Partisan selective engagement: evidence from Facebook

This study investigates the effects of variation in “congeniality” of news on Facebook user engagement (likes, shares, and comments). We compile an original data set of Facebook posts by 84 German news outlets on politicians that were investigated for criminal of-fenses from January 2012 to June 2017. We also construct an index of each outlet’s media slant by comparing the language of the outlet with that of the main political parties, which allows us to measure the congeniality of the posts. We ﬁnd that user engagement with congenial posts is higher than with uncongenial ones, especially in terms of likes. The within-outlet, within-topic design allows us to infer that the greater engagement with congenial news is likely driven by psychological and social factors, rather than a desire for accurate or otherwise instrumental information. © 2020 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ )


Introduction
It is well known that concerns about "echo chambers" and "filter bubbles"-citizens limiting their news consumption to belief-confirming or amplifying sources-have grown rapidly in recent years.The corresponding academic literature has grown as well; see Guess et al. (2018) for a recent review.They conclude that "selective exposure to like-minded political news is less prevalent than you think."Weak evidence for partisan selective exposure undermines the common suspicion that new media may be a primary cause of political polarization ( Sunstein, 2017 ;Bail et al., 2018 ).
But exposure to news, and how one responds to news, are of course not the same thing.It is entirely possible that politically congenial news 1 is not only more likely to be believed by partisan media consumers, but that congeniality could also lead to greater engagement with the news.Consumers may be more likely to endorse, pass on, respond to, or simply pay greater attention to more congenial news.These reactions could ultimately influence one's own political beliefs and actions, and those of fellow citizens, with important welfare effects.
In this paper, we study such selective engagement (a term we propose to parallel the already standard term "selective exposure").We examine how Facebook users' likes, shares, and comments on news posts vary depending on the political congeniality of the post.Selective engagement with congenial news would support those arguing that new media can be a significant factor driving polarization (e.g., Lelkes et al., 2017 ;Gentzkow et al., 2018 ;Levy, 2019 ;Schwarz, 2019 ;Shmargad and Klar, 2020 ).In addition, it would likely suggest that social motives dominate information motives among users-an aspect emphasized by recent studies that link social media to hate crime ( Müller and Schwarz, 2018a ;2018 b) and political protests ( Qin et al., 2017 ;Acemoglu et al., 2018 ;Enikolopov et al., 2019 ).
Studying selective engagement is challenging because the nature of most news stories varies in many ways.Both the political congeniality of the story, and other aspects of the story that may affect consumer interest, such as the topic's importance, vary across stories in ways that are typically difficult to observe and measure.We address these challenges by studying variation in engagement with news stories on a fixed topic: the lifting of the immunity of German politicians.Political immunity reduces the risk that members of parliament can be manipulated by the threat of arbitrary prosecution.The requirements to lift a politician's immunity are strictly regulated, and immunity is only lifted when there is strong suspicion of involvement in criminal activity, such as tax evasion, embezzlement, or child pornography.In line with previous evidence,2 we assume that the lifting of immunity for a politician in one's preferred party is bad or uncongenial news, whereas it is good or congenial news when the immunity of an ideologically distant politician is lifted.We verify the validity of this assumption in Section 3.3 .
Besides straightforward coding of congeniality, another key advantage of the topic of immunity being lifted is it that these events occur regularly, and are reported on frequently, by many outlets.During our period of investigation (January 2012 to June 2017), we observe 107 cases of German national and state representatives' immunity being lifted.Our sample of Facebook news pages includes 84 outlets, of which several can be classified as mainstream pages with centrist coverage, as well as some at the left and right ends of the political spectrum.
In Section 2 , we present a highly stylized model of engagement with political news, to guide the interpretation of our results.We show that if social media users are motivated by the desire to provide Friends with information that is useful for either instrumental or intrinsic reasons, then users should in general be more likely to engage with uncongenial news as with congenial news.If users have other motivations-such as expressing how they feel, signaling their party loyalty or the validity of their political views, or wishing to persuade Friends to support one's own preferred party-then users will be more likely to engage with congenial news.
Thus, the model clarifies that selective engagement with congenial news is unlikely to be motivated by the desire to share useful information.While intuitive, this point stands in contrast to the substantial strand of the media bias theory literature showing that selective exposure to like-minded news can be driven by rational information-seeking behavior ( Gentzkow et al., 2015 ).We do not model the implications of engagement for the beliefs of a user's Friends.But it is clear that if Friends are generally like-minded and do not fully account for a user's psychological and strategic motives for engagement, then engaging more with congenial news will tend to push one's Friends toward being more partisan, while engaging with uncongenial news equally or more than congenial news would have a moderating effect.
In Section 3 we discuss the data and method for constructing a measure of ideological similarity between each outlet and party.We use the Facebook Graph Application Programming Interface (API) to download all 2,042,415 posts by these outlets, including information on the posts' time of publication, content, type, and popularity.We then compare the language in these posts with the language used by the political parties in their election programs.Specifically, we compute the cosine distance between the text vectors for each of the 756 outlet-party combinations in our sample.This measure directly translates into a measure of congeniality in the context of our study: The larger the ideological distance, the more congenial the news about the lifting of immunity to an outlet's average reader.In addition, we develop a search routine that is based on a combination of keywords and time parameters to identify posts about the cases of liftings of immunity.
Our estimation results imply that users do engage more with more congenial stories.Users like, share, and comment posts about liftings of immunity more often when the posts relate to politicians from the opposite political camp, compared to posts about ideologically similar representatives.According to the baseline specification point estimate, an increase in congeniality by one standard deviation raises the average number of likes per post by 89%.The increases in shares and comments are 54% and 29%, respectively.Due to the within-outlet, within-topic research design of our study, these differences can most likely be explained by the psychological and social factors mentioned above, and not due to demand for decision-relevant information or simply the desire to hold accurate beliefs (see Garz et al., 2020 ).
An important potentially confounding factor is Facebook's news feed algorithm, which is designed to maximize engagement by providing the content that users prefer.The algorithm could select those political immunity stories that users find congenial (e.g., Claussen et al., 2019 ;Shmargad and Klar, 2020 ), and thus the greater engagement could be exclusively driven by the algorithmic exposure to congenial posts.If so, this would provide indirect evidence of partisan selective engagement.We also investigate this possibility by estimating our model with data from Twitter-when the platform did not have an algorithmically curated timeline-and obtain similar results.Thus it is unlikely that the greater engagement with congenial posts is completely driven by Facebook's news feed algorithm.
The interpretation of our main finding that users engage more with congenial posts for psychological and social reasons is only valid if the ideology of the outlets approximates the ideology of the users that engaged with the political immunity story posts in our sample.If this assumption is violated and the engagement with these posts is instead driven by users that do not share the outlets' ideology, instrumental or intrinsic motives would likely dominate.However, individual-level data on these users indicate that outlet and user ideologies are sufficiently matched in our context.
Furthermore, to complement the direct analysis of engagement, we compile a case-outlet panel and investigate the supply of posts, and find that supply is approximately unbiased.Thus the outlets do not cater to their readers by overreporting congenial cases, or such attempts are offset by differences in the costs of news production: Ideologically close outlets usually have more background knowledge about and better connections to the politician in question, which makes it cheaper to produce news items and therefore easier to post uncongenial stories here.
Our paper primarily contributes to the literature by shedding light on the mechanisms driving user engagement on social media.Bakshy et al. (2015) investigate, among other things, the distribution of average self-reported ideologies of users who share a given article.They find that these averages are highly polarized: Most articles are shared by users who are either consistently liberal or conservative.This alone does not imply the sharing of congenial news, but they note other results suggesting this (that Fox News articles are shared by users who are mostly conservative, and Huffington Post articles shared by mostly liberal users).An et al. (2014) obtain similar results, finding that self-reported partisans are more likely to share news from like-minded sources.Our paper confirms these results, but differs in that we study multiple types of engagement and, perhaps more importantly, that we focus on news on just one topic.Fixing the topic allows us to distinguish between rational-information forces and psychological/social explanations for selective engagement with congenial news.Our paper complements Pogorelskiy and Shum (2018) , who study news sharing and voting on an artificial social network.Their experimental set-up allows them to control for the true informational content of news especially cleanly.We confirm their finding that subjects are more likely to share congenial news, but provide evidence based on observational, real-world data.
In addition, we contribute to the literature on measuring media slant.Similar to the method proposed by Gentzkow and Shapiro (2010) , we compare the language of the most important parties in Germany with the language in the posts of the news outlets to construct an index of slant.The bias of German media outlets has been investigated before; for instance, based on vocabulary and party mentions ( Dallmann et al., 2015 ), the tonality of reports ( Dewenter et al., 2019 ), and slant perceptions of readers ( Polisphere, 2017 ).In contrast to these studies, we do not restrict our investigation to a small number of national, leading media but provide a measure of slant for a more comprehensive set of outlets.For instance, our sample covers all types of news media (i.e., online, print, and broadcasting) and includes the most important regional outlets.More importantly, by calculating the ideological congruence between outlets and individual parties, our investigation is not limited to the position of media in the political left-right spectrum.That is, our slant index accounts for all sorts of ideological differences, such as progressive vs. conservative, egalitarian vs. elitist, authoritarian vs. libertarian, or religious vs. secular.Covering these dimensions is an important aspect when investigating multi-party systems.
Finally, our findings contribute to the literature that investigates the role of media for political accountability, especially to research on news coverage about transgressions of politicians (e.g., Ferraz and Finan, 2008 ;Di Tella and Franceschelli, 2011 ;Puglisi andSnyder, 2011, 2015 ;Nyhan, 2014 ;Garz and Sörensen, 2019 ).In democratic societies, news media are considered crucial watchdogs, allowing voters to make informed decisions.A central question in this literature is how well media transmit politically relevant information.We evaluate this aspect in a social media environment, by studying the supply of news about politicians that are investigated for criminal behavior.

