Democratic erosion, negativity and the role of media

Larry M. Bartels (Reference Bartels2023) contends that political elites and not the citizens are the driving force behind democratic erosion: Leaders like Orban and Erdoğan come to power through democratic elections, and their increasingly authoritarian governments are electorally sustained because the citizens are mostly satisfied with their economic and social well-being, and not because they demand authoritarianism. Nonetheless, aspiring autocrats can take advantage of polarisation among the citizens. Aytuğ Şaşmaz et al. (Reference Şaşmaz, Yagci and Ziblatt2022) provide evidence that affective polarisation as well as economic anxiety can cause popular support for authoritarian institutions. Indeed, across countries, affective polarisation is correlated with the magnitude of democratic backsliding (Orhan Reference Orhan2022). As Milan Svolik (Reference Svolik2019: 23) explains, ‘their countries’ acute society-wide political conflicts raise the stakes in elections and, in turn, the price their supporters have to pay for putting democratic principles above partisan interests’. This is why backsliding leaders do not merely benefit from prior polarisation; they may actively try to further polarise their polities (Druckman et al. Reference Druckman, Peterson and Slothuus2013; Somer et al. Reference Somer, McCoy and Luke2021).

However, ways in which politicians actively polarise their polities and accentuate partisan identities are less well known. A well-established field of relevant research concerns negative campaigning. John G. Geer (Reference Geer2006: 23) defines negativity as ‘any criticism levelled by one candidate against another during a campaign’. In a meta-analysis of empirical studies, Richard R. Lau et al. (Reference Lau, Sigelman and Rovner2007) report that evidence is mixed as to whether negative campaigns are more effective in generating negative evaluations of their targets or in changing the vote decision. Maybe this is because not all negativity is received in the same way by the public. James N. Druckman et al. (Reference Druckman, Gubitz, Lloyd and Levendusky2019) separate civil from uncivil kinds of negativity, the latter consisting of slurs, threats of harm and disrespect. They demonstrate that incivility can backfire: Uncivil attacks against the opposite party coming from partisan media affiliated with one’s own side lead the individual to be less favourable towards his own party. The finding is driven by effects on individuals with conflict-avoidance (rather than conflict-seeking) traits. Studying the multiparty setting of EU polities, Philipp Mendoza et al. (Reference Mendoza, Nai and Bos2024) find that such backfiring results in withdrawal of support when there are ideologically close alternatives available to the voter.

Despite its mixed effectiveness in campaigns, negativity persists. Theoretical models predict that weaker candidates will be incentivised to run more negative ads, but candidates who are attacked will retaliate in the same way too. As a result, the use of negativity will spread in the course of a campaign, especially if the race is close (Nai Reference Nai2020; Skaperdas and Grofman Reference Skaperdas and Grofman1995). In any case, negative political messaging can increase affective polarisation (Lau et al. Reference Lau2017).

Much of the empirical evidence regarding the use and effectiveness of negativity presumes an information environment where political competitors are free to attack each other with negative messages. However, the more lopsided media landscape typical of more authoritarian polities provides a different setting. Studies of electoral authoritarianism emphasise the role of media in influencing the public opinion and underline the importance of its independence from political interests for vibrant democratic competition (e.g. Bermeo Reference Bermeo2016; Levitsky and Way Reference Levitsky and Way2002; Levitsky and Ziblatt Reference Levitsky and Ziblatt2018; Waldner and Lust Reference Waldner and Lust2018). When autocrats have control over mass media, they can achieve dominance ‘by distorting information flows’ on a constant basis, and if successful, ‘many citizens do not realize that they are being dominated’ (Guriev and Treisman Reference Guriev and Treisman2019: 123). These distortions can take up different forms. Arturas Rozenas and Denis Stukal (Reference Rozenas and Stukal2019) demonstrate that state media in Russia responds to adverse economic developments not so much by censoring or distorting the facts, but by systematically coupling bad news with mentions of external actors or other bad news happening in other countries, thus trivialising the negative data or diverting responsibility.

