The impact of public approval on the use of negativity throughout the electoral cycle

Abstract The literature has found that politicians who lag behind in public approval ratings during campaigns resort to more negativity. However, the actual impact of approval on the use of negativity during the electoral cycle has yet to be addressed. Furthermore, due to the short-lived nature of campaigns, current studies have been unable to establish a directional causal link between approval ratings and negativity. This article addresses these gaps by: (i) building a theory for understanding the impact of public approval on the use of negativity throughout the electoral cycle; and (ii) methodologically testing this impact on a time series basis. Using data on negativity in parliaments, the results confirm that low approval ratings lead to more negativity closer to elections in Belgium (2014–2020) and Croatia (2010–2021). In the UK (2010–2020), however, approval does not appear to be a significant predictor of negativity use. These findings have important implications for our understanding of the use of negativity by political actors outside campaigns.

Negativity in politics has been blamed for lower citizen turnout at elections (e.g. Lemert et al. 1999;Nai 2013), decreasing political trust (e.g. Thorson et al. 2000) and increasing polarisation (e.g. Iyengar et al. 2012). Despite knowing much about the impact that negativity has on citizens, not much is known about the impact of citizens' opinions on politicians deciding to go negative in the first place. This is surprising as there is plenty of literature on political behaviour that has confirmed the notion that politicians do respond to citizens' preferences (e.g. Wlezien and Soroka 2016). This begs the question: Could it be that citizens themselves (and their approval) are an important predictor of politicians deciding to use negativity in the first place?
In order to address this question, negativity needs to be treated as a dependent variable. Thus far, the only literature that has extensively tested negativity in such a way has focussed on negative campaigning. These scholars studied negativity during election campaigns by collecting and analysing attacks by political actors on their opponents (Nai and Walter 2015). They have theorised that one of the explanations for going negative is the status these actors have in terms of public approval, operationalised through opinion polls and/or election results. More specifically, it has been hypothesised that the frontrunners, who score higher in opinion polls and are expected to gain a larger share of votes compared to a previous election, are less likely to use negative campaigning. In contrast, it is suggested that actors who are lagging behind in opinion polls are more likely to go negative (Skaperdas and Grofman 1995). These expectations have found support in several empirical studies (e.g. Damore 2002;Elmelund-Praestekaer 2008, 2010Haynes and Rhine 1998;Maier and Jansen 2017;Nai and Sciarini 2018;Walter and Van der Brug 2013) and confirm the expectation that public approval is an important element that politicians evaluate before deciding to utilise negativity.
While these findings are fundamental for our understanding of the use of negativity, two gaps in the research need to be addressed. First, the above-mentioned studies only explored the impact of public approval during election campaigns. Hence, the conclusions we currently have are only tied to a specific moment in time. However, citizens' opinions about political actors are not only formed during campaigns but also during regular day-to-day politics. As such, politicians are exposed to public approval signals during the entire electoral cycle, such as polls in the media (see Oleskog Tryggvason 2020). The impact of these daily approval signals on the use of negativity is yet to be investigated. Second, due to the limited time frame of campaigns, researchers have predominantly had to approach the operationalisation of public approval in a static and binary fashion, where some actors are classified as frontrunners and others are considered stragglers throughout the entire campaign (e.g. Elmelund-Praestekaer 2010; Nai and Sciarini 2018;Walter et al. 2014). Therefore, the causal link which is presumed (i.e. public approval impacting negativity) is based on the congruence between the two and not on directional causation. Therefore, the intriguing results of these founding studies deserve some in-depth follow-up research. The aim of this article is thus to: (i) build a theoretical framework (following prospect theory of risk-taking) about the causal relationship between public approval and negativity during the entire electoral cycle, while (ii) methodologically studying the interplay between the two in a time series fashion. The main argument of this study is that public approval has an impact on political actors going negative not only during campaigns but also throughout the electoral cycle. However, it is expected that this impact of public approval on negativity shifts and changes during the cycle. More precisely, this article argues that as time during the electoral cycle elapses, the impact of public approval on the use of negativity grows. Those actors who enjoy high approval ratings are encouraged to use less negativity because they fear the potential risks that negativity runs (i.e. fewer votes at the upcoming election). In contrast, actors who have low approval ratings become risk-takers later in the electoral cycle, using negativity in the hope of increasing their approval while damaging that of their competitors.
To test this hypothesis, the study made use of data on negativity employed by politicians in Belgium (2014-2020), Croatia (2010 and the UK (2010-2020) during 'question time' sessions (QT) in their respective parliaments. The amount of negativity that politicians employed during QTs (t) was regressed on the most recent public approval ratings that preceded these QTs (t-1). The results showed that negativity used by politicians was, surprisingly, not always affected by approval polls. However, there was a significant effect of approval later in the electoral cycle and as the probability of an upcoming election increases. Furthermore, the notion that higher approval leads to less negativity and vice versa later in the cycle was identified in Belgium and Croatia but not in the UK, where the study found that as time elapsed during the cycle, both high and low approval actors tended to use more negativity. These results have important implications for our understanding of the use of negativity by political actors, while also contributing to the broader literature that studies the linkage between politics and citizens.