Theory and empirical strategy
We use the basic framework of Gentzkow et al. (2015) .The state of the world is binary, θ ∈ { L, R }, and news is binary as well . The state could be interpreted as the relative valence of a politician ( θ = L would imply the leftist politician has superior valence to the rightist).News about immunity for a politician is evidence about her/his relative valence.
Facebook user i has prior p i = P r( θ = R ) , then exogenously observes the news, n , and then makes a decision about whether to engage with it on Facebook or not ( e i = 1 or 0).We examine several possibilities for the media consumer's objective function to determine basic comparative statics to guide the empirical analysis and interpretation of results.The main goal is to clarify ideas, which is why we keep the model simple and the analysis largely informal.At the end of the section, we summarize the main results, explain how we can use the theory to interpret our empirical findings, and discuss potential differences between likes, shares, and comments.

Instrumental information
First, suppose the consumer is only interested in instrumental information: information improving the quality of another decision.Let X i ∈ { X L , X R } denote the other decision.This decision could be something as simple as which candidate to support in a public opinion poll.A politician's standing in the polls could affect their "political capital" and consequently real policy outcomes.If the consumer or her Friends were perfectly indifferent between actions prior to news, the news would always provide instrumental information.We focus on other cases in which the consumer has one strictly preferred action prior to observing news; without loss of generality suppose this is X R .
Specifically, suppose i chooses e i given n to maximize the expectation of: in which v ( X i , θ ) is i 's payoff from action X i in state θ , −i refers to the set of i 's N Facebook Friends, X i ( e i ) refers to the action of Friend i given e i , and c > 0 is a per-user cost of engagement.That is, i cares about her own action given the state, and the actions of those in her network.The cost of engagement may be from an attention cost to others that i internalizes.
The cost may also be due to i 's concern about bothering Friends, losing credibility with Friends, or Friends paying less attention to i 's engagements.There can also be a fixed cost related to clicking the like/share button or writing a comment, without loss of generality.Assume, for simplicity, that i believes that i 's Friends are only exposed to n due to i 's engagement.
This assumption is most appropriate for shares, but loosely applies to likes and comments, since these actions could attract attention to posts.
If the action matches the state, then v ( X i , θ ) = 1 , and v ( X i , θ ) = 0 otherwise.It is then optimal for i to choose X L given n 3 We say that n has instrumental value for i if n causes i to change her action from her ex ante optimal choice.4Thus, if p i > 0.5 and so X R is ex ante (prior to news) optimal, then n = n R has no instrumental value, and it is straightforward to show that n = n L has instrumental value if π > p i .Suppose all of i 's Friends had the same prior.Then it is clear that v ( X i ( e i ) , θ | n ) is unaffected by whether e i = 0 or , then e i = 1 imposes a cost and no benefit, and so e i = 0 is This condition requires that the expected benefit from providing instrumental information to a Friend (information that causes her optimal action to change) exceeds the cost of engagement, which is very plausible, so we assume that this condition holds.
Suppose i 's network is now heterogeneous.For simplicity, assume: That is, some of i 's Friends have priors favoring L symmetrically to those that favor R .Let f = P r( p i = p i ) , i.e., the fraction of i 's Friends who are like-minded.
In this case, n R has instrumental value for i 's left-leaning Friends, and n L has instrumental value for right-leaning Friends.
Due to the symmetry assumption, this value is the same for any given individual.Let α denote this value for an individual, which is equal to E given that n has instrumental value for i .
Engaging with news creates the same cost for each Friend, and an average benefit of f α if the news has instrumental value for right-leaning Friends, and an average benefit of ( 1 − f ) α if the news has instrumental value for left-leaning Friends.Thus, writing the engagement choice as a function of the news realization, e i ( n L ) = 1 is now optimal if f α > c , i.e. f > c / α, and e i ( n R ) = 1 is optimal if ( 1 − f ) α > c, i.e., ( α − c ) /α > f .Ignoring knife-edge cases, there are now four possibilities: (1) i engages with all news ( ( α It is natural to think that f is greater than 0.5, given the prevalence of homophily (e.g., Bakshy et al., 2015 ;Halberstam and Knight, 2016 ).Making this assumption rules out case 4, since min { c/α, ( α − c ) /α} ≤ 0 .5 .Thus, even allowing for heterogeneous Friends, right-leaning consumers are not more likely to engage with n R rather than n L .
Three additional points are worth noting before proceeding.First, if a user's political preferences or beliefs are sufficiently strong, then uncongenial information would be less likely to have instrumental value for the user, but may still be instrumental for like-minded Friends with weaker preferences.Users might be less likely to share uncongenial information for strategic reasons in this case.Second, if the set of actions were larger, then it would be more likely that sharing congenial information with like-minded Friends would have instrumental value.For example, if the action set included a middle option ("to abstain"), and like-minded Friends prefer this option ex ante, then sharing congenial news would be more likely since this could persuade Friends to take the right-leaning action. 5However, in general with richer action spaces, uncongenial news would likely continue to have just as much, or more, instrumental value, for the same reason that this is true for the binary case (uncongenial information would be more likely to change a user's optimal action).Third, if users take into account that Friends may see other Friends' engagement for the same news, each user could be less likely to engage with any given post, knowing that other Friends may also engage with it.The equilibrium might change to mixed strategies, but the sign of the relationship between congeniality and engagement would be the same.In addition, for users with the same priors, the most natural equilibrium would still be symmetric.