For scholars of electoral politics, the media is important not only as a site for documenting the political discourse in circulation but also as an actor that filters, frames and amplifies the discourse selectively. Focusing on the coverage of the same person (Donald Trump) in 12 Western democracies, Sjifra De Leeuw et al. (Reference De Leeuw, Azrout, Rekker and Van Spanje2020) find that pejoration is more common in countries with a polarised pluralist media system. For the USA, Matthew Gentzkow and Jesse M. Shapiro (Reference Gentzkow and Shapiro2010) find that audiences have a preference for like-minded news and media firms respond strongly to consumer preferences, which accounts for most of the explained variation in news coverage slant. By contrast, the identity of a newspaper’s owner explains far less of the variation.

However, in less competitive settings, it is typical for media bosses with wide-ranging commercial interests to use their media outlets to curry favour for access to government contracts. Alper H. Yagci and Cem Oyvat (Reference Yagci and Oyvat2020) observe that TV channels in Turkey owned by businessmen connected to the Erdoğan family devote very little time to opposition leaders during a campaign period. Focusing on Turkey’s 2011 election, Ali Çarkoğlu et al. (Reference Çarkoğlu, Baruh and Yıldırım2014) find that as the election approaches, pro-incumbent newspapers favour the incumbent and disfavour the opposition, and vice versa in opposition papers. Following a similar research design, Kerem Yıldırım et al. (Reference Yıldırım, Baruh and Çarkoğlu2021) look at four subsequent election periods in Turkey to examine the evolution of newspaper content. They find that negative mentions of opposition parties increase at each subsequent election, while the incumbent Erdoğan’s party establishes dominance in terms of both visibility and favourability.

These studies of media content focus on narrow time periods such as election periods and code news tone at the level of news stories. In this study, we direct our coding to the sentence level to detect more subtle changes in language, since a particular kind of language may creep at the sentence level into text that is coded for another category at higher levels like news stories. We use machine-learning methods to access and code every single sentence published online about the main opposition party in two newspapers over a period of about 13 years. Our methods allow us to go beyond the study of campaign periods, discover long-term trends and test how news content responds to real-time shocks. In addition, we offer a more nuanced differentiation of negative media content relating to the opposition, which is explained in the next section.

Theory and hypotheses

Turkey is an electoral regime that has been dominated by the Justice and Development Party (AKP, after the Turkish initials) since the party came to power under a parliamentary system in 2002. Party leader Tayyip Erdoğan first served as the prime minister and, gradually increasing his power, engineered a government system change to become the country’s first popularly elected, highly empowered president. Throughout, as an electoral strategy, Erdoğan exploited the social cleavage between religious conservatives, for whom his party bid to speak, and the seculars, who disproportionately vote for the main opposition CHP. This polarising strategy has so far paid off despite ups and downs in economic performance, generating increasingly authoritarian incumbency. V-Dem data indicate that during this time, Turkey has become one of the ‘worst offenders’ and a pioneer of the now-worldwide trend in terms of democratic backsliding (Nord et al. Reference Nord2025: 18–19). Thus, we are studying a crucial case.

We examine how media depictions of the main opposition party evolved during Turkey’s gradual authoritarianisation under Erdoğan’s leadership. At the centre of our interest is a difference between two kinds of negative description, ‘rival’ and ‘villain’, with different implications for the place of opposition in an electoral regime.

A political rival is quite like the self, just worse – ostensibly. Political actors claim to be better than their rivals and in doing so they recognise a common ground where two sides compete in delivering similar things for the society. The rival type of rhetoric does not problematise the legitimacy of the political opposition and sees it as a valid opponent within the political arena. A villain, on the other hand, is fundamentally different from the virtuous self – in being immoral or having illegitimate goals. The villain type of rhetoric may come with important consequences for democratic well-being. Parties, if seen as illegitimate by large segments of the population, can lead to distrust and dissatisfaction in democratic institutions. In addition, claiming that the opposition is immoral contributes to the strengthening of us-versus-them camps within the society, harming democracy through lowering interpersonal trust and enabling justifications of authoritarian action against the other camp.

The conceptual line between the rival and villain types is inspired by Carl Schmitt’s (Reference Schmitt2005) discussion of the political importance of friend–enemy distinctions, as well as by the emergency politics literature that builds upon Schmitt’s thought (Agamben Reference Agamben2005; Lührmann and Rooney Reference Lührmann and Rooney2021). It also sits well with the recent emphasis on the multimodal style of rule by hybrid regimes. As Ugur‐cinar (Reference Ugur‐cinar2023: 451) comments:

Deploying rival and villain types of rhetoric at different times may serve these varying ends.

