A New Look at Celebrity Endorsements in Politics: Investigating the Impact of Scandalous Celebrity Endorsers and Politicians’ Best Responses

ABSTRACT When a celebrity receives negative news coverage, his or her endorsements of politicians can pose negative consequences for the politicians. We investigated such negative consequences with the help of two experimental studies. In Study 1 (celebrity involved in tax scandal), we manipulated whether an endorsement was initiated by a politician or a celebrity (i.e., controllability) in a 2 × 2 between-subject experiment. We also manipulated politicians’ responses (i.e., no response vs. response). Study 2 was a conceptual replication of the first experiment (celebrity involved in a real estate scandal). Results of Study 1 revealed that politicians are perceived to be more in control of self-initiated endorsements than other-initiated ones. Perceived controllability, in turn, influenced feelings of anger and pity, eventually affecting voting intentions. For self-initiated endorsements, no response appears to be the best reaction. By contrast, public response is advised when the endorsement was initiated by another entity. Results were replicated in Study 2. However, particular responses of a political candidate revealed no influences in connection with a real estate scandal. We explain our findings by applying the theory of planned behavior, attribution theory, and situational crisis communication theory.

a collaboration with a celebrity endorser (e.g., via Facebook). Celebrities frequently announce their support without the approval or even the knowledge of the endorsed politician. Often, celebrities simply post their support on Twitter, Facebook, or YouTube (Loader, Vromen, & Xenos, 2016). Given the celebrity's elevated social status, such news gets picked up frequently and is circulated by traditional media-for instance, Forbes or the British Broadcasting Corporation, as in the previous examples (Clinton, Cameron). As a result, the news often reaches a far broader audience. However, it remains unclear whether such other-initiated endorsements precipitate the same effects as endorsements initiated by politicians themselves. Of course, when it comes to potentially harmful endorsements, whether the same effects can be expected is vital information for political campaigns.
Although some studies on celebrity endorsements in politics have examined the moderating conditions of endorsement effects, hardly any study has investigated the psychological mechanisms underlying those effects, the exception being Mishra and Mishra (2014). There is relevant research available on mechanisms with regard to celebrity endorsements in marketing communications (e.g., Kang & Herr, 2006). However, it is debatable whether such results can simply be transferred to the political sphere ( van Steenburg, 2015). Addressing those shortcomings, in this study we provide a new look at celebrity endorsements in politics by examining the influence of potentially harmful endorsements, either self-or other-initiated, as well as the psychological mechanism underlying their influence and politicians' best options for responding to them. Study 1 examined potential harmful consequences in connection with a comedian's tax scandal. Study 2 was a conceptual replication of the first experiment in which a celebrity was involved in a real estate scandal.

Positive and negative consequences of celebrity endorsements
The influence of celebrity endorsements is typically modeled as associative learning (Erdogan, 1999). In that scheme, individuals learn the association between a celebrity endorser and an object via (repeated) exposure. As a result, responses initially directed toward the celebrity are liable to become associated with the endorsed object, as well (Knoll, Matthes, Münch, & Ostermann, 2016). For instance, positive feelings toward celebrities enhance attitudes toward candidates when celebrities are paired with them (Chou, 2014;Pease & Brewer, 2008). Following the source effects model (McGuire, 1985), such positive feelings are conceived to originate from celebrities' attractiveness, credibility, and expertise. When people's involvement with a candidate or political issue is low, such effects should come to existence regardless of whether the celebrity and the candidate match (Kang & Herr, 2006). That is, regardless of whether the celebrity is similar or consistent with relevant attributes of the endorsed object (Bergkvist & Zhou, 2016). Given that people's involvement is high, only matching celebrities should provoke enhanced effects (Kang & Herr, 2006). According to the elaboration likelihood model, people elaborate more strongly on the endorsement when their involvement is high (Petty & Cacioppo, 1986). As a result, they consider the endorsement information only if it presents a relevant argument in evaluating the endorsed object (e.g., candidate). Only matching information is likely to be considered (Kahle & Homer, 1985;Kirmani & Shiv, 1998;Petty & Cacioppo, 1986). Besides involvement and match, political communication researchers investigated the moderating effects of gender, party preferences, celebrity liking, and general attitudes toward political ads. The gender of the celebrity did not moderate political endorsement effects (Morin et al., 2012). However, participants' gender did matter. There were weaker effects for female participants (Chou, 2014). In addition, effects were stronger for endorsements of parties consistent with one's own party preferences and endorsements by liked celebrities (Chou, 2015;Nownes, 2012). Also, effects were stronger for participants with poorer general attitudes toward political ads (Chou, 2014).
Just as positive responses toward celebrities can become associated with endorsed objects, negative responses can become associated, too. Research in brand and product marketing has focused on that aspect for quite some time (e.g., Carrillat et al., 2014;Till & Shimp, 1998;Um, 2013). Among such research, studies have consistently revealed the negative influences of celebrity transgressions on brands that they have endorsed: "An inherent risk in all celebrity endorsements is misbehavior by the celebrity that spills over to brands endorsed" (Bergkvist & Zhou, 2016, p. 651). Effects primarily surfaced if the celebrity or brand was responsible for the transgression. By contrast, negative effects did not emerge if responsibility could be attributed to external causes (Um, 2013). Such findings fit neatly with the aforementioned discovery that political celebrity endorsements have become increasingly initiated by other parties (Loader et al., 2016). When it comes to negative consequences, they can prompt differential effects compared with endorsements initiated by politicians.
Interestingly, research in political marketing has entirely neglected to examine the negative influences of celebrity endorsements, and no knowledge is available about whether effects established for consumer behavior hold true for voting behavior, too. In response, van Steenburg (2015) recently called for more empirical evidence of the existence or nonexistence of the socalled "voter-as-consumer" paradigm.

