (1), 4787. https://doi.org/10.1038/s41467-018-06930-7. Shmargad, Y., Sanchez, L., 2022. Social media influence and electoral competition. Soc. Sci. Comput. Rev. 40 (1), 4–23. https://doi.org/10.1177/ 0894439320906803. Stier, S., Bleier, A., Lietz, H., Strohmaier, M., 2018. Election campaigning on social media: politicians, audiences, and the mediation of political communication on facebook and twitter. Polit. Commun. 35 (1), 50–74. https://doi.org/10.1080/10584609.2017.1334728. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 17 https://doi.org/10.1007/s11109-015-9308-6 https://doi.org/10.1007/s11109-015-9308-6 https://doi.org/10.1017/S0007123424000164 https://doi.org/10.1017/S0007123424000164 https://doi.org/10.1016/j.euroecorev.2020.103459 https://doi.org/10.1177/0093650211426004 https://doi.org/10.1080/1369118X.2018.1428656 https://doi.org/10.1080/1369118X.2018.1428656 https://doi.org/10.1257/aer.104.7.2238 https://doi.org/10.4232/1.13957 https://doi.org/10.1093/jeea/jvad058 https://doi.org/10.1093/jeea/jvad058 https://doi.org/10.1016/j.jebo.2020.06.016 https://doi.org/10.1016/j.jebo.2020.06.016 https://doi.org/10.1093/restud/rdy028 https://doi.org/10.1016/j.euroecorev.2021.103772 https://doi.org/10.1016/j.ejpoleco.2022.102315 https://doi.org/10.1016/j.jpubeco.2016.08.011 https://doi.org/10.1093/oxfordhb/9780190469733.013.16 https://doi.org/10.1016/j.physa.2019.122795 https://doi.org/10.1016/j.physa.2019.122795 https://doi.org/10.1080/19331681.2018.1448735 http://refhub.elsevier.com/S0176-2680(25)00045-X/optPpByvmVBQr https://doi.org/10.1145/2508436.2508437 https://doi.org/10.1145/2508436.2508437 https://doi.org/10.1017/S0003055417000363 https://doi.org/10.1080/10584609.2018.1526238 https://doi.org/10.1080/10584609.2018.1526238 https://doi.org/10.1016/j.electstud.2023.102701 https://doi.org/10.1016/j.jet.2009.11.005 http://arxiv.org/pdf/2012.03327v1 https://doi.org/10.1257/aer.20191777 https://doi.org/10.1257/aer.20191777 https://doi.org/10.1016/j.ejpoleco.2025.102641 https://doi.org/10.1016/j.jebo.2024.106686 http://refhub.elsevier.com/S0176-2680(25)00045-X/sref49 https://buildmedia.readthedocs.org/media/pdf/textblob/dev/textblob.pdf https://doi.org/10.1016/j.ejpoleco.2024.102619 https://doi.org/10.1016/j.ejpoleco.2024.102619 https://ora.ox.ac.uk/objects/uuid:33f7d85c-a25e-4796-aebf-a0e601fc84c1 https://doi.org/10.1080/19331680801975359 https://doi.org/10.1177/2053951718811844 https://doi.org/10.1287/mnsc.2020.3740 https://doi.org/10.1371/journal.pone.0289341 https://doi.org/10.1016/j.jebo.2024.106856 https://doi.org/10.1016/j.respol.2019.103861 https://doi.org/10.1016/j.respol.2019.103861 https://doi.org/10.1080/1369118X.2017.1301521 http://refhub.elsevier.com/S0176-2680(25)00045-X/sref61 https://doi.org/10.1111/joes.12629 https://doi.org/10.1111/joes.12629 https://doi.org/10.4232/1.13790 https://doi.org/10.1111/1475-6765.12373 https://doi.org/10.1016/j.electstud.2010.03.011 https://doi.org/10.1016/j.jpubeco.2015.03.009 https://doi.org/10.1038/s41467-018-06930-7 https://doi.org/10.1177/0894439320906803 https://doi.org/10.1177/0894439320906803 https://doi.org/10.1080/10584609.2017.1334728 Tumasjan, A., Sprenger, T.O., Sandner, P.G., Welpe, I.M., 2011. Election forecasts with twitter. Soc. Sci. Comput. Rev. 29 (4), 402–418. https://doi.org/10.1177/ 0894439310386557. Tyran, J.-R., 2004. Voting when money and morals conflict: an experimental test of expressive voting. J. Publ. Econ. 88 (7–8), 1645–1664. https://doi.org/10.1016/ S0047-2727(03)00016-1. Vepsäläinen, T., Li, H., Suomi, R., 2017. Facebook likes and public opinion: predicting the 2015 Finnish parliamentary elections. Gov. Inf. Q. 34 (3), 524–532. https:// doi.org/10.1016/j.giq.2017.05.004. Zhuravskaya, E., Petrova, M., Enikolopov, R., 2020. Political effects of the internet and social media. Annual Review of Economics 12 (1), 415–438. https://doi.org/ 10.1146/annurev-economics-081919-050239. A.I. Aziz and I. Bischoff European Journal of Political Economy 88 (2025) 102685 18 https://doi.org/10.1177/0894439310386557 https://doi.org/10.1177/0894439310386557 https://doi.org/10.1016/S0047-2727(03)00016-1 https://doi.org/10.1016/S0047-2727(03)00016-1 https://doi.org/10.1016/j.giq.2017.05.004 https://doi.org/10.1016/j.giq.2017.05.004 https://doi.org/10.1146/annurev-economics-081919-050239 https://doi.org/10.1146/annurev-economics-081919-050239 Social media campaigning and voter behavior–evidence for the German federal election 2021 1 Introduction 2 Review of literature 3 Hypotheses 4 Data and methodology 4.1 Data collection and processing 4.1.1 Offline data: opinion polls 4.1.2 Online data: social media activity 4.2 Empirical strategy 5 Results 5.1 Robustness check: the role of sentiment 6 Discussion 7 Conclusion CRediT authorship contribution statement Declaration of competing interest Appendix Declaration of competing interest Data availability References