For operationalisation purposes, we conceive the concepts of villain and rival as consisting of subcomponents, in turn. Disaggregating them in this manner will enable observers to more reliably recognise when they see them. Accordingly, the villain category can be considered to apply to descriptions of Turkey’s main opposition party CHP as ‘illegitimate’ or ‘immoral’. Illegitimate designations include allegations or insinuations to the effect that the opposition (1) is against democracy and supportive of military coups; (2) cooperates with terrorist organisations or their supporters; and (3) obstructs national interests or the national will. Immoral designations include allegations that the main opposition (4) is lacking in qualities such as integrity, virtue or honesty; (5) does not respect the general moral or religious values popularly held in the society; and (6) is involved in corruption and abuse of power.

These two subcomponents together make up the villain category. Next come two other subcomponents that make up the rival category: These are descriptions that paint the opposition as ‘incompetent’ or ‘incohesive’. Incompetent refers to designations of the main opposition party as being incapable of creating positive results. This includes allegations relating to (1) defeats, failures, negligence or lack of merit; (2) mismanaging the economy (in the past), the municipalities or the public institutions; (3) unpopular policies by CHP administrations such as lay-offs and protests against them; and (4) the party or its leader’s positional volatility, indecisiveness and ambiguity. Lastly, incohesive designations refer to internal turmoil that gives a negative signal about the party’s capacity to govern and its leader’s effectiveness. This includes (5) instances of factions, internal conflict and struggles for leadership; (6) resignations, expulsions from the party and internal disciplinary actions; and (7) allegations of the party’s unfair treatment of its members.

The relationship of the various categories discussed is visualised in Figure 1. The figure also presents exemplary sentences for each category.

Figure 1.

Villain, Rival and Other Depictions of the Opposition

Based on this conceptual framework, we hypothesise that the growth of authoritarianism will be reflected as a change over time in three ways in the media content about the opposition. Where

V denotes the daily number of sentences about the opposition coded as a ‘villain’;

R denotes the daily number of sentences about the opposition coded as a ‘rival’;

N denotes all ‘negativity’, and equals V + R;

P denotes the daily number of other sentences about the opposition, which may be neutral or positive, and therefore enable the opposition to enjoy a ‘platform’;

T denotes the ‘total’ (and not only opposition-related) online newspaper content size;Footnote ¹

we predict the following hypotheses:

Potentially, the three phenomena can be observed simultaneously in data from a given newspaper. But they represent distinct types of disadvantage for the opposition. ‘Negativity’ is compatible with business-as-usual campaigning, and it is an endemic, if not commendable, part of more democratic electoral politics, too. ‘Vilification’, on the other hand, is a vicious form of negative discourse in which the opposition is disproportionately described as an enemy of the people and the state, which therefore warrants to be crushed, rather than a regular electoral rival. ‘Deplatforming’ represents a subtle form of disadvantage, where the opposition becomes invisible in the political discourse. While negativity and vilification could be useful in consolidating pro-incumbent leaners, they could be ineffective or even counterproductive on that part of the electorate with more centrist inclinations. To reduce political mobilisation among the latter, deplatforming could be a useful method, contributing to a widespread feeling that there is no convincing, feasible electoral alternative to the incumbents.

Why do we choose these particular measures over simple sentence counts or alternative ratios? Consider the following scenario: During intense political episodes, i.e. around scandals or election times, there may be an increase in villain designations relating to the opposition, as a count or as the ratio of overall opposition-related discourse. However, rival designations may be increasing even faster. Our first measure, which is the count of villain designations as a ratio of all negative content about the opposition, is designed to distinguish vilification intensity from such scenarios of rivalry-driven increases in negativity. Similarly, during intense political episodes, we may observe an increase in the count of negative designations relating to the opposition simply because politics becomes more salient in the news cycle. However, neutral or positive designations may be increasing even faster. Our second measure, which is the count of negative designations as a ratio of all content about the opposition, corrects for this possibility while measuring negativity. Lastly, neutral or positive designations of the opposition may remain unchanged as a sentence count or as a ratio of all content relating to the opposition. But this may be happening while the opposition is increasingly marginalised as an item in the overall discourse. Our last measure, which is the count of other (neutral or positive) opposition-related sentences as a ratio of total newspaper content size, corrects for this possibility.