Election campaigns
Negativity is one of the key strategies used by political actors during election campaigns to gain election victory. It can be defined in a 'directional' way (Walter and Vliegenthart 2010) as any form of political attack in which one political actor criticises the competition during the campaign (Geer 2006). Citizens are exposed to these attacks through TV adverts, election posters, TV debates, etc., during campaign periods. Exposure to this negativity in politics can have harmful effects on the democratic process (e.g. adversely affecting political trust; Mutz and Reeves 2005) and also on society as a whole (e.g. increasing affective polarisation; Iyengar et al. 2012). There is extensive research on negativity during campaigning, in which scholars have attempted to establish what leads politicians to resort to such a campaign strategy. Given the impact that negativity has on citizens, these researchers have also looked at how citizens and their approval ratings have an impact on the use of negativity. Among the very first studies was one by Harrington and Hess (1996), who empirically showed for the US that candidates who were ranked low by citizens on the valence scale (i.e. competence) tended to be more negative. This prompted the hypothesis that greater use of negativity and low public approval may be spuriously related, that is, they are both driven by candidates' inabilities (Harrington and Hess 1996: 221).
This hypothesis, however, was quickly abandoned, with scholars arguing that it is simply the low approval of the public that leads actors to employ negativity. Skaperdas and Grofman (1995), for example, had argued that in a two-candidate race, the candidate that lags in opinion polls is more likely to utilise negativity. The reason for this outcome is simple: if the runner-up wants to be the frontrunner, they need to discredit the current frontrunner in the perception of the electorate. As such, using negativity to attack the frontrunner is a rational strategy, as it may lower citizen approval of the frontrunner (Lefevere et al. 2020;Seeberg and Nai 2021). Another option for the runner-up would be to focus the campaign on positive messages instead (praising their own achievements/commitments), but this would hardly be an effective strategy if the public already has a more positive image of the frontrunner (Damore 2002). This argument was also tested beyond mere two-candidate races and found to be true for US primary races, where multiple candidates run to become their party's presidential candidate (Haynes and Rhine 1998).
As the literature on negative campaigning started to move beyond the US, the impact of public approval on negativity was also tested in European states. These studies also identified a significant relationship between public approval and the use of negativity during campaigns in Europe. To be precise, parties and candidates across Europe use more negativity if they are expected to lose the election (Elmelund-Praestekaer 2008, 2010; Maier and Jansen 2017; Nai and Martínez i Coma 2019; Nai and Sciarini 2018; Walter and Van der Brug 2013). 1 Most recently, Nai (2020) identified the same relationship, not only in Western (US and European) campaigns but also across the globe. Therefore, it is safe to conclude that the hypothesis about the relationship between negativity and public approval stands in the majority of cases: low public approval is associated with candidates going negative.
Despite these findings, it remains unclear what happens with negativity once campaigns are over. Negativity is employed not only during campaigns but also in regular day-to-day politics (Ketelaars 2019). Venues in which political actors are present between campaigns, such as the media or parliaments, present opportunities for politicians to utilise negativity during their parliamentary term. We know this from various other fields of political science that have shown how, for example, there is a heavy negativity bias in the news (Soroka and McAdams 2015), resulting in negative communication by politicians who aim to obtain media access (Haselmayer et al. 2019). At the same time, the literature on parliamentary behaviour has also established that parliaments, for example, are venues in which clashes and conflicts between actors take place (Otjes and Louwerse 2018;Vliegenthart and Walgrave 2011). Therefore, it is safe to say that there is negativity throughout the electoral cycle, leading us to wonder whether public approval also impacts this strategy.
In addition, public approval has been operationalised in a static way in the current research, which makes it difficult to establish the direction of causality, whereby approval impacts negativity (but see Maier and Jansen 2017; Nai and Martínez i Coma 2019). Scholars have tended to either classify actors as frontrunners or losers based on average public approval polls during elections (e.g. Nai and Sciarini 2018) or have calculated scores about their standing by, for example, comparing polls during elections to the vote share from the previous election to determine expected gains or losses (e.g. Elmelund-Praestekaer 2010; Walter et al. 2014). While this approach allowed scholars to establish that there is a congruent relationship between low public approval and high use of negativity, recalling the initial point of Harrington and Hess (1996), it can be argued that this relationship may not be directionally causal from approval to negativity. It could be argued, for example, that during the actual campaign citizens approved of the actor who was less negative, making them the frontrunner and/or putting them in a gain position. Still, it should be noted, however, that the current approach to the operationalisation of public approval is understandable, as campaigns outside the US tend to last for only a few weeks, which can make it difficult to study public approval dynamically (a similar problem was also encountered by public opinion scholars; see Wlezien and Soroka 2016).