Intrinsic information
Another possibility is that the Facebook user's objective is a function of information for intrinsic and not instrumental reasons.That is, it is possible that users want to hold beliefs that are as accurate as possible, and want their Friends to do so as well.The form of the objective function in Eq. ( 1) could still be applied in this case, with X i now denoting i 's posterior probability that state R is true, and the states L and R correspond to θ taking values of 0 and 1, with v (.) denoting a loss function that increases in the distance between X i and θ .In this case, the benefit of news, which is the impact on accuracy of beliefs, would be greater when the news conflicts with priors.This is because given a prior that favors R , beliefs change more after n L rather than n R .6Hence, again i would be more likely to engage with n L so long as a majority of Friends have like-minded priors favoring R .Thus engagement with uncongenial news ( n L ) seems equally or more likely than engagement with congenial news.Engagement with congenial and uncongenial news would be equally likely for this case if Facebook users were heterogeneous due to differences in tastes rather than differences in priors.

Psychological utility, persuasion, and signaling
Next, we consider several cases in which i 's objective does not depend on the accuracy of i 's Friends' beliefs.The implications for engagement in each of these cases are straightforward, so we omit mathematical analysis and simply discuss the effects that seem most plausible.
First, assume that user i does not gain utility from news informing an action or accuracy of beliefs, but instead i 's news utility increases when the news confirms i 's priors or supports an outcome that i hopes will occur, perhaps creating anticipation utility.Gentzkow et al. (2015) refer to this case as psychological utility.Clearly in this case i would be more likely to engage in the form of liking more congenial news, if i has any preference for the honest expression of this psychological utility, which is plausible ( Brennan and Hamlin, 1998 ;Abeler et al., 2016 ).Similarly, if i internalized the psychological utility of like-minded Friends, i would be more likely to engage with more congenial news to draw Friends' attention to this news.
Another reason for engaging with news unrelated to the desire for others to be well-informed about the state is that i may wish to use the news to persuade others to take a particular view about the state.If i is, for example, a committed rightist, i may wish to maximize the beliefs of others that θ = R independent of the realization of n .In this case, clearly it is optimal to share n = n R only.This case could easily be formalized by making u i a function of Friends' beliefs that θ = R .This case would be even more relevant when i 's Friends are more diverse or even left-leaning.This mechanism is perhaps also most relevant to news stories being shared.
Another mechanism that may drive engagement orthogonal to the desire to provide useful information is signaling.One thing i may wish to signal to like-minded Friends in her network is that i is a strong partisan (to show loyalty, similarity, etc.).Choosing e i ( n R ) = 1 and e i ( n L ) = 0 would be an informative signal of gaining psychological utility from n = n R and sharing this for expressive value.This case could be formalized by making i 's type ("partisanship", e.g., strength of prior favoring R ) uncertain, assuming that more partisan types are more interested in persuasion or the psychological utility of themselves or others, and that less partisan types get utility from a reputation term that increases in Friends' beliefs of that type being more partisan.This case is more likely to be relevant when homophily is more common (when i 's Friends are largely politically like-minded).
An additional signaling possibility is the following.Suppose i may have processed some information inaccurately in arriving at i 's current belief p i and i cares about her reputation with her Friends for processing information accurately.News of n R would be positive evidence in support of prior accuracy of interpretation by i , and n L would be negative evidence.Thus, again i would be more likely to engage with n R , and again this is easily formalized with a reputation term (this time, representing Friends' beliefs about i 's information processing ability, which would increase in consistency of new information with how i processed past information).This case is perhaps more likely to occur when i 's Friends are diverse since non-like-minded Friends are more likely to need convincing that p i > 0.5 is a positive signal about i 's ability.Both signaling cases are more likely for the more noticeable types of engagement, shares, and perhaps also comments, but could also explain "liking" engagement.

Connection between theory and empirical approach
A summary of these results is as follows.Assuming that there is at least some homophily (more of a user's Friends are politically like-minded than not): (1) If engagement is motivated by providing typically like-minded Friends with intrinsically or instrumentally useful information, social media users are in general more likely to engage with uncongenial news.
(2) If engagement is motivated by other factors (expressive or anticipation utility, persuasion, or signaling), then users are more likely to engage with news that is more congenial.
Both ideas refer to the relationship between user engagement and congeniality at the individual level , describing a situation where the message of a post is in line with the views of an individual user.However, our empirical analyses in Section 4 relate to congeniality at the outlet level , evaluating if the news is compatible with the outlet's ideology.That is, in our main estimating Eq. ( 7) , the dependent variable is the user engagement with the political immunity story post published by outlet n on case c , and the key explanatory variable is the ideological distance between outlet n and party p of the politician in question (and hence, the congeniality of that post).
It is crucial to consider different scenarios when interpreting the empirical outlet-level results in light of the individuallevel theory.A first scenario is that the user engages with a post from a like-minded outlet followed by this user (i.e., a "subscribed" outlet that shares the user's ideology).In this case, the ideological distance between outlet n and party p is a good proxy for the ideological distance between the user and the party involved in the political immunity story.Second, the user may also follow the Facebook page of an outlet with a different ideology.In this scenario, the outlet-level congeniality of the post does not capture the congeniality of that post for the user, because outlet and user ideologies are not matched.Third, the user engages with a post shared by a Friend or recommended by Facebook.If this post is from an outlet with a similar ideology as the user, the outlet-level measure of congeniality is again a good proxy for the congeniality of the post from the user's perspective.Fourth, a post shared by a Friend or recommended by Facebook is from an outlet with a different ideology than the user, in which case congeniality at the outlet-level differs from the congeniality for that user.This fourth scenario could be particularly relevant if the user's network of Friends is more heterogenous.
Given that our empirical approach investigates congeniality at the outlet level, any results would be misleading if the second and/or fourth scenarios dominate in practice.Thus, to link the theoretical considerations to the empirical analyses, we need to make the assumption that the outlet's ideology is the same as the ideology of the users engaging with the political immunity story posts published by that outlet.Previous empirical evidence from US newspapers supports this assumption, showing that the average slant of a paper usually matches the ideology of the consumers in its area of circulation (e.g., Gentzkow and Shapiro, 2010 ;Puglisi and Snyder, 2015 ).
How plausible is it that outlet and reader ideologies are matched in our context?It is arguably much easier for online users than newspaper readers to switch between outlets with different ideologies.Thus it is not sufficient to test if the ideology of the average reader matches that of the outlet, but we need to verify that those users who actually engaged with a political immunity story post share the ideology of the outlet publishing it.Similar to Bond and Messing (2015) , we examine these users' like profiles to determine if they are "fans" of the Facebook pages of the political party involved in the post.We describe the procedure at the beginning of Appendix C in detail.We do not use the individual-level data in our main specifications because some users do not have their like profiles public, which may introduce selection bias.The data are nonetheless useful to compare the outlets' average slant with the ideology of the relevant users.According to Fig. C1, users that engaged with a political immunity story post are more likely to be "fans" of the involved party if the outlet is on average slanted towards that party.Thus the figure suggests that the engagement with the posts in question primarily comes from users that share the ideology of the publishing outlet, which allows us to rule out that the second and fourth scenarios described in the previous paragraph play an important role here.With congruent ideologies, we are therefore able to use our main theoretical predications about individual motivations for engagement with (un-)congenial posts to interpret the empirical outlet-level results.
Note that congeniality can be empirically measured in two ways, by looking at (a) bad news for an ideologically distant party and (b) good news for a close party.Higher engagement with either type of news is likely an indication that psychological and social factors matter, since both types are congenial to partisans.However, investigating good news for a close party would be less clean-cut, as motivations for intrinsic or instrumental information might also be at play here, at least in the context of a multi-party system.The reason is that ideological differences between parties are sometimes small (e.g., there have been political overlaps between the Greens ("Grüne") and the Left Party ("Linke"), or between the Christian ("CDU") and Free Democrats ("FDP")), and voters in multi-party systems often have a small set of close parties they consider.Thus engagement with good news for a close party could be an indication that users seek information to decide which party exactly to vote for.Focusing on political immunity stories avoids this kind of confounding, as these stories fall in the category of bad news for an ideologically distant party.