Theoretical framework
The theory of planned behavior (Ajzen, 1991) and attribution theory (Weiner, 1985) served as our theoretical framework. Whereas the former can explain why people intend to vote for a candidate due to their attitudes (e.g., Echabe & Garate, 1994;Hennessy, Delli Carpini, Blank, Winneg, & Hall Jamieson, 2015;Singh, Leong, Tan, & Wong, 1995), the latter can explain the formation of emotional reactions to potentially harmful endorsements that influence those attitudes (Glaser & Salovex, 1998;Weiner, 1985). We also applied situational crisis communication theory, which offers insights into politicians' best responses to potentially harmful endorsements (Coombs, 2007).

Theory of planned behavior
We applied the theory of planned behavior to explain why people intend to vote for a specific candidate, that is, to explain on what their vote is primarily based on. The theory has been successfully applied for this purpose in numerous studies (Rosema, 2004). It assumes that intention to vote for a candidate can be influenced by (a) attitudes toward voting for the candidate or attitudes toward the candidate, (b) subjective norms and social pressure related to voting for the candidate, and (c) perceived behavioral control over voting for the candidate (Ajzen, 1991). When it comes to voting, the behavior is usually performed anonymously, meaning that people cannot be held accountable for their behavior and that social pressure is generally low. As such, social norms are hardly a factor in predicting voting intention (Rosema, 2004). Perceived behavioral control is even less of one. Specifically, control has been found to contribute little to the prediction of intentions when the respective behavior poses no difficulties for its performance (Ajzen & Fishbein, 2005). Accordingly, perceived behavioral control has been found to be largely irrelevant when it comes to voting choices (Ajzen, 1991). By contrast, attitudes toward candidates have turned out to be a powerful predictor of voting intention (Wojciech, Falkowski, & Kaid, 2000). In Rosema's (2004) words, How people vote appears to be determined strongly by the degree to which they like or dislike the competing parties or candidates. Usually, voters simply vote for the party or candidate they like best. Such attitudes should thus be central in models of voting (p. 59; Figure 1) In support, various meta-analyses have consistently detected a particularly strong relationship between attitudes and behavioral intentions in regard to people's voting behavior (Kim & Hunter, 1993).
In tracing the origins of attitudes, social psychologists have generally assumed that attitudes result from emotions in response to encounters with attitude objects (Olson & Kendrick, 2008). More specifically, an emotional response can become conditioned to the attitude object, thereby resulting in the psychological tendency to evaluate the object with some degree of favor or disfavor (Eagly & Chaiken, 1993). This conceptualization fits neatly with the findings of various studies showing that emotions elicited by political candidates are a powerful predictor of attitudes toward them (e.g., Abelson, Kinder, Peters, & Fiske, 1982;Chang, 2001;Marcus, 1988;Ottati, Steenbergen, & Riggle, 1992;Ragsdale, 1991), as reviews by Glaser and Salovex (1998) and Marcus (2000) have demonstrated. In fact, "Affective reactions play a major, if not the dominant role in candidate selection" (Glaser & Salovex, 1998, p. 159). Following such thinking, in this article we explain voting intention according to attitudes toward candidates and attitudes themselves according to emotions elicited by candidates. This strategy upholds the theory of planned behavior, which views emotions as an important antecedent of attitudes (Ajzen & Fishbein, 2005).