Electoral cycle
In order to address the two gaps identified, this article analyses the occurrence of negativity throughout the entire electoral cycle (1), while hypothesising about the directional causal relationship between public approval and negativity (2). As such, my definition of negativity follows campaigning studies, in which it is seen as occurrences of politicians publicly attacking each other. However, my definition does not specify the time period in which the attacks occur (i.e. it may be during a campaign or between campaigns) nor does it specify that criticism must be directed towards political competitors (i.e. criticism of partners and party colleagues are also included).
This broader definition is suitable for the purpose of this study as it allows us to analyse negativity in politics that takes place between actors in day-to-day politics (parliamentary debates, press releases, media interviews, cabinet meetings, etc.). Furthermore, this definition encompasses the strategic use of negativity, where actors direct negativity to internal party or coalition politics. For example, in systems where politicians are elected in single-member districts, parties may tolerate internal attacks, as this possibly allows a party to preserve its seat in the parliament (Proksch and Slapin 2012). In addition, coalition partners in government may go negative towards their partners to prevent drift from coalition agreements which may hurt their re-election chances (e.g. Höhmann and Sieberer 2020).
In order to investigate this broader definition of negativity and its interplay with public approval, the study employs the prospect theory of decision making, which argues that individuals who expect to gain are risk-averse, while individuals who expect to lose are willing to take risks (Kahneman and Tversky 1979). To determine if one is expected to gain or to lose, a reference point is chosen. Compared to that reference point, a decision-maker is either in a gain situation or in a loss situation, which determines whether they opt for a risk-taking or risk-averse decision strategy (Vieider and Vis 2019). For example, if the deviation from the reference point is beneficial for the decision-maker (gain), they become risk-averse in order not to damage the current favourable position. In turn, if the deviation from the reference point is not favourable (loss), they become a risk-taker.
Applying this general framework of prospect theory to negativity in politics, it can be argued that the reference point is the approval rating of political actors. An actor with a high approval rating would be in a gain situation, while an actor with a low approval rating would be in a loss situation. Therefore, we can hypothesise that politicians will decide whether to risk using negativity or not based on their approval rating. The risk of using negativity is grounded in the backlash effect associated with negativity, as citizens may not approve of actors who go negative. For example, a meta-analysis of the literature on the effects of negativity by Lau et al. (2007Lau et al. ( : 1180Lau et al. ( -1183 found that those who attack others usually experience a decrease in citizen approval (out of 40 situations, 33 were characterised by decreasing approval for the attacker). This is in line with some contemporary studies that also looked at the backlash effect (e.g. Nai and Maier 2021), which confirmed that some voters may choose to vote for a less negative candidate (Walter and van der Eijk 2019). Therefore, given that going negative does not always lead to a desirable outcome, actors with high public approval are likely to be risk-averse and not willing to go negative. In contrast, actors with low public approval are willing to risk going negative in the hope of decreasing the approval of their competitors and/or increasing their own approval.
As mentioned above, this part of the theory is already addressed in the literature. However, this current framework is lacking some nuances, if we want to expand our understanding of the use of negativity during the entire election cycle. More specifically, there are reasons to argue that in addition to the reference point (which mainly concerns the approval rating), proximity to the next election also affects the use of negativity. We know from studies on campaigning that, as the election date approaches, more negativity will be employed (Damore 2002;Nai and Sciarini 2018). However, since campaigns are limited in time, scholars have been unable to provide a deeper theoretical explanation for the joint impact of approval and elections on negativity use (i.e. approval ratings and the election date were usually studied separately). Therefore, it is fundamental to evaluate such interplay given that there are reasons to expect shifts in political responsiveness to the public throughout the electoral cycle (Pardos-Prado and Sagarzazu 2019). This is why this study extends the theoretical notion that approval impacts negativity during the electoral cycle by borrowing the prospect theory concept of probability weighting (see Vieider and Vis 2019). 'Probability weighting' means calculating whether the loss or gain outcome is going to occur and adapting risk behaviour accordingly (e.g. higher probability of gain leads to higher risk aversion). It is argued that the probability weighting of the outcome is associated with the electoral cycle (i.e. proximity to the next election). As time elapses during the electoral cycle, there is an increasing probability that the election will be held, meaning that an actor with high approval has a higher chance of achieving electoral victory and suffering from the backlash effect if they go negative. This is unlike early in the cycle, when such probability is low despite having gains (i.e. there are no elections to be won). In contrast, an actor with low approval knows that with a higher probability of losing the election as time elapses, the more risk they will have to take, but this is not necessary early in the cycle.