Types of engagement
So far, we have talked about Facebook user engagement mostly in general terms, but likes, shares, and comments can be expected to vary in how strongly they respond to the congeniality of the content, if the engagement is motivated by expressive or anticipation utility, persuasion, or signaling.When a user likes a post, this action is visible below the post itself.There is a relatively low probability that Facebook will select the action of liking into Friends' news feeds, unless they are followers of the publishing outlet.The cost of engagement is low, because liking is done with one click and only some Friends' attention is drawn.
Sharing a post implies that the content will be reposted on the user's timeline.The likelihood that Facebook selects this action into Friends' news feeds is comparatively high.Thus the cost of engagement is higher than in the case of likes, even though sharing also requires just one click.
Commenting is an action that becomes visible below a post and also has a relatively high probability of appearing in Friends' news feeds.This type of engagement certainly has the highest cost of engagement, because the user needs to write something rather than just click a button, in addition to the internalized attention costs of Friends.Comments are used to add information to a post, and since this information could be either critical or confirmatory, this behavior could occur whether the post is congenial or not.
If engagement is motivated by social or psychological factors, we therefore expect liking to be most strongly associated with the congeniality of a post.The relationship between congeniality and sharing is presumably less strong, due to the higher cost of engagement.Commenting involves the highest cost and might not necessarily be used to confirm a post message, which is why we expect the weakest and perhaps most ambiguous relationship here.

Data
Our data cover the time from January 2012 to June 2017.Before 2012, we do not observe much activity on the Facebook pages of the outlets in our sample.In fact, many outlets did not have an official Facebook page before 2012.Our period of investigation ends in June 2017, shortly before we started collecting the data.

Politicians under criminal investigation
In Germany, members of the national and state parliaments are generally protected from judicial and police measures.Unless apprehended while committing an offense, a representative can only be prosecuted or arrested if the parliament grants authorization.A few parliaments adopt lists of cases in which prosecution is possible without an explicit approval of the committee in charge; the authorities merely have to notify the parliament 48 h before taking action.However, in most cases, judicial or police measures have to be authorized.Usually, the act of liftings somebody's immunity consists of two procedural steps.The parliamentary committee in charge first issues a formal request, after which the final decision on the lifting of immunity is made by parliament.Both steps might take place at the same day, but in many cases several days pass between the request and the decision.
Lifting somebody's immunity implies that the prosecution has sufficient evidence to initiate criminal proceedings.Some proceedings relate to minor transgressions (e.g., defamation, driving under the influence), whereas others pertain to severe felonies (e.g., child pornography).The prominence of the politicians varies as well, ranging from ordinary members of state parliaments to former Federal President Christian Wulff.In addition to the legal consequences that are associated with a lifting of immunity, such as indictment and conviction, public approval and intra-party support often decline.In some cases, politicians under investigation are excluded from their party or have to step down.
Most information on cases of liftings of immunity come from official parliamentary records.In a few cases, we complement missing data by publicly available information from other sources (e.g., press archives and search engines); see Garz and Sörensen (2017) for further details on collecting the data.Between January 2012 and June 2017, there are 107 cases pertaining to 80 politicians that belong to 9 different parties, after excluding six cases of politicians without party affiliation.

Facebook news pages and posts
Newspapers, news magazines, and newscasts nowadays provide access to their content online, often complementing their traditional ways of distributing information.Similar to pure Internet news portals, these outlets have Facebook representations which they use to post content.A Facebook post is an individual entry in a page's feed.In addition to its message, a post may include a photo, a video, or a link to external content.News outlets are usually interested in getting Facebook users exposed to and engage with their posts.User engagement refers to liking, sharing, and commenting the post, as well as clicking on external links.In the case of news pages, these links usually redirect the user to full articles on the outlets' websites.
Our selection of outlets is guided by audience reach and the goal to fully cover the political left-right spectrum.To construct a sample that fulfills these criteria, we include all media outlets that focus on general or politics news, have an official Facebook page, and are listed in at least one of the following rankings: (a) the 75 most visited news websites according to the German audit bureau of circulation (Informationsgesellschaft zur Feststellung der Verbreitung von Werbeträgern, IVW ), (b) the 75 most popular outlets on Facebook and Twitter according to the social media monitoring project 10,0 0 0 Flies ( http://www.10 0 0 0flies.de/), or (c) the Süddeutsche ranking of the most-liked Facebook pages by politically interested users ( https://bit.ly/2COYrAV).The IVW is one of Germany's most established provider of audience reach data.Its ranking is based on the number of page impressions of news websites in June 2017.Thus, it does not specifically reflect the popularity of outlets on Facebook but more generally online.The 10,0 0 0 Flies ranking also refers to June 2017, but it is based on the number of likes, reactions, shares and comments on Facebook, as well as the number of likes and retweets on Twitter, generated by websites with their content.The ranking by Süddeutsche is the most specific one to the context of this study.Referring to the time from October 2016 to March 2017, their ranking lists the 100 Facebook pages most often liked by politically interested Facebook users; it consists of a separate list for each of the six most popular German parties these users also like.Based on these criteria, our sample consists of 84 Facebook news pages (many outlets appear in all three rankings).The selection includes Facebook representations of all national newspapers, all national news magazines, the most important regional newspapers, the most important national newscasts, and the largest online news outlets.The sample also comprises known left-and right-wing outlets, such as Compact-Magazin, Junge Freiheit, Junge Welt, PI-News , and taz (Die Tageszeitung).See Table A1 for a full list of outlets.
We use the Facebook Graph API to download the entire content of these pages, including all 2,042,415 posts.The number of posts varies across outlets, with a minimum of 2154 ( Monitor ) and a maximum of 92,955 ( N24 ).Given these quantities, we have a sufficiently large sample to accurately estimate each outlet's ideology, even for those outlets that were less active on Facebook.For each post, we record the date and time of publication, the message text, the type (i.e., link, status, photo, video, or event), the link to the underlying news article, the link to a possible picture, as well as the number of likes, shares, and comments. 7We download the data in the second half of August 2017-at least six weeks after the most recent posts-to guarantee that we register the "final" engagement measures (99.9% of user engagement takes place in the 15 days after a post is published; see Lee et al., 2018 ).