Attribution theory
Attribution theory encompasses several theories, as Folkes's review (Folkes, 1988) has shown. They are concerned with modeling how individuals make sense of events in their environments by attributing causes to the events. Prominent examples are Kelley's covariation theory (1967) or Weiner's (1985) attributional theory of achievement motivation and emotion. We concentrate on the latter because Weiner (1985) is explicitly concerned with the explanation of emotions as a result of attributional processes (Folkes, 1988) and we are interested in the emergence of emotions as important antecedents of attitudes (Ajzen & Fishbein, 2005). According to that theory, individuals are interested in gaining knowledge about causes of events in their environment to more effectively manage their well-being (Lazarus, 1991). In short, if an event is perceived to be successful, then people are likely to enact behavior that reinstates the event's cause. By contrast, if an event is undesired, then people are likely to demonstrate behavior that alters the event's cause. In that sense, specific emotional reactions, which form as a result of those causal attributions, trigger people's behavior (Weiner, 1985). Therefore, depending on people's specific causal attributions, the theory can help to predict people's emotional reactions.
Although countless definitions are available, an emotion can generally be understood as "an episode of interrelated, synchronized changes in the states of all or most of the five organismic subsystems in response to the evaluation of an external or internal stimulus event" (Scherer, 2005, p. 697). The five organismic subsystems encompass cognitive, neurophysiological, motivational, motor expressive, and subjective feeling components (Scherer, 2005). Regarding the cognitive component, emotions form as a result of the continuous assessment of events in one's environment (Scherer, 2001). Put differently, the specific type of emotion experienced-that is, the subjective feeling-depends on the evaluation of the event-that is, the cognitive component (Scherer, 2001).
According to attribution theory, people first and foremost evaluate whether an event is a success or failure (Weiner, 1985). In addition, they attribute a specific cause to either result. Starting with the premise that endorsements of scandalous celebrities are failures, recipients are likely to assess the politician's responsibility for the endorsement: could he or she control the endorsement or not (Coombs, 2007)? Weiner (1985) calls this the controllability assessment, which refers to the fact that people can either be in control of events that end in failure or that the events can occur despite a person's ability and volitional control (Weiner, 1979). In the political arena, politicians are clearly in control of self-initiated endorsements implemented of their own will; however, no one will think politicians to be responsible for endorsements clearly initiated by others (Weiner, 1985).
The controllability assessment is crucial because perceived controllability is uniquely related to a set of emotions. In particular, it determines whether people direct anger or pity toward others, or as Weiner (1985, p. 562) put it, "We feel anger toward the lazy [and able] and therefore punish lack of effort, but we feel pity toward the unable and therefore do not punish lack of ability." In other words, anger results when another party is regarded as the cause of a negative state or event, whereas pity results when no other party is responsible (cf. Kepplinger, Geiss, & Siebert, 2012;Nerb & Spada, 2001;Um, 2013;Weiner, 2006). Several studies have supported these propositions, as Weiner's (2000) review has illustrated.
Based on that body of evidence, we proposed a first set of three hypotheses: H1: People perceive politicians to be more in control of self-initiated harmful endorsements than of other-initiated harmful endorsements.
H2: The greater the perceived controllability of a potentially harmful endorsement, the greater the feeling of anger directed toward the endorsed politician.
H3: The greater the perceived controllability of a potentially harmful endorsement, the lesser the feeling of pity directed toward the endorsed politician.
In light of previous research showing that emotional reactions experienced upon encountering a candidate influence attitudes toward the candidate, both anger and pity should strongly affect attitudes (Abelson et al., 1982;Chang, 2001;Glaser & Salovex, 1998;Marcus, 1988Marcus, , 2000Ottati et al., 1992;Ragsdale, 1991). Because anger is of negative valence and pity of positive, the former is assumed to make attitudes less positive, whereas the latter is assumed to make them more positive (Kühne, Schemer, Matthes, & Wirth, 2011;Weiner, 1985). Therefore, if a negative endorsement prompts anger because the politician initiated the endorsement, then the attitude toward him or her should become more negative. The opposite mechanism can be theorized for pity in the case that an endorsement is initiated by another. Two more hypotheses articulate these ideas more formally: H4: The greater the feeling of anger directed toward an endorsed politician, the more negative one's attitude.
H5: The greater the feeling of pity directed toward an endorsed politician, the more positive one's attitude.
As outlined, various meta-analyses have consistently confirmed that attitudes toward a candidate or party are a particularly strong predictor of voting intention (Kim & Hunter, 1993). Accordingly, candidate attitudes are assumed to influence voting intention, as also assumed by the theory of planned behavior (Ajzen, 1991). As a result, we formulated an additional hypothesis: H6: Attitudes positively affect voting intention.
It has also been proposed that the impact of the initiator of the endorsement, the politician or another initiator, and the politician's response (see next paragraph) is moderated by party identification. According to cognitive approaches explaining the formation of emotions, people are more likely to engage in effortful evaluations of stimulus events as the relevance of the respective stimulus increases (Scherer, 2001). Among other things, relevance of information refers to the importance of information to the needs, goals, or values of a person, including his or her social identification (Scherer, 2013). By extension, identification with a group (e.g., a party) is understood as "a feeling of oneness with a defined aggregate of persons, involving the perceived experience of its successes and failures" (Mael & Tetrick, 1992, p. 814). Due to that feeling of oneness, information is perceived to be highly relevant when it relates to a group with whom a person identifies (Mackie, Worth, & Asuncion, 1990). As reviews by Fleming and Petty (2000) and van Knippenberg (1999) have shown, various studies have demonstrated that this dynamic, in turn, precipitates a more effortful processing and evaluation of presented information. Ultimately, more strongly identifying recipients should be more motivated to effortfully evaluate the situation, thereby heightening the impact of the evaluation on perceived controllability and subsequent variables (Chang, 2012). In other words, we hypothesized that: H7: The impact of initiator of the endorsement on controllability and the impact of politician's response on controllability are both moderated by party identification, meaning that their impact increases in parallel to party identification.

Situational crisis communication theory
Provided that politicians are involved in potentially harmful endorsements, whether initiated by themselves or by others, it is of immediate relevance to know how to best respond in those situations. For assistance, situational crisis communication theory (Coombs, 2007), which addresses the protection of reputational assets during a crisis, can provide valuable insights. According to the theory, people or organizations should respond to crises depending on their degree of crisis responsibility. Response strategies range from the denial or diminishment of responsibility in the case of low responsibility to taking full responsibility in the case of high responsibility (Coombs, 2007). Importantly, people or organizations are advised to always answer for their actions in times of crisis. The reasoning for this recommendation is "that negative information is likely to be accepted [and attributed to involved people or organizations] by observers at face value unless countervailing information is also provided" (Bradford & Garrett, 1995, p. 878). As a result, the vast majority of literature addressing crisis communication suggests some kind of response to a negative event as early as possible (e.g., Benoit, 1997;Bradford & Garrett, 1995;Coombs, 1998Coombs, , 2007Dean, 2004;Hale, Dulek, & Hale, 2005;Williams & Treadaway, 1992). Accordingly, practitioners almost always employ one of the suggested response strategies (Kim, Avery, & Lariscy, 2009). However, it remains unknown and has hardly been tested whether no response may be an equally beneficial or even better solution in certain cases. As Geiss (2016) has recently mentioned, any response whatsoever might signal some kind of control or responsibility for a negative event, regardless of whether that conclusion is true, and become tantamount to an admission of guilt. Following the framework provided by Weiner (1985), that outcome should increase anger and decrease pity, thereby resulting in more negative attitudes in either case. It thus seems debatable whether responding to potentially harmful endorsements is actually better than not responding at all. We consequently sought to answer whether politicians should respond to potentially harmful endorsements by revoking them and denouncing the endorsers or by not responding whatsoever, or more formally: RQ1: Should politicians respond to potentially harmful endorsements?
Our complete theoretical framework, including all hypotheses and the research question, is summarized in Figure 1.

Method
To test the hypotheses and form an answer to the research question, we conducted a 2 × 2 between-subject experiment with an additional control condition in Austria. The independent variables were initiator of the endorsement (i.e., self-vs. other-initiated) and politician's response (i.e., no response vs. response). We obtained the moderator party identification as a posttreatment measurement (Carpentier, Roskos-Ewoldsen, & Roskos-Ewoldsen, 2008). It was unaffected by the experimental manipulation, F(4, 279) = 1.17, p = .33; one-way analysis of variance comparing the five experimental conditions.