A key assumption behind our three main hypotheses is that the gradual deepening of authoritarianism in Turkey under Erdoğan can be proxied by the simple passage of time, to be modelled as a linear time trend. We take this to be a reasonable assumption in the absence of a daily varying measure of authoritarianism that should not rely on media indicators that it is supposed to predict. In addition, our data will allow us to test a number of more specific hypotheses.

One hypothesis concerns variation across media outlets, and it will partly serve as a validation check for our measures. Existing empirical studies attest that, among the two newspapers we examine, Sabah is more pro-incumbent, whereas Hürriyet is rather mainstream, in terms of both their editorial direction and survey-based analysis of the political preferences of their respective audiences (Çarkoğlu and Yıldırım Reference Çarkoğlu and Yıldırım2024, see Figure A7). Hypothesis 4 will exploit this established finding to explore varying patterns across the newspapers. If our measures are measuring what we theorise, they should be observed more intensely in the pro-incumbent Sabah rather than the centrist Hürriyet. Hence,

We should also note that Hürriyet, while traditionally having a more centrist audience, changed hands in May 2018 to become more aligned with a pro-government position (evidenced in greater detail in the Supplementary Material). Hence, we also expect to see changes in Hürriyet after May 2018. Since this date roughly corresponds to a change in the country’s government system in addition to the newspaper’s ownership, we do not list this expectation as a hypothesis to be tested, but we will comment on the available evidence.

Another hypothesis concerns the election periods. We argue that political actors have mixed incentives about the extent and type of negativity they want to direct towards each other during election periods, compared to the rest of the timeline. A straightforward expectation could be the intensification of negativity around elections. Extant literature is replete with findings pointing to increases in negativity over the course of election campaigns (for a review, see Nai Reference Nai2020); however, studies of campaign periods only cannot test how campaign periods differ from the rest of an incumbent’s term. And the opposite expectation is also theoretically plausible. In light of the previously noted backfiring mechanism (Druckman et al. Reference Druckman, Gubitz, Lloyd and Levendusky2019) and mixed effectiveness of negativity (Lau et al. Reference Lau, Sigelman and Rovner2007), a nastier style would not necessarily pay off during a period of heightened citizen interest in political news. It should also be noted that in relatively authoritarian environments, ongoing, perennial attacks against the opposition are already an established part of the game. These attacks signal and, by intimidating potential opponents, contribute to an environment of uneven competition. In such a context, election periods need not be relatively nastier. To the contrary, politicians may want to utilise election periods as an opportunity to sound more moderate to the voters so as to form a ‘big tent’ appeal, and signal to international observers that serious, competitive elections are being held. After all, leaders who have been designated as populists are known to adapt their discourses due to voter expectations (Taraktaş et al. Reference Taraktaş, Esen and Uskudarli2024). In sum, election periods could be indistinguishable or exceptional in either direction, and we will test the following hypothesis.

Next, we consider how these political discourse indicators respond to two sources of political distress: economic shocks and security challenges. To politically cope with such events, the incumbent leadership may turn to using more negative language about the opposition. This may come in the form of defensive comparisons where the incumbents claim that government performance would have been worse under hypothetical governments ruled by (current) opposition parties, hence featuring a rival discourse. Negative language may also come in the form of attacks on the opposition, linking them to active efforts to undermine national security or economic stability. The incumbents may employ such vilifying discourse to divert attention away from government failures, and even take advantage of the situation by scapegoating the political opposition as the real culprit for the country’s problems, to justify authoritarian measures and restrict opposition space. It is expected that security challenges, more than economic shocks, would be associated with vilification of the opposition, though one could plausibly expect either an increase or a decrease in the vilification tone of negative content during times of economic shock. We will test the following hypotheses:

Data collection and coding

We will examine how vilification, negativity and deplatforming of the main opposition party CHP have proceeded in media discourse in Turkey during a time of gradual authoritarianisation. Our data population is the entire digital news content of the two newspapers from September 2009 to August 2022, dates imposed by data access constraints. This excludes opinion columns and only includes news reporting of what has happened, as well as what has been said by other political actors in relation to the main opposition party. We code at the sentence level and include every single sentence directly about the main opposition party published (online) in two newspapers over a period of about 13 years.