Hypotheses
This theoretical outline leads to two overarching suggestions. First, changes in public approval of actors at time t − 1 (reference point) have an impact, stimulating these actors to use more or less negativity at time t (risk), forming the first hypothesis: H1: Low public approval leads political actors to use more negativity, compared to high approval, which leads political actors to use less negativity.
Second, the decision to use negativity based on previous public approval depends on the electoral cycle (probability weighting). The risk of negativity is expected to be low at the start of the electoral cycle due to the low probability of a new election. Naturally, as time elapses during the cycle, the probability of winning or losing the upcoming election increases, leading actors with low approval to become risk-takers, while those with high approval become increasingly risk-averse. As such, this forms the second hypothesis: H2: The effect of public approval on the use of negativity (see H1) becomes stronger as time elapses during the electoral cycle.

Cases
The hypotheses were tested in three different country case studies in Europe: Belgium, Croatia and the UK. One of the most profound differences across these three countries concerns the party systems in which politicians function and compete between themselves. Belgium is characterised by an extremely fragmented multi-party system that requires politicians to cooperate to gain office. The UK, in contrast, has a two-party system in which one of the two dominant parties assumes office on its own. Croatia, somewhere between the two extremes, has a multi-party system (like Belgium), but most parties group into two blocks -each of which is led by a dominant party which cooperates with the other smaller partners (like the UK).
In each country, the focus is on politicians going negative in federal/ national parliament during question time sessions (QTs) (Belgium: Vragenuur; Croatia: Aktualno Prijepodne; UK: Prime Minister's Questions). QTs are useful to study the impact of approval ratings on politicians, since we know that QTs tend to be the parliamentary activity most exposed to the media. This allows politicians to communicate with voters (Salmond 2014). As such, it comes as no surprise that research has demonstrated that politicians use QTs in order to achieve their vote-seeking goals. For example, a survey of politicians revealed that they believe citizens do pay attention to their activities during QTs due to the media exposure and their own promotion of questions on social media (Soontjens 2021). Studies of parliamentary speeches have also shown that politicians use more emotional rhetoric during QTs compared to other debates, and they do this to appease voters (Osnabrügge et al. 2021). Studies have also shown that politicians from the opposition use QTs to discredit the government in the eyes of the electorate .
There are, however, substantive differences in how QTs are structured in each of these countries (e.g. Serban 2022), which provides a diverse setting to test the theory and hypotheses. In Belgium, QTs take place every week and MPs put questions in groups (based on a topic) to one or several members of the government. Once their questions are answered, MPs are granted a rebuttal. Each party group is permitted to ask questions, regardless of their size in the parliament. Although Croatia has a similar approach, QTs only take place once every two to three months and last for an entire day. The number of questions (out of a total of 40) is allocated to each party group based on the share of seats in the parliament. As such, this approach does not allow small third parties a lot of room for manoeuvre during QTs. Similar to Belgium, MPs are allowed a rebuttal to cabinet members' answers. Finally, the Prime Minister's Questions (PMQs) in the UK take place every week, where opposition leaders and MPs from both the majority and the opposition ask questions of the PM. Only opposition leaders are permitted a rebuttal. Similar to Croatia, PMQs also tend to favour the two main parties, with smaller parties in the opposition often being neglected.