Study design and manipulation of independent variables
Experimental stimuli encompassed four versions of a half-page news article. The article strongly resembled online versions of a well-known Austrian newspaper in writing and layout (Appendix 1). All articles reported a tax scandal of a well-known Austrian male comedian that the articles claimed would reflect negatively on a certain woman politician recently endorsed by the comedian. The story was completely fictitious, of course, which ensured a fully controlled manipulation. However, both the comedian and politician are actual persons selected based on the results of a pretest administered to 23 undergraduate students, roughly half of whom were women. Participants rated seven Austrian celebrities and seven Austrian politicians in terms of their familiarity as celebrities and politicians, the celebrities' plausibility of endorsing politicians, and the celebrities' plausibility of being involved in a scandal, all on 7-point scales. We selected the comedian because he ranked among the most well-known celebrities (M = 4.65, SD = 2.52), as well as that participants considered it possible that he would endorse a politician (M = 4.79, SD = 1.80) and be involved in a scandal (M = 3.24, SD = 1.30). By contrast, we selected the politician due to her young age, which was rather appealing to the undergraduate participants, and because the pretest revealed her to be poorly to moderately familiar (M = 2.04, SD = 1.66), which rendered pre-existing and fixed attitudes unlikely.
We manipulated the initiator of the endorsement by reporting that it was on the behalf of either the politician (i.e., self-initiated) or the celebrity (i.e., other-initiated). The article clearly stated that the politician (celebrity) initiated the collaboration by publicly announcing it on her or his Facebook page. Stating the manipulation as clearly as possible, the article additionally explained that the endorsement received public media coverage on either the politician's or the celebrity's Facebook page. Meanwhile, we manipulated the politician's response by writing that the politician either revoked the endorsement and denounced the endorser (i.e., response) or did not want to comment on the issue at all (i.e., no response). We chose the particular kind of response according to situational crisis communication theory (Coombs, 2007), which suggests denouncing an external source in the case of no or low responsibility. Altogether, even if the politician was responsible for the endorsement of two of the four conditions, she was clearly not responsible for the tax scandal.
Importantly, the resulting four versions of the article (i.e., self-initiated with a response, self-initiated with no response, other-initiated with a response, and other-initiated with no response) did not differ except in terms of the manipulation of the two variables. There was a fifth condition as a control that received no experimental treatment; its purpose was merely to gauge whether the endorsement was actually harmful. Compared to the control, all four experimental conditions should have led to decreased positive attitudes and intentions to vote for the candidate.

Participants
A total of 280 undergraduate students participated in the study. Current research shows that student samples yield valid results and do "not pose a problem for study's external validity" (Druckman & Kam, 2011, p. 41, emphasis in original). A post hoc power analysis revealed sufficient power to detect all assumed effects (power > .80 ;Faul, Erdfelder, Buchner, & Lang, 2009;Field, 2009). Participants' mean age was 23 years (SD = 3.78), and most were women (84%); however, gender did not affect any of the mediating variables or the dependent variable. In terms of political orientation, participants were, roughly, evenly spread regarding their identification with the Social Democrats, the party of the political candidate (M = 3.60, SD = 1.54; 7-point scale). Participants received course credits for taking part in the study.

Procedure
The study took place at the university's computer laboratory, which featured eight work stations separated from each other by nontransparent, noiseabsorbing dividers. Upon arriving at the laboratory, participants completed an informed consent form, each took a seat at one of the computers, and each began to read one randomly assigned version of the online news article or the control. We told participants that the respective article was recent, that they should read it as carefully as possible, and that they could take as much time as they liked (Coombs & Holladay, 2008;Fang, Singh, & Ahluwalia, 2007). Afterward, we presented participants with an online questionnaire that they completed before we thanked and dismissed them from the laboratory.

Measurement
The exact wording of all measurements can be found in Appendix 2. We measured attitude toward the candidate on a 7-point semantic differential as suggested by Ajzen (2006). We adopted five items from Pease and Brewer (2008) and Veer et al. (2010); they included contrastive pairs such as badgood, negative-positive, and unlikeable-likeable. Reliability was excellent (α = .91, M = 4.20, SD = .96).
We measured intention to vote for the candidate by asking participants how likely they were to vote for her, provided that there was an election next Sunday (M = 2.84, SD = 1.39). We adopted the item from Pease and Brewer (2008) and ranged responses from very unlikely to very likely on a 7-point scale (cf. Ajzen, 2006).
We assessed anger and pity toward the political candidate by adapting items from Weiner, Graham, and Chandler (1982), as well as from the state-trait anger expression inventory (Azevedo, Wang, Goulart, Lotufo, & Benseñor, 2010). We measured each emotion according to three items on a 7-point scale, ranging from I strongly disagree to I entirely agree. Items read, for instance, "I was very annoyed at [name of the politician]" or "I felt sorry for [name of the politician]." Reliability was fine for both measures (anger: α = .86, M = 2.31, SD = 1.26; pity: α = .86, M = 2.86, SD = 1.46).
We measured perceived controllability by adapting four items from the 7point revised causal dimension scale (McAuley, Duncan, & Russell, 1992). According to Weiner (1985), perceived controllability refers to the "controllability of a cause for a negative outcome" (p. 562). The cause of the negative outcome is the celebrity endorsement and the tax scandal. There would be no negative outcome if there would be just the celebrity endorsement without the scandal and there would be no negative outcome if there would be just the scandal without the endorsement. It is the combination of both that matters. Accordingly, we measured controllability of the cause as whether the politician was in control of the celebrity and the tax scandal becoming associated with her. For instance, "[name of the politician] had no control over the tax scandal becoming associated with her" or "[name of the politician] could not prevent [name of the celebrity] from becoming associated with her." Considering that the scale consisted of four items only, reliability was passable (α = .63, M = 3.51, SD = 1.28).
Last, we measured party identification by asking participants how strongly they identified with the Social Democrats, the politician's party (Weisberg, 1980). Identification was middling (M = 3.60 SD = 1.54).
In addition to the dependent variables and measurement of the moderator, the questionnaire contained two questions that served as a manipulation check (Sigall & Mills, 1998)