Retrieval of political immunity story posts
We are interested in retrieving those posts that address an imminent, requested, or realized lifting of immunity.It is not difficult to find many of these posts because there are no synonyms for the German word "Immunität".In addition, the term is always used as a noun or compound noun (e.g., "Immunitätsaufhebung") and there are no word corruptions.Thus, the term "Immunität" (truncated at the end) is a very sharp and effective keyword in this context.We focus on the post message, which has a similar function as the headline or sub header of a traditional news article.Journalists include buzzwords in the post message to signal its topic to the reader.The underlying news article very likely addresses the lifting of somebody's immunity if the word immunity is included in this message. 8In addition, there are posts that address the lifting of somebody's immunity without explicitly using the word "Immunität".Such posts are usually published around the date of the request or decision to lift the immunity, discussing the intentions of the prosecution to open criminal proceedings, the particularities of the transgression, and potential or actual consequences of the case.It is also straightforward to retrieve these kinds of posts by searching for the name of the politician in question and comparing the date of the post with the date of the lifting of immunity.To not omit many true positives, we do not initially restrict the search to posts that contain both the first name and the last name, because the media often only uses the last name of the politician.An exception are German last names that are very common (e.g., Beck, Müller) or are also used as other words, such as "Mächtig" (powerful) or "Junge" (boy), which is when the media usually uses the full name to avoid misunderstandings.In these cases, we retrieve only those posts that contain the first name and the last name.
Based on these considerations, we develop a simple routine to search the downloaded page data for posts that are likely about liftings of immunity.The routine identifies all posts that (a) contain the name of the politician in question and the German word for immunity (truncated at the end) in their message text, article link, or picture link, or (b) only include the name but are published within seven days before and after the date of the request or decision to lift this politician's immunity.This search procedure retrieves 1291 posts.Manually removing a few false positives-which are mostly due to homonymy-decreases this number to 1115 posts.Fig. A1 in Online Appendix A shows the distribution of these posts across outlets.Five outlets never posted about a political immunity story ( Lausitzer Rundschau, MAZ, Neue Westfälische, Südwest Presse , and WDR ), whereas the television news channel N24 had a record number of 88 posts.Importantly, about half of the outlets in our sample had more than 10 posts, which implies that we can estimate the relationship between the congeniality of news and user engagement based on a large number of diverse outlets.
Table 1 , Panel B, provides further summary statistics.On average, a post receives about 84.1 likes, 16.9 shares, and 56.4 comments.Most of the posts simply include a link (81.2%), followed by status updates (11.0%).Few posts include a photo (6.9%) and even less a video (0.8%).
We assume that the content of the retrieved posts is bad news for the accused politician.Allegations of criminal behavior are usually associated with losses in vote shares (e.g., Welch and Hibbing, 1997 ;Costas-Pérez et al., 2012 ;Hirano and Snyder Jr., 2012 ).In some cases, such accusations cause politicians to refrain from running for reelection or to resign prematurely ( Garz and Sörensen, 2017 ;Larcinese and Sircar, 2017 ).The language used in the political immunity story posts reflects the detrimental implications of the act.Applying Rauh's (2018) sentiment dictionary for German political language suggests that these posts are disproportionately negative.On average, the post messages contain 1.09 positively but 2.05 negatively connotated terms.Reports about a lifting of immunity always provide "actual news", even if the public has heard rumors about the criminal behavior before.The act of lifting somebody's political immunity implies that the prosecution has collected hard evidence that is likely sufficient for a conviction (about 50% of these cases result in a guilty verdict; see Garz and Sörensen, 2017 ).Thus it is plausible to assume that political immunity stories are negative publicity for the politician in question, and by extension, for the politician's party and supporters.However, it is conceivable that the outlets could attempt to defend the (criminal) behavior of an ideologically close politician, especially when the transgression is politically motivated.As a consequence, readers of these outlets might not perceive such posts as uncongenial.We manually identify all posts that defend the accused politician and verify that these posts do not pose a problem to our approach of measuring congeniality (see robustness checks in Online Appendices B and C).Specifically, we tag all posts that clearly (a) solidarize with the accused, (b) deplore the behavior of prosecuting authorities and parliamentary committees, or (c) contest the meaningfulness of prevailing law. 9We also consider but do not pursue the option to tag these posts by using an automated approach.On the one hand, we are not aware of an automated procedure that could accurately evaluate these criteria.It is possible to measure the sentiment of the posts, as mentioned in the previous paragraph, but it would be a stretch to assume that a post would defend the accused politician if there was a positive sentiment.On the other hand, the number of posts that need to be evaluated is small enough for a manual approach to be feasible.There are 69 posts that meet at least one of the above-mentioned criteria (ca.6.2% of all retrieved posts).In most cases, these posts pertain to left-wing and Green politicians that were accused of violating the right of free assembly, either by participating in anti-Nazi demonstrations or anti-nuclear movements.

Similarity between outlets and parties
We use an approach similar to that proposed by Gentzkow and Shapiro (2010) to construct our measure of media slant.Specifically, we compare the language in the outlets' Facebook posts10 with the language used by the parties in their election programs.Before an election, German parties usually publish a document that provides details on their goals for the upcoming legislative session.These documents are a central element of the campaigns and thus widely discussed in the public.They circulate under the label election program or, in some cases, election platform, party program, party manifesto, or government program.We use all programs pertaining to the two national elections during our period of investigation (2013 and 2017), as well as the available state-level programs in that time.
The election programs are used to identify characteristic terms that are typically used by the parties.With an average of 295,971 words per party, the documents provide a sufficient amount of text.We clean the texts (i.e., lower case transformation, word stemming, as well as removal of punctuation, numbers, stop words, formatting, and party references) and generate a matrix representation of the processed terms.We evaluate the importance of these terms by computing the product of the relative term frequency and the inverse document frequency (TF-IDF), which is a standard statistic in information retrieval for this kind of task (e.g., Jurafsky and Martin, 2008 ).Simply put, the TF-IDF is largest for terms that are often used by one but not the other parties; i.e., terms that appear frequently and uniquely in the parties' election programs.The exact formula is in the notes to Table A2; the table itself shows each party's 20 highest ranked terms to illustrate the approach.For instance, characteristic terms of the Left Party (Linke) are "erwerbslos" (unemployed), "superreich" (super-rich), and "neoliberal" (neoliberal), whereas typical terms by the right-wing party NPD are "Vaterland" (fatherland), "Massenzuwanderung" (mass immigration), and "Ausländerkriminalität" (crime committed by foreigners).The Greens (Grüne) emphasize "Kohleausstieg" (fossil fuel phase-out), "Klimakrise" (climate crisis), and "Atomausstieg" (nuclear phase-out); and characteristic terms of the market-liberal FDP are "Vertragsfreiheit" (freedom of contract), "Schulfreiheitsgesetz" (autonomy education act), and "Träume" (dreams).These examples also show that it is not necessary to compute the TF-IDF for two-or threeword phrases because of the common usage of compound nouns in the German language.
Next, we concatenate the 2,042,415 Facebook posts by the outlets in our sample and clean the message texts in the same way as the election programs.We compare the language of the outlets with the language of the parties by computing the cosine similarity (cp.Jurafsky and Martin, 2008 ) between the cleaned post messages and the most characteristic terms of the parties: Based on terms t , we compute this statistic for each of the 756 outlet-party combinations.The term vector x represents the post messages of outlet n , whereas y captures the language used by party p .Since we are interested in those terms that are highly characteristic for the language of each party, we focus on terms with particularly high TF-IDF values.Specifically, we only use the top 0.1% of each party's characteristic terms, according to the TF-IDF statistic.Selecting this cut-off is arbitrary and subject to a trade-off.11Using only few party-specific terms increases the chances that we treat posts as neutral even if they are actually biased, as we might exclude terms that are indicative of a party's ideology.An overly generous cut-off decreases the differences between the parties, because it involves including terms that are not often used, that are simultaneously used by the other parties, or both.Inspection of the retained terms suggests that the top 0.1% of the distribution of TF-IDF values balances the trade-off well.This cut-off selects between 48 (AfD, NPD) and 86 (FDP, SPD) terms per party, or a total of 582 expressions.These differences between parties reflect the uniqueness and length of the election programs.Thus parties with a broader range of unique topics have a greater weight when constructing the slant index.
We plot levels of slant by outlet and party in Figs.A2 and A3.The graphs indicate that the outlets generally vary in their tendency to use ideologically relevant terms.For instance, we observe the lowest tendency to use terms for the regional outlet Badische Zeitung (cosine similarity cumulated over parties = 0.0034).The socio-critical blog Nachdenkseiten , for which we measure the highest cumulated cosine similarity (0.0399) exceeds this value by a factor of 10.However, there are also differences in the extent of bias towards individual parties.For example, we observe outlets devoting between 4.1% ( Osthessen News ) and 23.4% ( Zeit ) of their overall bias to the green party (Grüne), and between 0% (multiple outlets) and 10.9% ( Compact-Magazin ) to the far-right NPD.
In the regressions, our measure of congeniality is the cosine distance between an outlet and a party (i.e., dis t n,p = 1 − si m n,p ). 12 To ease the interpretation of the results, we rescale the cosine similarity using its empirical minimum and maximum before computing the distance: The resulting measure varies between 0 and 1, with 1 reflecting the greatest observed ideological distance between an outlet and a party, and 0 indicating the greatest similarity.The empirical distribution of this distance measure (cp.Fig. A4) implies that there is much variation in the congeniality of political immunity story posts, which is useful when estimating the effects on user engagement.