Data
In order to study the share of negativity employed by actors during QTs, a quantitative content analysis was performed on transcripts from these sessions in parliament. QTs were sampled by selecting one in each month from January 2010 to December 2020 (2021 for Croatia), which resulted in the following sample: 103 QTs in Belgium (30.5% of all Belgian QTs), 43 QTs in Croatia (100%) and 115 QTs in the UK (32.7%). Every speech contribution was scrapped as an observation (N = 20,044) during these QTs from official parliamentary websites (for Belgium, dekamer.be; for Croatia, edoc.sabor.hr; for the UK, hansard.parliament.uk). This 'raw' data showed what each politician said during a particular QT (Belgium N = 6,634; Croatia N = 5,679; UK N = 7,731). Therefore, this data included every possible speech contribution, from formal speeches, such as questions and answers, to informal ones, such as interruptions and speakers' interventions (see Fernandes et al. 2021). Protocol speeches, such as speakers moderating the debate (only transcribed in Croatia), or PMs listing their engagements at the start of every QT (only in the UK) were not included in this data.
Coders familiar with these countries and languages (Croatian, Dutch, English and French) were trained and tested for six weeks in recognising negativity during QTs, that is, in each speech contribution (see Online Appendix A, supplementary materials, for a detailed description of the training and Krippendorff 's alpha scores). Following the new definition of negativity (without the campaigning and the competitor dimensions), the main dependent variable was operationalised as an attack that contains a criticism of a political actor. As such, this definition encompasses: (i) attacks regardless of the moment in time at which they occur; and (ii) attacks in any type of direction (e.g. majority MP criticising the PM). This is important, as attacks also occur outside campaigns, and parties or coalition governments may be incentivised to use internal criticism, as mentioned above (see Höhmann and Sieberer 2020;Proksch and Slapin 2012).
Once the coders reached satisfactory reliability in the final two weeks of training, they proceeded to code negativity in transcripts from QTs in all three countries. Each speech contribution was coded according to whether there was negativity in it or not. In other words, the data collected indicated whether a politician had attacked someone in a speech. The final number of speech units with at least one attack was 6,643 and they accounted for 33.1% of all speech contributions (32.7% in Belgium, 36.8% in Croatia and 30.9% in the UK). 2 As such, on a descriptive level, we can see that QTs are not always about conflict, as approximately two thirds of the speech contributions were neutral. This indicates that negativity in parliaments is possibly employed strategically, with actors evaluating whether they should use it or not. See Online Appendix B for examples of negativity and neutrality in data.
Next, to gather data on the public approval of actors at different times, the study relied on external polling data sources that preceded the QTs sampled in each country. When selecting this data, the most important criterion was that these polls were publicly available and had a media presence. This ensured that the actors were exposed to them and as such that they constituted a good proxy for public approval. In Croatia, the study used CroDemoskop polling data from polls conducted by PromocijaPlus for a major Croatian private television channel (RTL), while in the UK it relied on polling data from Ipsos MORI, which is featured in the Evening Standard newspaper. All polls were conducted regularly each month (with some gaps) on a representative sample of Croatian/ British citizens, in which they are asked to indicate their preferences for parties and politicians. They were available for the majority of the time frame studied (2010-2020 for the UK and 2010-2021 for Croatia).
The situation in Belgium was less straightforward, not only because polls are not conducted regularly but also because of the country's split into Dutch-speaking (Flemish) and French-speaking (Walloon) regions, with different actors competing in each region. This is why several polling sources were used to generate approval ratings of actors on both sides and why they only cover the period between 2014 and 2020. Notably, the polls were conducted by several organisations (mostly Ipsos) for major public (VRT; RTBF) and commercial (VTM; RTL) media outlets in both regions, which asked Belgian citizens about their political preferences. To ensure that actors from all three countries were exposed to these polls, a brief investigation of their posts on Facebook revealed that the majority of them had acknowledged these polls in their social media posts at least once (see Online Appendix C).
Using the above-mentioned data, a new dataset was generated, where units of observation were parties nested in QTs per country. For example, the Flemish Christian Democratic party at a QT from December 2020 in Belgium constitutes one observation. In Belgium, in total, twelve parties that regularly engaged in QTs were included (2 Christian-Democratic, 3 Liberal, 1 Radical Right, 1 Radical Left, 1 National, 2 Socialist, and 2 Green), while in Croatia and the UK the focus was on the two dominant parties, one from the left (Social Democratic Party of Croatia; Labour Party, UK) and one from the right (Croatian Democratic Union; Conservative Party, UK). 3 In the case of a party not speaking during a particular QT in Belgium, it was dropped for that particular observation, while the two dominant parties in Croatia and the UK were active in every QT.

Negativity
This is the main dependent variable, which indicates the share of negativity employed by a particular actor during a particular QT. Therefore, if the Flemish Greens in Belgium made 10 speech contributions during a QT in May 2020 and, out of these, 5 were negative, this variable has a value of 0.50, that is 50% of all speech contributions had negativity in them. For descriptive statistics see Table 1. Public approval This is the main independent variable, which represents the share of approval that actors had in polls at the time of the observation. Given that the study observed politicians on the party level, the value of this variable represents the share of citizens that indicated they would vote for this party if elections were to be held today or tomorrow. 4 If elections were held before a QT and no polling had yet been conducted, then the share of votes that the party received at the election was used here (only in the UK).

Electoral cycle
This variable indicates how many months have passed since the last parliamentary election. Therefore, for a QT that took place in July 2015 in the UK, the value of this variable would be 2, as the previous election was held in May 2015. Therefore, the bigger the value, the closer the next election would be, allowing us to deduce the possible probability weighting (see theory) of politicians about when to go negative during the cycle. 5

Statistical method
The study used time series regressions to test the hypotheses presented in the theory section. Time series are the best method to predict causal inferences that are time dependent (influence of t − 1 on t) and are also frequently used in dynamic representation research to test how politicians respond to shifting public opinion (Beyer and Hänni 2018: 29). As such, the dependent variable (share of negativity by an actor during a specific QT) is placed at t, while the independent variable (on public approval) is placed at t − 1. To test the expectations, party approval rating at t − 1 was evaluated in interaction with the electoral cycle looking at the impact this interaction had on the negativity share at t for each country. Given the frequency of polls in Croatia and the UK, public approval was simply lagged monthly (e.g. comparing Ipsos MORI ratings from September 2016 to a QT from October 2016), while in Belgium the nearest lag possible was used (e.g. Flemish public broadcast poll published in the first week of May 2018 was compared to the nearest QT in the third week of May 2018). 6 All regressions were run with fixed-effects on the party level (i.e. controlling for differences between parties), which is particularly important because it treats changes in approval for each party separately. Several control variables were included in these models, namely whether the observed actor is in government or in opposition (not for the UK, as Conservatives were, for the most part, in office, which was controlled through fixed-effects). Furthermore, there was also a control for the main competitor's approval (not in Belgium, due to the fragmented nature of the party system). Therefore, when observing the approval ratings of the Labour Party in the UK, the approval of the Conservative Party was also accounted for. For Belgium, there was also a control for regional differences, that is, whether a party was from Wallonia or Flanders. In all three countries, a lagged dependent variable on negativity use in a prior sampled QT was used to control for a potential temporal dependence, that is the autocorrelation of negativity (which accounts for the fact that some actors may be negative in general during QTs).