We employ a relatively labour-intensive procedure of supervised machine learning, where a context-specific coding scheme based on a priori theoretical constructs was validated in an iterative fashion and applied by a sizeable number of human coders to prepare training data, which was then used to generate automated machine coding. Unsupervised machine-learning techniques that utilise little human input have the advantage in terms of speed and cost, especially if they rely on off-the-shelf dictionaries or pretrained word embeddings. However, they may have limitations such as insensitivity to word ordering or context, and a general difficulty to validate the results (Grimmer and Stewart Reference Grimmer and Stewart2013; but also see Rodriguez and Spirling Reference Rodriguez and Spirling2022 for recent developments).

Our data collection and coding process is explained in greater detail in the Supplementary Material and summarised here. Data collection started with a keyword filtering of all online news content from two newspapers to obtain all sentences relating to the main opposition party CHP and its leaders. Then came tokenisation into sentences, utilising the VNLP Library for Python, developed to handle natural language processing steps specifically for Turkish (Turker et al. Reference Turker, Ari and Han2024). The resulting corpus to be examined consists of 184,369 sentences from the Sabah newspaper and 153,651 from Hürriyet.

The preparation of a codebook was the next step. Accordingly, each sentence could be coded positive for one of five mutually exclusive binary categories: illegitimate, immoral, incompetent, incohesive and null (i.e. none of the above). As explained earlier, the illegitimate and immoral categories count as villain rhetoric. The incompetent and incohesive categories count as rival rhetoric. Null sentences featuring none of these are considered as featuring neutral or positive descriptions of the main opposition party.

Once prepared by the researchers and tested, the codebook was given to a team of 14 human coders, excluding the researchers, to code a training dataset. While coding for training data, inter-coder reliability was ensured through an iterated process that began with tests that lasted until a Krippendorff’s alpha coefficient (and a very close Cohen’s alpha) of 0.69 was achieved across coders working on common test data.Footnote ³ These are chance-corrected statistics reporting a measure of how much agreement is achieved above what could be expected by coincidence and relative to a maximum of full agreement (Geisler and Swarts Reference Geisler and Swarts2019: 161).

We thus obtained manually coded training data, which we used to train the BERT sequence classification algorithm (Schweter Reference Schweter2020) for Turkish. The algorithm utilises part of the training data to make inferences to test against the other part. The coding found in the training data was consistent enough to allow the algorithm to achieve an algorithm evaluation success score (F ₁ measure) of 0.70, calculated as the harmonic mean of the proportion of actual positives identified correctly (i.e. recall) and the proportion of positive predictions actually correct (i.e. precision, see Powers Reference Powers2020: 2). The entire corpus of 184,369 sentences from Sabah newspaper and 153,651 for Hürriyet were then machine-coded based on the trained algorithm, generating the data we utilise for the analyses in this study.

Our methods do not allow us to distinguish sentences or sentence parts that are direct quotes from indirect reporting of discourse and events. A look at our raw data would reveal that quotes dominate much of the news content, and most of the quotes seem to come from incumbent party figures like Erdoğan himself. Nonetheless, which quotes to cover and how long would require an editorial decision by the newspaper management too. In this regard, the data we work with are not a direct measure of the incumbent discourse towards the opposition but a measure of what the audience can hear through two major media outlets. Our assumption is that as the incumbent language towards the opposition becomes more negative, this would be reflected in media coverage of the opposition featuring sentences with more negative descriptions of the opposition party; therefore, the latter can be taken as an indicator of the effective incumbent discourse. In addition, the editorial differences between newspapers will be reflected as a variation in sentences across the two newspapers, as well as across different observation periods for the same newspaper that went through ownership and editorial change.