Manipulation check
The manipulation check indicated the successful manipulation of both independent variables. Eighty-seven percent of the participants in the selfinitiated group perceived the endorsement as initiated by the politician, whereas only 21% reached that conclusion in the other-initiated group. The difference was highly significant, χ 2 (1) = 95.89, p < .001, and demonstrated a large effect (r = .66). Furthermore, 88% of participants in the response group reported that the politician responded to the scandalous behavior, compared to only 8% in the no response condition. Again, the difference was highly significant, χ 2 (1) = 139.04, p < .001, and its effect size was even larger (r = .80). Moreover, the endorsement was actually deemed harmful; participants in all four experimental conditions evaluated the politician as being about a half to one scale point worse on the applied attitude scale compared to the control condition (p < .05, Bonferroni post hoc tests, r = -.27 to -.38). The conditions did not significantly differ in terms of age, F(4, 275) = .43, p = .79; gender, χ 2 (4) = 3.74, p = .44; or party identification, F(4, 275) = 1.17, p = .33.

Hypotheses and research question
We tested the hypotheses and research question by calculating a path model using Mplus software (Muthén & Muthén, 1998. We dummy coded the two categorical independent variables (i.e., self-initiation = 0 and other-initiation = 1; no response = 0 and response = 1). Zero-order correlations of all constructs appear in Table 1. The model depicted in Figure 2 presents unstandardized, as well as standardized, path coefficients in that same order. Fit indices revealed an acceptable overall model fit (Byrne, 1989;Hu & Bentler, 1999), with a standardized root mean square residual (SRMR) of .06, a root mean square error of approximation (RMSEA) of .06, a chi-square of χ 2 (33) = 58.23 (p < .01; χ 2 /df < 2), and a comparative fit index (CFI) of .88. All path coefficients were statistically significant and pointed in the expected direction.
The impact of the initiator of endorsement and politician's response on controllability appeared to be moderated by party identification, as we had assumed (H7). This finding emerged from a highly significant three-way interaction (b = .57, SE = .23, p < .001, unstandardized effect), whose specific nature appears in Figure 3. The figure depicts the impact of the endorsement initiator on controllability at levels of the politician's response (i.e., no response vs. response) and party identification, following the pick-a-point approach (Hayes & Matthes, 2009). For party identification, values of 4, 5, or 6 points on the applied 7-point scale represented moderate, moderate-to-high, and high levels of party identification. For reasons of clarity, we did not depict low levels of party identification (1, 2, or 3) in the figure because there were no significant effects of the initiator of the endorsement and politician's response at levels less than an identification of four, adding no value to the figure (Johnson-Neyman technique; Hayes, 2013). We calculated graph values using the Statistical Package for the Social Sciences PROCESS macro (Hayes, 2013).
As Figure 3 shows, recipients perceived politicians to be more in control of self-initiated endorsements than other-initiated ones, which supports H1. However, that effect was primarily evident when the politician responded to the endorsement. In other words, the politician's response reinforced the effect of the initiator of the endorsement on controllability; in the case of selfinitiation, a response increased perceptions of controllability compared to no response, and participants thought that the politician was even more responsible for the endorsement. By contrast, a response decreased perceptions of controllability compared to no response in the case of other-initiation, and participants thought that the politician was even less responsible. This effect appeared to be stronger in parallel to the level of identification, which accounted for the three-way interaction supporting the seventh hypothesis.  R 2 = .14 R 2 = .02 R 2 = .34 R 2 = .24 R 2 = .07 Figure 2. Unstandardized/standardized path coefficients estimating the impact of harmful celebrity endorsements on intention to vote for the candidate (* p < .05; ** p < .001) (Study 1). Note. Calculating interaction effects by multiplying the main effects and adding the main effects and their product as predictors to the model, path coefficients of main effects do not represent independent effects but dependent effects. Specifically, they display the influence of the respective predictor when the other interacting predictors are zero. As Figure 2 illustrates, controllability affected anger positively (b = .37, SE = .07, p < .001) and pity negatively (b = -.17, SE = .09, p < .05), which supports H2 and H3. Both emotions, in turn, influenced attitudes toward the politician; the impact of anger was negative (b = -.39, SE = .04, p < .001), whereas the impact of pity was positive (b = .18, SE = .04, p < .001). Both results support H4 and H5. We also tested whether the effects of anger and pity persisted when adding the direct effect of party identification on attitude toward the candidate. Past research revealed a substantial influence of party identification on candidate evaluations (e.g., Hart, Ottati, & Krumdick, 2011). The model was recalculated adding the direct effect (SRMR = .04; RMSEA = .04; χ 2 (32) = 45.43, p = .06; χ 2 /df < 2; CFI = .94). The direct effect was significant and positive (b = .12, SE = .03, p < .01). More importantly, the effects of anger (b = -.36, SE = .04, p < .001) and pity (b = .17, SE = .04, p < .001) persisted almost unreduced in their size. That is, emotions significantly added to the explanation of candidate attitudes next to the influence of party identification. Finally, attitudes positively influenced intention to vote for the candidate, as assumed in H6 (b = .71, SE = .09, p < .001). Our model ( Figure 2) eventually explained 34% of the variance of attitudes toward the politician and 24% of the variance of intention to vote for the candidate.
In addition to the significance assessment of single-path values, we also calculated indirect effects. It was thus possible to test whether initiator of the endorsement indirectly affected intention to vote for the candidate via controllability, emotions, and attitudes. Again, we calculated indirect effects and corresponding 99% bias-corrected confidence intervals at different levels of the moderators of politician's response (i.e., no response vs; response) and party identification (i.e., 2, 3, 4, 5, and 6 points on the 7-point scale). We calculated confidence intervals by drawing 5,000 bootstrap samples and following the procedures of Hayes and Preacher (2013). Results appear in Table 2. Findings matched the previously examined interaction insofar as Note: Point estimates represent unstandardized effects. CI: confidence interval; LL: lower limit; UL: upper limit only moderately, moderately-to-highly, and highly identifying recipients showed highly significant increased intentions to vote for the politician when the endorsement was other-initiated compared with when it was selfinitiated; all confidence intervals were entirely above 0. Again, however, this result emerged only when the politician responded to the endorsement, which underscores the reinforcement effect of the politician's response point estimates can be considered to be of medium size given that they present triple-mediated effects. In addition, about 13% of the participants may have missed the treatment (cf. manipulation check). Hence, effects may have been stronger if all participants would have been able to correctly perceive the treatment. There was neither a moderated (β = -.18, SE = .87, p = .50) nor an unmoderated (β = .10, SE = .61, p = .10) significant direct effect of initiator on intention to vote for the candidate.