Results
The results of the time series regressions are shown in Table 2. From the base models (1, 3, 5), we can observe a negative influence on the use of negativity when approval increases in Belgium, Croatia and the UK. However, none of these relationships show statistical significance, leading to the rejection of H1. While there is strong confirmation in the literature that public approval is an important indicator of negativity during campaigns, it appears that we cannot apply this logic across the full electoral cycle.
However, once the interaction between public approval and the electoral cycle is added, we start observing different results. As we can see Note: † p < 0.1 * p < 0.05; ** p < 0.01; *** p < 0.00. from these interaction models (2, 4), the public approval of actors depends significantly on the electoral cycle. As we move further away from the previous election and towards the next, politicians become responsive to public approval, exhibiting less or more negativity. Specifically, in Belgium and Croatia, an increase in public approval (t − 1) later in the electoral cycle leads to less use of negativity by politicians in a subsequent QT (t). However, in the UK, an increase in the public approval (t − 1) of parties in interaction with the electoral cycle (Model 6) remained an insignificant predictor of more or less negativity (t). In fact, there is a strong autocorrelation of negativity in the UK, with politicians employing similar shares of negativity during QTs, regardless of approval. This means that there is support for H2 in Belgium and Croatia, but not in the UK. However, the strength of the effect of approval does not increase steadily as time elapses (see Figure 1). Rather, the public approval effect in interaction with the electoral cycle only becomes significant after a certain threshold. This means that while H2 holds, the effect of approval is only significant approximately 2.5 years after the previous election was held. In fact, approval in the first years after an election had no significant impact on politicians going negative during QTs in Figure 1. Marginal effect (y-axis) of public approval in the interaction with the electoral cycle. note: intervals below 0 indicate significance in the interaction (see Brambor et al. 2006)./the x-axis represents the electoral cycle (Belgium and the uK, 5 years; croatia, 4 years).
Belgium and Croatia. As reported in Table 2, during the whole electoral cycle, approval has no impact in the UK. The following presents detailed results for each country, inspecting exact negativity use based on the realistic low and high approval values in each case. 7 In the interaction model (2), Belgian politicians behaved as expected in H2 (Figure 2; Online Appendix D.1). If a certain party in Belgium enjoys high public approval in the final months of the electoral cycle (22% at t − 1), this results in 40% negative interactions by this party's members at the final QT (t) before the election. In contrast, if a party has low public approval at the end of the cycle (8% at t − 1), it uses approximately 65% negativity (t). As such, politicians are 25 percentage points more negative if they have low approval. Thus, the higher the approval of politicians, the less likely they are to utilise negativity during QTs, but the difference between high and low approval only emerges later in the electoral cycle.
For Belgium, the difference only becomes visible at the start of the second part of the five-year electoral cycle (approximately 26-28 months since the last election) and further increases as we move towards the election. Note that for high approval, the approach of the elections does Figure 2. predicted negativity (share in all speech contributions by a party) during Qts (t) throughout the electoral cycle based on high or low approval (t − 1). note: the x-axis represents the electoral cycle (Belgium and the uK, 5 years; croatia, 4 years)./Vertical lines indicate the 90% confidence interval holding other variables at their mean. not change anything, and the share of negativity remains consistent (see the flat line in Figure 2). This allows us to conclude that, in Belgium, it is politicians whose parties have low approval that start to behave differently midway through the electoral cycle, using more negativity as elections get closer.
Croatian politicians behaved similarly to Belgian politicians and thus also confirm H2 (Figure 2; Online Appendix D.2). If a party enjoys high approval in Croatia in the final months of the electoral cycle (30% at t − 1), politicians from this party will use approximately 50% negativity during the final QT (t). This is 20 percentage points less than a party with low approval (20% at t − 1), which results in 70% negativity during the final QT (t). Like Belgium, this effect only occurs later in the cycle. In Croatia, this difference in the use of negativity becomes visible during the second half of the four-year electoral cycle (approximately 29-30 months after the last election). It should also be noted that, in Croatia, high approval has little effect on the use of negativity as the elections near, rather it is low approval that leads to more negativity as elections approach.
The slightly delayed effect in Croatia (i.e. during the second half of the cycle) compared to Belgium (i.e. at the start of the second half) may be due to the stability of the majority in parliament. With the two-block nature of the party system in Croatia, parties have greater certainty about when the next election will take place, so the impact of approval only becomes visible later in the electoral cycle. This is unlike Belgium, where a large number of parties in government produces uncertainty about cabinet stability and the duration of the electoral cycle. On average, cabinets in Belgium do not make it halfway through their maximum possible tenure (Bergmann et al. 2022). This is likely to lead to a stronger effect of approval earlier on, with politicians anticipating elections sooner, rather than later.
Finally, in the UK, we can observe that, regardless of approval, politicians go more negative as time elapses during the cycle (Figure 2; Online Appendix D.3). As such, politicians in both high and low approval scenarios at the end of the cycle (42% and 31% at t − 1 respectively) employ approximately 45% negativity at the final QT (t). The difference between high and low approval is not visible at all during the electoral cycle (see also Figure 1). There are several possible explanations for this observation. Above all, the UK has a two-party system in which conflict takes place primarily between the two main parties. Therefore, the risk associated with using negativity may be lower due to the lack of alternative parties to which voters could turn (see also Elmelund-Praestekaer 2010). As such, it appears that this context does not allow public approval to be a fundamental predictor of the use of negativity (see the conclusion for more discussion on the UK exception).