Once coding the data at the sentence level, we sum the number of sentences corresponding to different categories of discourse for each day and then obtain daily outcomes of interest, such as the ratio of negative sentences to all sentences about the main opposition (negativity), or the ratio of vilifying sentences to all negative sentences about the opposition (vilification) for a given day. All of these measures are expressed in percentage terms. We thus acquire a time series data consisting of daily observations. Some days have missing values due to online data access problems or zero denominators of ratios. To replace them and have an uninterrupted time series, we interpolate from the trend. Averaging over our different outcome measures, less than 3% of all values are interpolated. To replicate without interpolation, we will also run tests on data collapsed to the weekly level for robustness.

Analysis and findings

We begin our analysis by presenting Figure 2, a stacked area graph for the distribution of each coded category over time, collapsed at the weekly level for ease of visualisation. The figure suggests that early in the period under analysis, the two newspapers start out with comparable levels of opposition-related content, as a ratio of their overall content. However, the pro-incumbent Sabah gradually gets ahead in this regard, driven by negative and increasingly vilifying content. The military coup attempt of 16 July 2016 seems to mark a critical juncture, leading to a period of higher vilifying content for this newspaper. The coup attempt failed to overthrow the incumbents and led to a State of Emergency that lasted until the transition to a presidential system of government in June 2018. There is a marked decrease in the overall opposition coverage of Hürriyet starting with the second half of 2018, which roughly corresponds with this transition, as well as a change of ownership for this newspaper to be more in line with incumbent interests.

Figure 2.

Data Aggregated at the Weekly Level

Figure 2 Long description

Data Aggregated at Weekly Level Weekly Average Sentence Counts. Two stacked area graphs. Graph 1: Hurriyet. Horizontal axis label: Date. Range: 2016 to 2022. Vertical axis label: Hurriyet. Range: 0 to 200. Legend categories: Other, Rival, Villain. The stacked area shows frequent spikes in the early part of the timeline, with several peaks reaching close to 200. After the early period, the stacked area stays mostly near the baseline with smaller spikes, generally below about 50, with occasional higher spikes. Graph 2: Sabah. Horizontal axis label: Date. Range: 2016 to 2022. Vertical axis label: Sabah. Range: 0 to 200. Legend categories: Other, Rival, Villain. The stacked area shows repeated spikes across the timeline, with multiple peaks rising well above 100 and some peaks approaching 200. Spikes appear more frequent and taller in the later part of the timeline compared with the earlier part. Comparison. Hurriyet shows the largest spikes concentrated earlier in the timeline and lower levels later. Sabah shows more frequent high spikes across the timeline, including later periods, with peaks that reach close to 200.

t-tests reveal that over the entire period, Sabah has a higher mean vilification ratio (62.37 ± 0.38) than Hürriyet (59.51 ± 0.44) (t(9,492) = −4.88, p = 0.000). Sabah also has a much higher negativity ratio (57.23 ± 0.36) than Hürriyet (32.88 ± 0.30) (t(9,494) = −51.42, p = 0.000). In other words, a much greater part of the opposition-related content is indeed negative in Sabah compared to Hürriyet, and a greater part of this already higher negativity consists of villain rather than rival designations.

We continue with diagnostic plots focusing on how key outcomes behave around election times. Since we suspect outcomes to be displaying long-term trends, for this periodic analysis, we first detrend the outcomes by subtracting the 1-year moving average (centred on the observation day) from each day’s raw value. We then define a time period extending to 120 days before and after each election, giving 241-day periods, and average the detrended values across all such election periods, obtaining one representative 241-day period. Within this window, if we focus on a campaign period defined as the 60 day-period before each election, t-tests would reveal that the representative 60-day campaign period displays somewhat higher detrended ratios of vilification (0.19 ± 1.37) than the remaining days for Sabah (−2.24 ± 0.80) (t(239) = 1.52, p = 0.064) as well as for Hürriyet (0.18 ± 1.32 campaign vs −2.47 ± 0.92 for the rest, t(239) = 1.49, p = 0.068). The representative 60-day campaign period also displays higher detrended ratios of negativity (7.16 ± 0.76) than the remaining days of the studied representative period (−1.75 ± 0.54); significantly (t(239) = 8.65, p = 0.000) for Sabah, and perhaps also for Hürriyet (0.76 ± 0.64 campaign vs −0.55 ± 0.49 for the rest, t(239) = 1.39, p = 0.08). Lastly, the neutral or positive sentences as a detrended ratio of total newspaper political content display no significant change between campaign periods (3.62 ± 1.07) and the rest (1.53 ± 0.92) in Sabah (t(239) = 1.22, p = 0.111), while in Hürriyet it is significantly higher in campaign periods (7.75 ± 1.37) compared to the rest (0.26 ± 0.67) (t(239) = 5.32, p = 0.000). For ease of comprehension, the patterns are also visualised in Figure 3. Displayed lines of fit are derived from interrupted time series analysis regression, where treatments are assumed to be taking place 60 days before each election as well as on election days themselves, and the lagged version of the outcome is included as a predictor.