Discussion
As predicted, the results of Study 1 showed that endorsement initiation (selfvs. other-initiated) was of crucial importance regarding participants' perceived controllability of a potentially harmful endorsement. In line with attribution theory (Weiner, 1985), the particular assessment of the situation (perceived controllability) had an influence on participants' emotional reactions. More precisely, when participants regarded the politician to be responsible for a harmful endorsement, they pitied her to a lesser extent. In contrast, controllability affected anger positively, revealing that participants were annoyed at the politician when they thought that she was responsible for the endorsement. Emotions, in turn, positively (pity), respectively, negatively (anger) influenced participants' attitudes toward the politician and, eventually, influenced particular intentions to vote for the politician. However, the politician's response to the endorsement, as well as voters' party identification, moderated the decisive situation evaluation. The former changed the situation to be evaluated, whereas the latter made cognitive evaluation more relevant. As for party identification, the impact of the initiator of the endorsement and politician's response increased with increased party identification. Furthermore, only moderate to highly identifying voters appeared to be significantly influenced.
Furthermore, the results show that if the politician initiated the endorsement, then a revoking, denouncing response reinforced the fact that the politician initiated and was in control of the endorsement. By contrast, if the endorsement was initiated by another-in our study, a celebrity-then the response emphasizes that the politician was not in control. As outlined, perceptions of high control corresponded with anger and prompted decreased voting intentions, whereas low control prompted pity and increased intentions to vote for the candidate. Important to note, denouncing and revoking an other-initiated, harmful endorsement does not lead to more positive attitudes and voting intentions than no endorsement (control condition). All things being equal, it is better not to have a flawed endorsement than to have one and repudiate it.
In Study 1, participants were exposed to one news article and a particular scandal only. Previous research suggests that using multiple stimuli (e.g., news stories) may increase ecological validity in media psychological studies (Reeves, Yeykelis, & Cummings, 2016). Furthermore, the importance of replication in experimental research has been emphasized (Matthes et al., 2015;Stroebe & Strack, 2014). Thus, to test the robustness of the effects detected in Study 1, we conducted a conceptual replication of the experiment with the help of a second study.

Method
The second study presents a conceptual replication of the first study, the only difference being that the politician was involved in a real estate scandal instead of a tax scandal and that controllability was measured more comprehensively (cf. Appendix 1). A total of 212 undergraduate students participated in the study at the University of Vienna (Austria). Participants' mean age was 22 years (SD = 7.11), and most were women (81%). Measures were as reliable as in the first study: attitude toward the candidate (α = .93, M = 4.07, SD = 1.05); intention to vote for the candidate (M = 2.63, SD = 1.40); anger (α = .87, M = 2.39, SD = 1.39); pity (α = .89, M = 2.90, SD = 1.53); controllability (α = .74, M = 3.88, SD = 1.21); and party identification (M = 3.43, SD = 1.59).

Manipulation check
The manipulation check indicated a successful manipulation of both independent variables. Eighty-two percent of the participants in the self-initiated group perceived the endorsement as initiated by the politician, whereas only 9% reached that conclusion in the other-initiated group. The difference was highly significant, χ 2 (1) = 112.39, p < .001, r = .73. Furthermore, 89% of the participants in the response group reported that the politician responded to the scandalous behavior compared to only 7% in the no response condition. Again, the difference was highly significant, χ 2 (1) = 136.48, p < .001, r = .82. Furthermore, the conditions did not significantly differ in terms of age, F(3, 208) = 1.50, p = .22; gender, χ 2 (3) = .03, p = .99; or party identification, F(3, 208) = .04, p = .99.

Hypotheses and research question
We tested the hypotheses and research question applying the same procedures as in study 1. This time there was no significant three-way interaction between initiator of the endorsement, politician's response, and party identification when calculating their impact on controllability (b = .28, SE = .15, p = .13, unstandardized effect). There were neither any two-way interactions. That is, party identification did not show any moderating influence, as assumed in H7. In addition, politician's response did not alter the effect of initiator of the endorsement. We consequently calculated the path model without the interaction terms, looking solely at the main effect of initiator of the endorsement. Results are depicted in Figure 4. Zero-order correlations of all variables included in the model are presented in Table 3. Fit indices revealed a good overall model fit (SRMR = .06; RMSEA = .05; χ 2 (9) = 14.15 (p = .12; χ 2 /df < 2), CFI = .98).
Results were in line with study 1: Recipients perceived politicians to be less in control of other-initiated endorsements as compared with selfinitiated ones, supporting H1 (b = −1.15, SE = .15 p < .001). Controllability affected anger positively (b = .30, SE = .07, p < .001) and pity negatively (b = −.31, SE = .08, p < .05), which supports H2 and H3. Both emotions, in turn, influenced attitudes toward the politician. The   impact of anger was negative (b = −.25, SE = .05, p < .001), whereas the impact of pity was positive (b = .23, SE = .04, p < .001). Both results support H4 and H5. Finally, attitudes positively influenced intention to vote for the candidate, as assumed in H6 (b = .79, SE = .07, p < .001). Eventually, the indirect effect of initiator of the endorsement on intention to vote for the candidate came to an estimate of .13. The corresponding 99% bias-corrected confidence interval did not included zero (.06, .25). In addition, there was no direct effect.