Score in the political left-right spectrum
To illustrate the outcome of the computations, we rank the outlets in the political left-right spectrum.For that purpose, we use left-right scores of Germany's political parties provided by the Manifesto Project ( Volkens et al., 2019 ).These scores are based on content analyses of the parties' 2013 and 2017 election programs, following a standardized procedure that allows comparisons over time and across countries.As shown in Fig. 1 , the Pirate Party (Piraten), the Greens (Grüne), the Social Democratic Party (SPD), and the Left Party (Linke) can be found left of the mean score, whereas the Union parties (CDU/CSU), the Free Democratic Party (FDP), and the newly founded Alternative for Germany (AfD) are right of the center. 13ext, we regress the cosine similarity between outlet n and party p on a constant and outlet and party fixed effects μ n and θ p : The posts of some outlets are generally more similar to the parties' election programs than the posts of other outlets; and the election programs of some parties are generally closer to the outlets' posts than those of other parties.Using the residuals ɛ n, p from Eq. ( 4) accounts for these differences.Our measure of an outlet's position in the left-right spectrum is the sum of the outlet's similarity residuals ɛ n, p weighted by the parties' score p on the left-right scale (as shown in Fig. 1 ): scor e n = P p=1 ε n,p scor e p (5) By construction, the outlets' left-right score n is bounded between -1 and 1. Negative values suggest that an outlet's post messages are slanted towards the parties left of the population mean, whereas positive values indicate slant towards the right side of the spectrum.
In Fig. 2 , we compare the resulting left-right score of the national outlets with the popularity of their Facebook pages.The distribution of these outlets resembles a bell curve.The graph suggests that the most popular outlets-such as Bild, Spiegel Online , and Tagesschau -are fairly balanced; i.e., have left-right scores close to zero.Outlets at the left and right ends of the spectrum are less popular on Facebook.We do not find such a pattern when looking at the regional outlets in Fig. 3.14 This is plausible because regional outlets usually cater to the views of consumers in local news markets ( Gentzkow and Shapiro, 2010 ).These markets vary in size, which is a main factor of the regional outlets' popularity on Facebook.In addition, the newspaper versions of the regional outlets are often local monopolists; as such they have incentives to cater to a broad ideological spectrum rather than a single ideology ( Gentzkow et al., 2015 ).The regional outlets also have a smaller range of left-right scores, which implies that their Facebook pages are ideologically more balanced than the national ones, as can be expected in locally concentrated news markets.Note, however, that we compute and display the left-right scores only for illustrative purposes, mainly to validate that our approach of measuring ideology yields plausible results.Our main explanatory variable-the ideological distance shown in Eq. (3) -has very similar means and standard deviations when comparing national and regional media.Thus we use variation in ideology from both types of outlets when estimating the relationship between congeniality and user engagement.
It might seem puzzling that Facebook pages belonging to the same brand (i.e., Spiegel and Spiegel Online , as well as Zeit and Zeit Online ) have slightly different left-right scores.Using different pages operated by different social media editors is most likely part of product differentiation strategies that allow media companies to target different audiences (e.g., Anand et al., 2007 ;Gal-Or et al., 2012 ).Thus it is not implausible for these outlets to have different left-right scores.
Overall, the positions of the outlets in the left-right spectrum shown in Figs. 2 and 3 concur with their reputations.For example, we obtain large positive scores for the right-wing outlets Junge Freiheit, PI-News , and Compact , whereas the socialist newspaper Neues Deutschland , the socio-critical blog Nachdenkseiten , and the cooperative-owned taz (Die Tageszeitung) exhibit large negative scores.

Supply of posts
To begin, we investigate the supply of posts, given the availability of news material.For that purpose, we estimate the effect of the congeniality of the case on the number of posts on a given case, by a given outlet: where dist is the cosine distance between outlet n and party p of the politician involved in case c , as calculated in Eq. ( 3) .Note that the congeniality of the case derives from the ideological congruence between the outlet and the party affiliation of the politician in question: Cases in which an outlet is slanted towards the party of the politician are uncongenial, because the lifting of immunity is bad news for the average reader of this outlet.In contrast, cases are congenial if an outlet is ideologically different from the party of the politician.Thus, high values of dist reflect congenial constellations, whereas small values indicate uncongenial cases.We estimate this effect conditional on outlet and case fixed effects.The outlet fixed effect μ n captures unobserved differences across outlets; for example, due to popularity, social media strategy, or affinity for the topic.The case fixed effect ϕ c accounts for the particularities of the case, such as the point of time of the lifting of immunity, the severity of the transgression, the party of the suspect, or the reputation of the politician, which might all result in differences in news value.We compute two-way clustered standard errors by outlet and case.
We use different versions of the dependent variable when estimating Eq. ( 6) .The basic version simply counts all posts per outlet and case.In the second version, we exclude posts that refer to multiple issues (i.e., posts also addressing other topics than the lifting of somebody's immunity).For instance, the Facebook pages of newscasts sometimes use a post to list the main topics of their upcoming show.The third version excludes posts related to multiple politicians from different parties.The outlets sometimes report on multiple politicians simultaneously, because the cases occur at the same time or to compare a case with a previous lifting of immunity.
It would be optimal to use estimation procedures that account for the distributional characteristics of these count variables, such as Poisson or negative binomial models.We use OLS to estimate Eq. ( 6) though, because maximum likelihood estimators are biased when modeling panel data with (two-way) fixed effects; see Greene (2004) on the incidental parameters problem.In fact, maximum likelihood estimates often fail to converge with the data at hand, because there are many outlet-case combinations without any variation in the amount of posts.
Results are summarized in Table 2 .All specifications indicate a negative relationship between the ideological distance and the supply of posts.Thus the number of posts slightly decreases when the congeniality of the case to the outlet's average reader increases.However, the estimated coefficients are statistically insignificant and imply very small effect sizes.For example, the coefficient of −0.412 in Column ( 1 ) indicates that a one standard deviation increase in distance (0.193) decreases the number of posts by 0.079.This decrease corresponds to approximately 7.1% of the standard deviation of the amount of posts.
Fig. B1 supports the interpretation that the congeniality of the case does not affect the number of posts related to that case.To rule out that this finding is caused by our specific approach of capturing congeniality, we evaluate the robustness of the estimates when we construct the distance measure in a different way.We obtain similar results when using alternative TF-IDF cut-offs to construct this measure (Tables B1 and B2) and when we use a measure of congeniality based on the Polisphere (2017) index of slant (Table B3).As discussed in Section 3.3 , a small fraction of posts defends the accused.These posts might have a different congeniality, but our estimates do not substantially differ when we distinguish between these posts and "ordinary" political immunity stories (Table B4).To rule out that our results are driven by outliers, we exclude the most active outlets (Table B5) and drop the most prominent cases (Table B6).Excluding outlets and cases with zero posts does not affect the results either (Table B7).Another concern is that OLS estimates are biased when the dependent variables are count variables.However, we obtain qualitatively similar results when we estimate linear probability models with binary dependent variables that indicate whether the outlet posted about a case or not (Table B8).Another possibility is that only certain outlets adjust their supply of posts to the congeniality of the case, but we do not find any differences between leftand right-leaning or between centered and more extreme outlets (Table B9).It is also conceivable that outlets do not adjust the number of posts but change certain qualitative elements.For instance, outlets could be more inclined to include a call to action in a post when the case is more congenial.This is not the case though (Table B10).In addition, there is no evidence that the congeniality of the case affects the characteristics of the posts, such as the length of the post message, the usage of photos and videos, and the share of posts published on Sundays or at night (Table B11).
In conclusion, there is no evidence that outlets cater to the preferences of users by posting more about more congenial cases.Absence of evidence is not necessarily evidence of absence though.It is possible that attempts by outlets to satisfy consumer demand for congenial posts are offset by other factors.For instance, cost advantages could make it attractive to post about less congenial cases: Outlets that are ideologically close to a party likely have more background knowledge about and better connections to politicians from that party, which implies lower costs of producing news items about them.In contrast, it might be more difficult for outlets to obtain certain information when politicians from ideologically distant parties are involved.Such cost differences influence the supply of news items, which in turn affects the activity of outlets on Facebook.