Robustness checks
In order to check the robustness of these findings, several other tests were run. First, the potential differences between government and opposition were explored (Online Appendix E). Second, models were run using the categorical variable on years since the last election (Online Appendix F). Third, the impact of proximity to the end of the parliamentary term and actual elections was tested (Online Appendix G). Fourth, the dependent variable was transformed into a count variable (i.e. absolute number of negative speeches; Online Appendix H). Fifth, individual approval impact on PMs/opposition leaders in the UK was explored (Online Appendix I). Sixth, trends in public approval were tested (was the approval of an actor rising or falling; Online Appendix J). Seventh, the impact of different elections was explored (Online Appendix K). Finally, tests were run dropping certain controls (Online Appendix L), including the year 2020 due to the COVID-19 pandemic (Online Appendix M). In general, most of these tests provided further confirmation of the results and relationships that have been presented above.

Reverse causality
Before discussing and summarising the conclusions, it is also worth exploring whether reverse causality occurs, that is, whether more or less negativity affects public approval. To do this, additional regressions were run that were identical to those reported above but they treated citizen approval as the dependent variable (t) and negativity as the independent variable (t − 1). These regressions showed that there was no impact of politicians' use of negativity in parliaments on their approval ratings in all three countries (Online Appendix N). However, a strong autocorrelation in approval ratings was identified. This means that actors who do well will likely continue to do well in the following approval polls, regardless of the negativity used during QTs. As such, while it makes sense for politicians to respond to their approval rating during QTs, it cannot be expected that citizens' behaviour follows a similar pattern (i.e. citizens watching QTs and adapting their preferences). The implications of this are discussed in the closing section of this article.