Discussion
The purpose of Study 2 was to replicate the findings of the first experiment. As in Study 1, the results supported hypotheses one through six, showing that initiation of the endorsement decisively affected controllability and, in turn, particular emotional reactions of the participants. Anger was positively related to more negative attitudes toward the candidate; pity affected candidate attitudes more positively. Again, positive (negative) attitudes showed the expected effects on voting intentions. However, we did not find the threeway interaction between initiator of the endorsement, politician's response, and party identification we detected in Study 1. This insignificant finding may be explained in two different ways. First, Study 2 may have simply failed to exactly replicate the three-way interaction detected in the first experiment. As pointed out by Stroebe and Strack (2014) "failures to replicate are puzzling, but in most sciences, empirical findings cannot always be replicated" (p. 60). Nevertheless, nonreplications are equally important to replications in understanding the boundaries of an effect, as well spurring additional research. With the conflicting evidence at hand, future research should explore these aspects in depth. Second, participants were exposed to a different type of scandal in Study 2. Although the case used in Study 1 (tax scandal) was similar to the scandal in Study 2 (real estate scandal), the cases differed regarding their social costs for the public. Social costs for the public were rather high in Study 1 (tax evasion damaging Austria's finances in the amount of 370.000 €), compared to Study 2 (no costs for the public, but only for an individual investor in the amount of 370.000 €). The particular response and identification with the party (Social Democrats) of the politician obviously played a more decisive role in combination with a scandal that caused significant social costs for the public (Study 1). One may argue that, especially, a social democratic or labor politician's response becomes important when scandalous actions involve high (compared to rather low) social costs for the public (also see Entman, 2012). However, future studies should test this assumption in more depth varying types of scandals, politicians' party affiliation, as well as response strategies (also see General Discussion).

General discussion
Thus far, researchers have completely ignored the possible negative consequences of celebrity endorsements in politics. Accordingly, we aimed to investigate the impact of potentially harmful endorsements on voters' attitudes and voting intentions and tested whether self-or other-initiated endorsements affected voters differently, as well as how politicians should best respond to such endorsements. In addition, we sought to reveal the underlying psychological mechanism by modeling perceived controllability, as well as anger and pity, as emotional responses.
In agreement with attribution theory (Weiner, 1985), we confirmed that politicians are perceived to be more in control of self-initiated endorsements than other-initiated ones. Perceived controllability, in turn, influenced feelings of anger and pity. The more that participants thought that the politicians were responsible for a harmful endorsement, the less that they pitied the politicians and the more that they were annoyed at them. That is, the specific type of emotion experienced depended on the respective cognitive evaluation of the situation (i.e., perceived controllability). This finding is in line with current cognitive approaches to emotion formation (Scherer, 2001). Such research holds that formed emotions, in turn, lead to worsened or enhanced attitudes, which are based largely on the emotional responses experienced upon encountering an attitude object (Olson & Kendrick, 2008). Eventually, attitudes prompt respective intentions to vote for the candidate (Ajzen, 1991). In sum, the initiation of a harmful endorsement (i.e., self-vs. otherinitiated) indirectly influences voting intentions, whereby people are more likely to vote for a politician when a harmful endorsement was initiated by somebody other than the politician.
However, the findings of Study 1 also suggest that under particular circumstances (i.e., tax scandal with rather high social costs) the politician's response to the endorsement, as well as voters' party identification, moderated the decisive situation evaluation. The former changed the situation to be evaluated, whereas the latter made cognitive evaluation more relevant. As for party identification, the impact of the initiator of the endorsement and politician's response increased with increased party identification. Furthermore, only moderate to highly identifying voters appeared to be significantly influenced. Taken together, the results of Study 1 support our reasoning that people are more likely to engage in effortful evaluations of celebrity endorsements in politics if they identify with the respective politician or party-that is, if the politician or party is personally relevant to them (Ellemers & Haslam, 2012). More effortful processing then leads to stronger effects or else enables effects in the first place (Scherer, 2001). However, these processes may also be dependent on the type of scandal allegations, because we did not find any effects of candidate responses and party identification in Study 2 (real estate scandal). Additional research is needed to clarify the role of particular political responses and party identification.
In terms of responses to harmful endorsements, it has been unclear how politicians should best react. Whereas the vast majority of literature recommends some kind of response no matter what (e.g., Benoit, 1997;Coombs, 1998Coombs, , 2007, recent approaches to political scandals have suggested that a response may signal some kind of responsibility and is tantamount to an admission of guilt (Geiss, 2016). The findings of Study 1 imply that both suggestions are appropriate depending on who initiated the endorsement; if the politician initiated the endorsement, then a revoking, denouncing response reinforced the fact that the politician initiated and was in control of the endorsement. By contrast, if the endorsement was initiated by another -in our study, a celebrity-then the response emphasizes that the politician was not in control. As outlined previously, perceptions of high control corresponded with anger and prompted decreased voting intentions, whereas low control prompted pity and increased intentions to vote for the candidate. It is important to note that denouncing and revoking an other-initiated, harmful endorsement does not lead to more positive attitudes and voting intentions than no endorsement (control condition). All things being equal, it is better not to have a flawed endorsement than to have one and repudiate it. Again, Study 2 showed no particular effects of the politician's response strategy and future research is needed.
The effects we found were mostly highly significant and moderate to large in terms of their size (Cohen, 1992). We are thus quite confident that there are no issues with Type 1 error in our study, as mainly studies with small effect sizes are prone to fail replication (Killeen, 2005).