Estimation and results
We estimate versions of the following model to explore the effect of the congeniality of posts on users: in which we use each of our three measures of user engagement (likes, shares, and comments) related to post i , published by outlet n on case c as left-hand side variables. 15Again, dist captures the ideological distance between the outlet and party p of the politician in question (and hence, the congeniality of a story on lifting immunity).The outlet and case fixed effects, Notes: N = 1115.OLS estimates.The column headers denote the dependent variables.All models include outlet and case fixed effects.The control variables include the type of the post, the length of the post message, the outlets' monthly average number of likes over all published posts, day of the week and hour of the day fixed effects, the overall and the outlet-specific number of previous posts on the same case, the number of days since the first post on the same case, a dummy to capture posts that refer to multiple politicians of different parties, and a dummy to capture posts about multiple topics.Standard errors (in parentheses) are clustered by outlet and case.* p < 0.10, * * p < 0.05, * * * p < 0.01.
μ n and ϕ c , account for unobserved outlet-and case-specific characteristics.The variable vector X includes controls for the type of the post (i.e., link, photo, video, status, and event), the length of the post message (number of characters), and a dummy variable to capture posts that refer to multiple politicians of different parties.Another binary variable captures posts that cover multiple topics because user reactions to such posts do not necessarily refer to the lifting of somebody's immunity.In addition, dummy variable sets account for the hour of the day and the day of the week, because the timing of publication likely affects user engagement as well.We also include the number of days since the first post on a case, as well as the overall and the outlet-specific number of previous posts on the same case to capture variation due to potential effects of the news cycle.Finally, we include the outlets' average monthly number of likes over all published posts to account for differences in popularity over time and across outlets.Again, we cluster standard errors by outlet and case.Here the effective numbers of outlets and cases are 79 and 49, respectively: Some outlets never posted about a political immunity story, and various cases did not receive any posts.Technically, it would be possible to include month and year fixed effects, as well as a time trend polynomial to account for general, time-related patterns.However, such patterns are almost entirely absorbed by the case fixed effects because most posts are published at the time of the lifting of immunity.Including these variables does not affect the results but leads to extremely large variance inflation factors due to multicollinearity, which is why we prefer to omit the additional controls.
Table 3 shows the results of estimating Eq. ( 7) .In general, the estimates do not substantially differ for the models with (Columns 4 to 6) and without (Columns 1 to 3) control variables.Throughout, the coefficient of the distance variable has a positive sign.Thus the estimates suggest that users engage more with posts, the greater the level of congeniality of the posts.Assuming that outlet and user ideologies are approximately matched (cp.Fig. C1), the positive relationship implies that the users' engagement is likely motivated by psychological and social factors, as discussed in Section 2 .According to Columns ( 4 ) to ( 6 ), a one standard deviation increase in distance (0.193) raises user engagement by approximately 74.4 likes, 9.2 shares, and 16.4 comments.In relative terms, likes, shares, and comments increase by 88.5%, 54.3%, and 29.1%, respectively.The coefficients are significant at the 5% and 10% level in the case of likes and shares, and insignificant for comments.Both the size of the coefficients and their significance levels are compatible with the theoretical expectations formulated in Section 2.5 .That is, we observe the strongest association for likes, which involve the lowest costs of engagement.The relationship is slightly weaker for shares, presumably because of the somewhat higher engagement cost.The low correlation between comments and engagement can likely be explained by the relatively high engagement cost and the fact that comments are not necessarily used to endorse the content.
Graphical evidence confirms the positive relationship between the congeniality of the posts and user engagement (see Fig. C2).A series of robustness checks suggests that our results are not substantially affected by specific choices made when constructing our measure of congeniality.Specifically, we obtain qualitatively similar results when we use alternative TF-IDF cut-offs to select ideologically relevant expressions (Tables C1 and C2), construct a measure of congeniality based on differences in the left-right spectrum (Table C3), or create a distance measure based on the Polisphere 2017 ranking of outlets (Table C4).We evaluate if posts that defend the accused affect the results, which is not the case (Table C5).We also verify that our results are not driven by outliers in the engagement variables (Fig. C3), particularly prominent cases (Table C6), or two outlets with extreme left-right scores (Table C7).The user engagement could be driven by posts including a call to action (e.g., "What do you think about the accusations?"),but removing these posts does not change the results either (Table C8).
Overall, the robustness checks confirm the results of the baseline specification, but the level of the robustness varies over the different engagement measures.We find the most robust effects for likes.The evidence is slightly less robust for shares, and most specifications do not indicate significant effects for comments.Again, this pattern matches the theoretical predictions discussed in Section 2.5 .
As discussed in Section 2 , the findings can be best explained by psychological and social factors.It "feels good" to encounter congenial news, whereas it "feels bad" to receive uncongenial information.The instrumental information concept used in many theoretical models of demand-driven media bias is not compatible with the result that consumers prefer congenial information within the same outlet.If consumers were behaving rationally and were seeking information with instrumental value, we should find either (a) no within-outlet relationship between congeniality and user engagement or (b) users to be more engaging with uncongenial posts.This is not the case though.The absence of such a finding does not necessarily imply that the effect it not present, but it could mean that psychological and social factors are dominant.