Discussion and conclusion
The main objective of this study was to explore how public approval impacts politicians' decisions to use negativity during the entire electoral cycle. Through an empirical analysis in three diverse European countries, the study confirmed its hypotheses in two countries. For Belgium and Croatia, the study found that, as time elapses during the cycle, actors become responsive to public approval through their use of negativity. They resort more to negativity later in the cycle if their approval rating is low, while actors with high approval use less negativity than others and keep their low negativity use constant. In other words, when parties in Belgium and Croatia have a low approval rating in the second part of the electoral cycle, their party members are more likely to use negativity during QTs in parliament. However, in the UK, this effect was insignificant.
The findings from the UK are interesting and deserve more attention, given that they contradict the current literature. As was proposed in the previous section, a potential cause for such different behaviour could be attributed to the two-party system (and the electoral system itself), which makes the backlash effect in the UK less prevalent compared to Belgium and Croatia. In other words, the potential gain of winning over volatile voters by going negative outweighs the potential loss of partisan voters. This is because partisan voters in two-party systems cannot abandon their party as easily as voters in a multi-party system, due to a lack of alternatives. Note that this is different in the US (also a two-party system), where 'candidate images and issue appeals have the potential to counteract partisan preferences' (Dalton 2021), as opposed to the UK, where partisan voters remain loyal. This potentially explains why, in the US, unlike the UK, scholars find an effect of approval ratings on negativity use.
At the same time, UK politics has also experienced increasing voter volatility among the general public in the last decade due to 'electoral shocks' such as Brexit (Fieldhouse et al. 2020). Moreover, approval polls in the UK now have a long-standing history of making incorrect predictions about electoral outcomes (Mellon and Prosser 2017). This was especially problematic during the closer races that took place in the studied period (see Jennings and Wlezien 2018). All of these factors may have also contributed to the lesser effect of approval ratings in the UK, where both sides try to persuade increasingly volatile voters by using negativity. Finally, the divergent findings in the UK might simply be due to the nature of PMQs, where the two main rivals argue back and forth with each other, reinforcing negativity.
Despite the differences between the UK and the other countries, this study contributed to the current literature on several levels. First, the article advanced the theoretical framework regarding politicians' use of negativity. The well-established framework of prospect theory provided a foundation for understanding the relationship between approval and negativity from a longitudinal perspective during the electoral cycle. While previous studies offer a theoretical understanding of the impact of approval ratings on negativity during mostly short-lived campaigning periods, this study revealed that there is also an impact of approval between elections, but only later in the electoral cycle. In Belgium, which has a fragmented party system with high uncertainty regarding snap elections, politicians become more responsive to approval halfway through the term. In Croatia, where majorities in parliaments tend to be stable, politicians pay particular attention to approval during the second part of the cycle.
Second, this article advanced the current literature by using a different methodological approach that explored the relationship between approval and negativity. Due to previous studies being oriented towards short-lived campaigns, they were unable to establish a directional causal link between approval and negativity. This article approached this relationship from a longitudinal perspective which made it possible to lag public approval behind the use of negativity. This allowed the assessment of whether public approval impacted the decision to use more or less negativity during subsequent QTs.
Third, this article also contributes to the public opinion literature. While previous studies showed that politicians respond to public opinion by shifting policy in a direction that the public wants (e.g. Sevenans 2021), this study identified that politicians shift communication based on signals they receive from the public. This demonstrates that politicians do pay attention to citizen approval closer to elections and that they are not only willing to change their policy based on what the public wants, but rather, in this particular case, they changed and adapted their communication strategy, probably trying to influence citizens' opinions about them. However, in line with previous studies that found weak citizen knowledge regarding politicians' engagements during QTs (Soontjens 2021), the study demonstrated that citizens' approval of politicians is not affected by negativity used during QTs.
On this final point, one may ask why politicians resort to more or less negativity based on approval if this does not significantly change the approval rating? One potential explanation could be grounded in the fact that politicians may think that more or less negativity works, despite approval not shifting significantly. In other words, low approval simply leads actors to employ more negativity in the hope that electoral gain will ultimately be achieved at the election (i.e. that it will pay off in the long run). For example, a low-approval actor who uses negativity may receive praise for this on social media, get more attention in the news (see Haselmayer et al. 2019), etc., leading them to believe that this is the way to go. In turn, high-approval actors may feel that their good standing is due to them being less negative, reinforcing risk-aversion: why change a winning strategy.
In this regard, it should be acknowledged that the main limitation of this study is that it does not explore this reverse causality puzzle in greater detail, as it was beyond the scope of this particular study. Other venues in which negativity is employed during the electoral cycle (e.g. on social media) may have a role to play, revealing different relationships and having a greater impact on citizens. Therefore, it is recommended that future research investigate other venues in which negativity is used (and during the entire electoral cycle), assessing how negativity in these venues is conditioned by approval but also whether negativity has an impact on citizens' approval.

Notes
1. For less significant findings, see Walter et al. (2014) and Hansen and Pedersen (2008). 2. While these descriptive findings are interesting and deserve attention, they are beyond the scope and goal of this article. 3. While a point can be made that third parties in Croatia should be included in this data due to the country's multi-party system, there was no reason to expect that public approval has an impact on these parties. Croatian third parties predominantly group around the two main ones for the election and post-election formation of a coalition. Therefore, it could be argued that these parties are more affected by the approval ratings of the two dominant parties. Furthermore, these parties have limited possibilities to calculate how to approach strategically QTs, given that some of them will be granted a few or no questions during QTs. 4. PVDA-PTB is the only party in Belgium that runs for office both in Flanders and Wallonia, so the share of citizens that intend to vote for this party is calculated by looking at the average of citizens that intend to vote for it in both regions. 5. Different operationalisation of the electoral cycle (where months indicate time left until the regular end of the parliamentary term or the date of the actual elections) was used as a robustness check (Online Appendix G). 6. To ensure that lags in Belgium stayed within a reasonable time frame of influence, every approval poll that was included in Belgium dated back to maximum one month before a particular QT (as in Croatia and the UK). 7. The realistic low and high approval values are based on standard deviation from the mean approval per country indicated in Table 1. For example, with the mean approval of 15% in Belgium, which has standard deviation of 7%, we can conclude that the realistic high approval in Belgium is around 22%, while the realistic low approval is around 8%.
Vargiu and Alessandro Nai. The paper also benefitted greatly from comments that came from anonymous reviewers and members of the M 2 P (Media, Movements & Politics) research group at the University of Antwerp, especially Stefaan Walgrave and Julie Sevenans. I also thank PromocijaPlus for providing a share of data used in this paper.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by the University of Antwerp Research Fund (BOF) under grant number 43834.