Implications
Results suggest several implications. In terms of theory, Weiner's (1985) attributional theory of achievement motivation and emotion provided a fruitful framework to model the effects of scandalous celebrity endorsers on intention to vote for the endorsed candidate. Moreover, it provided much needed insights about the mechanisms that mediate the relationship between scandalous celebrity endorsers and voting intention: It is the voters' assessment of controllability of the endorsement (i.e., responsibility for the endorsement) that determines the basic emotional reaction directed toward the politician (anger or pity). Emotions, in turn, affect attitudes and intention to vote for the candidate. The fact that there were no direct effects of the endorsement (full mediation or indirect only mediation) underscores the importance of these mediating mechanisms (Zhao, Lynch, & Chen, 2010). Investigating such mediating mechanisms, enabled the detection of the indirect effects in the first place (Hayes, 2013). Future studies are encouraged to use this framework, too (see also Um, 2013).
Looking at situational crisis communication theory (Coombs, 2007), our results implicate that no response might be added as a further valid response strategy in times of crisis. Given that a politician is clearly coresponsible for the negative outcome of an event, this strategy proved itself better than simply denouncing the other party involved in Study 1 and had no negative effects in Study 2. Admittedly, showing no response will leave voters to evaluate the situation unassisted. They are likely to accept any negative appraisal presented by media (Bradford & Garrett, 1995). However, responding at all emphasizes that the politician was at least partially in control of the endorsement, and thus responsible for its negative outcome (Weiner, 1985). As a result, attitudes toward the politician become even more negative as compared to no response (Study 1). It is important to note that there may be strategies that are still more effective than no response. Researchers should, therefore, consider applying a broader range of politicians' responses. In that way, they can test whether there are valuable alternatives to reacting to harmful endorsements. Apologizing for a self-initiated endorsement may be a third option, next to no response or revoking the endorsement. When politicians admit full responsibility for endorsements, they can blunt feelings of anger (Coombs, 2007). Although that strategy is rarely used in communication practices, it has proven to be effective (Kim et al., 2009).
Looking at political practice, politicians need to be aware of the possible negative consequences of celebrity endorsements. Findings show that celebrities involved in scandals negatively affect perceptions of politicians whom they endorse, regardless of whether the politician or celebrity initiated the endorsement. Moreover, politicians need not be involved in a scandal to be negatively affected, but appear to be guilty by association (Carrillat et al., 2014;Puente-Díaz, 2015). At the same time, they can reduce the negative impact of the endorsement on their image by reacting appropriately under particular circumstances. If there is clear evidence that an endorsement was initiated on behalf of politicians, then politicians may decide to not comment on the endorsements (instead of revoking and denouncing the celebrity) to prevent potential unwanted consequences like becoming associated with a celebrity scandal. By contrast, the results of Study 1 suggest that politicians may decide to respond to harmful endorsements, as long as there is evidence that a celebrity initiated them on his or her own. The politician's response then emphasizes his or her inability to control the situation, which generates pity and increased positive attitudes as well as voting intentions (Weiner, 1985).
The growing engagement of celebrities in politics (Chou, 2014;Pease & Brewer, 2008;Štechová & Hájek, 2015), as in the earlier examples involving Hillary Clinton and David Cameron, show that harmful endorsements are quite likely in today's politics. Indeed, their occurrence will likely increase. Contemporary means of communication enable celebrities to easily post political endorsements on Facebook and Twitter without a politician's even knowing. Celebrities talk directly to their fans, and such messages are liable to find audience in traditional news media (Loader et al., 2016). As a result, politicians and their staff are far from in control of every endorsement made. At the same time, politicians face competition for the online attention of voters from celebrities, which makes it increasingly difficult to get their own messages out (Loader et al., 2016).

Limitations and future research
The study is certainly subject to some limitations. Because we used a student sample, future studies should include older and less educated people, as well. Nevertheless, some authors have argued that celebrities may be particularly helpful when it comes to engaging young citizens with politics (Loader et al., 2016), for their kind of appeal stimulates young people in a culture familiar to them and that they consider to be theirs (Austin, Van de Vord, Pinkleton, & Epstein, 2008). Accordingly, young citizens, such as university students, may be the group of voters to which celebrity endorsements are most relevant after all (Boon & Lomore, 2001;Payne, Hanlon, & Twomey, 2007).
Furthermore, future studies should examine the influences of different types of politicians (e.g., male candidates), different kinds of scandals, and different celebrities (male/female, actors, sport celebrities etc.), to test if these findings can be replicated. They could also measure whether being linked to a scandal does not only affect emotional responses, but also cognitive responses like, for instance, the assessment of competency (von Sikorski, 2016;von Sikorski & Hänelt, 2016). Competence is considered as one of the most important trait attributes on which people evaluate politicians. Furthermore, it was shown to substantially influence voting decisions (Todorov, Mandisodza, Goren, & Hall, 2005).
Scholars should consider also examining politicians that are better known -that is, ones with whom voters share strong relationships. Although we employed a real politician-namely, a member of the national parliamentshe was only poorly to moderately known. As a result, our study provides valuable insights especially for up-and-coming young politicians or politicians who are not well known. In that sense, it remains untested whether results hold true for famous, well-known politicians. As research from sport marketing has revealed, familiarity with endorsed brands or events tempers negative consequences of harmful celebrity endorsements (Doyle, Pentecost, & Funk, 2014), and the same could pertain to endorsements in politics.
Moreover, we examined a female politician only. However, research from negative campaigning suggests that voters are more receptive to negative messages when men are under attack, instead of women. That effect is due to the fact that negative messages are more congruent with stereotypes of men (Fridkin, Kenney, & Woodall, 2009), and the same could hold true for harmful endorsements of men, whose effects may even be stronger than for women. Furthermore, future research should also examine the role of trust in news media when studying the effects of celebrity news. Trust may moderate the effects observed here (see Daniller, Allen, Tallevi, & Mutz, 2017;Kohring & Matthes, 2007). Last, future research may examine how negative endorsements by individuals or groups that are predominantly evaluated in negative ways affect political candidates. A politician's best reaction to such an endorsement (e.g., racist endorsements for Mr. Trump by racist groups in the recent Presidential election) may considerably vary from the findings of this study.
In that respect, we hope that our research inspires further studies examining the negative consequences of celebrity endorsements, particularly given the timeliness of the topic and the areas of research that remain unaddressed.