Alternative explanations
The positive relationship between the congeniality of posts and user engagement could be subject to reverse causality if ideologically motivated social media editors exploit certain features of Facebook to advocate their own or the outlet's political agenda.There are several ways in which outlets can promote individual posts to increase the chances that these posts appear in users' news feeds, which in turn increases the chances of engagement.Outlets can pay Facebook to "boost" posts, they can pay click farms to influence organic exposure, they can use social bots, they can pay influencers to spread the word (e.g., celebrities), and they can slant message texts.Editors could systematically promote those posts that support the outlet's ideological goals and disregard the posts that contradict these goals.Specifically, outlets could use these tools to sway audiences with a different ideology.We cannot conclusively rule out this kind of behavior, but various pieces of evidence contradict the hypothesis that outlets attempt to persuade users on a large and successful scale.First, our finding that the supply of posts is approximately unbiased does not support this kind of behavior.Second, as discussed in Appendix B, we do not find that the outlets manipulate the salience of posts in congenial cases by changing post characteristics (i.e., length, duration of coverage, and use of photos/videos) and the timing of publication (i.e., weekdays vs. weekends, night vs. day).If the actions of the editors were driven by ideological goals, we should already observe a bias at this stage.Third, persuasion would imply that outlets target partisan or uncommitted users with content that is not congenial to them, in which case we would not observe a positive relationship between congeniality and user engagement in the data.While not impossible, it is therefore unlikely that our results are driven by an ideologically motivated promotion of posts.
It is also conceivable that other actors could attempt to persuade audiences by using bots, fake accounts, or trolling.This kind of influence is mostly known in the context of elections and politically motivated disinformation campaigns.We cannot rule out that political immunity stories are used to manipulate Facebook audiences.However, the posts in our sample often relate to low-level politicians and are published outside of election times, so it is unclear if there is enough at stake for disinformation campaigns to influence our results at a large scale.The robustness checks that exclude the most prominent cases offer some reassurance here (cp.Table C6).In addition, bots and trolls tend to share content from dubious sources, but the vast majority of posts in our sample were published by legacy media.Our heterogeneity analyses indicate that the relationship between congeniality and user engagement is mostly driven by centered outlets (cp.Table C9, Columns 4 to 6), which also suggests that political bots are unlikely to be a big factor here.
Another alternative explanation relates to Facebook's news feed algorithm, which predominantly selects congenial content into users' news feeds.The algorithm likely predicts the affinity of users to certain outlets, parties, and politicians, as well as preferences for news categories (e.g., politics vs. sports) and topics within news categories (e.g., posts about transgressions of politicians vs. posts about policy making).It might also be possible that the algorithm distinguishes between congenial and uncongenial news within the same outlet in the specific context of liftings of immunity.However, as Bakshy et al. (2015) show, the algorithmic selection has a smaller effect on exposure to congenial content than individuals' choices.To evaluate the role of the algorithm, we collect data on engagement with political immunity stories on Twitter (see Appendix C for details).In contrast to Facebook, Twitter did not use an algorithm to expose users to tweets until January 2016.The platform only began to test an algorithmically curated timeline in February 2016.Before that, tweets were simply shown in reverse chronological order to users. 16As Table C10 shows, Twitter users did engage more with more congenial political immunity stories prior to February 2016, despite the lack of algorithmic exposure.However, we also find that the liking of congenial tweets has increased afterwards, whereas we do not observe significant changes for retweets and replies.Thus the algorithmic content selection might amplify user engagement with congenial content, but it is unlikely that our Facebook results are exclusively driven by the news feed algorithm.Unfortunately, we cannot quantify how much of the effect estimated in Eq. ( 7) is driven by the news feed, since the Facebook and Twitter algorithms differ in certain details.
A related possibility is that social media editors exploit tools (e.g., paid reach, influencers) to raise the chances that users are exposed to congenial content.Outlets could steer different posts to different audiences, depending on the content.For instance, an outlet might explicitly target right-leaning users with political immunity stories about left-wing politicians, whereas left-leaning users could be targeted with posts about right-wing politicians.This kind of strategy would aim to maximize engagement by catering to user preferences.As with Facebook's news feed algorithm, editor-driven exposure could amplify the positive relationship between congeniality and user engagement.Similarly, outlets could take direct measures to increase engagement metrics (e.g., social bots, engagement baiting).Checking the political immunity story posts in our sample, we find that a small fraction of posts (4.8%) include calls to action (e.g., "What do you think about this?").Excluding these posts does not change our results though (Table C8).We also check if there are other forms of engagement baiting (e.g., "Tag a friend that…", "Like this if you think that…", "Share with 10 friends if you…") but do not find any in our sample of political immunity story posts.

Conclusion
We investigate the effect of variation in congeniality in Facebook posts on user engagement.Using data on posts about the lifting of politicians' immunity allows us to distinguish between messages that confirm and contradict the beliefs of the average readers of 84 German news outlets on Facebook.Our results indicate that the supply of posts is approximately unbiased, whereas we find systematic differences in user engagement.Posts with congenial messages receive substantially more likes than uncongenial messages.To lesser degree, this also applies to shares, but not necessarily to comments.These differences across the engagement metrics match our theoretical predictions.
It would be optimal to explicitly account for potential supply-side factors that affect users' exposure to posts, such as ideologically motivated social media editors or Facebook's news feed algorithm.However, it is unlikely for such factors to be large confounders in our within-outlet, within-topic approach, because exposure only translates into engagement if user preferences are compatible.Another limitation of our research design is the specific kind of news that we investigate.It is unclear whether our findings can be generalized to other contexts than transgressions of politicians.Regardless, the results have important theoretical implications.We show that psychological and social factors play a major role in shaping user engagement.However, many models of demand-driven media bias do not account for such factors (e.g., Burke, 2008 ;Chan and Suen, 2008 ;Sobbrio, 2014 ;Oliveros and Várdy, 2015 ;Fang, 2016 ).This is problematic because these models lead to different conclusions about the welfare implications of media bias than models that do account for psychological factors (e.g., Mullainathan and Shleifer, 2005 ;Gentzkow and Shapiro, 2006 ;Bernhardt et al., 2008 ;Stone, 2011 ): If preferences for congenial news are driven by the desire to hold accurate or otherwise instrumental information, the proliferation of ideologically diverse media is relatively likely to cause consumers to be better informed.If preferences for congenial news are based on psychological and social factors, as our paper suggests, concerns about filter bubbles leading to biased beliefs are more justified.

Fig. 1 .
Fig. 1.Left-right scores of political parties.Notes: The figure shows left-right scores of Germany's main political parties, based on data from the Manifesto Project ( Volkens et al., 2019 ).The Manifesto Project derived theses scores by analyzing the content of the parties' 2013 and 2017 election programs.More negative values indicate stronger leanings to the left, whereas more positive values imply stronger leanings to the right.For comparison, the Democratic and Republican parties in the US scored −20.58 and 32.97, respectively, on the same Manifesto Project scale (pertaining to the 2016 US presidential elections).For a recent journalistic comparison of German and US political parties see Chase (2017) .

Fig. 2 .
Fig. 2. Popularity and slant of Facebook news pages, national outlets.Notes: The page likes refer to August 2017.The score on the x axis is computed by mapping language similarities between outlets and parties to left-right scores of the parties provided by the Manifesto Project (see Section 3.4 for details).

Fig. 3 .
Fig.3.Popularity and slant of Facebook news pages, regional outlets.Notes: The page likes refer to August 2017.The score on the x axis is computed by mapping language similarities between outlets and parties to left-right scores of the parties provided by the Manifesto Project (see Section 3.4 for details).

Table 1
Summary of the data.

Table 2
Supply of political immunity story posts and congeniality of cases.= 8988 (107 cases, 84 outlets).OLS estimates.The column headers state the dependent variables.All models include outlet and case fixed effects.Standard errors (in parentheses) are clustered by outlet and case.p < 0.10, * * p < 0.05, * * * p < 0.01. *

Table 3
User engagement and congeniality of posts.