Figure 3.

Key Outcomes around Election Periods

Figure 3 Long description

The image consists of four scatter plots analyzing trends in vilification and negativity around election periods for Sabah and Hurriyet newspapers. The first plot titled ′Vilification in Sabah′ has the horizontal axis labeled ′days to election′ ranging from negative 120 to positive 120 and the vertical axis labeled ′detrended vilification ratio′ ranging from negative 5 to positive 5. The plot shows a general upward trend in vilification as the election approaches, with data points scattered throughout the range. There are clusters of points around the 0 to 60 days before the election, indicating increased vilification during this period. The second plot titled ′Vilification in Hurriyet′ follows the same axis labeling as the first plot. It also shows an upward trend in vilification as the election approaches, with a noticeable cluster of points around the 0 to 60 days before the election. The third plot titled ′Negativity in Sabah′ has the same axis labeling as the previous plots. It shows a slight upward trend in negativity as the election approaches, with data points scattered throughout the range. There are clusters of points around the 0 to 60 days before the election, indicating increased negativity during this period. The fourth plot titled ′Negativity in Hurriyet′ follows the same axis labeling as the previous plots. It shows a similar upward trend in negativity as the election approaches, with clusters of points around the 0 to 60 days before the election. Overall, these plots illustrate that both vilification and negativity tend to increase as elections approach, with notable clustering of data points indicating heightened activity during the campaign period.

From this analysis focusing on election times, one could conclude that vilification progressively intensifies before elections, and so does negativity. However, this would be true only in comparison to the immediately surrounding time periods, which may be misleading, especially considering that the immediate post-election periods come with a lull in the intensity of political competition, maybe due to a honeymoon effect and simple fatigue.

For a comprehensive view that would allow us to test our hypotheses, we will run regression models where data from the entire time period are included, and the daily outcomes of interest are estimated separately for each newspaper. For our daily time series data, we will use generalised least-squares approach to estimate the parameters in a linear regression model where the errors are assumed to follow a first-order autoregressive process:

\begin{equation*}{y_t} = \beta {X_t} + {u_t}\end{equation*}

where x may refer to any number of (exogenous) independent variables, and where the errors satisfy

\begin{equation*}{u_t} = \rho {u_{t - 1}} + {e_t}\end{equation*}

and the e_t are independently and identically distributed as N(0, σ²). To deal with serial correlation in the error term u, we employ the Cochrane–Orcutt transformation. Durbin–Watson tests show that serial correlation becomes insignificant with this transformation (Table A1 in the Supplementary Material). For robustness, we will later employ other AR(1) models. We use Huber–White standard errors to deal with heteroscedasticity all along.

Among the predictor variables of our regressions, we include a linear time trend in the form of a day counter. Remember that inclusion of a time trend is equivalent to detrending both the left-hand and right-hand variables entering into a regression in terms of the coefficient estimates and, asymptotically, the standard errors. We are also interested in the coefficient for the time trend itself as a proxy indicator for growing authoritarianism over time. Dickey–Fuller tests show that no unit root is present in the data series once the trend is accounted for (Table A2 in the Supplementary Material).

Other variables in the baseline regression include a dummy variable indicating the 60-day period before elections. We also include a dummy variable indicating the 30-day period after elections, since diagnostic plots suggested that immediate post-election periods may be exceptional. We add a ‘coup attempt’ dummy indicating that the observation belongs to the period starting with 15 July 2016. ‘System change’ is a dummy variable indicating that the observation belongs to the period starting with 24 June 2018, when a constitutional referendum resulted in a switch from a transitory semi-presidential to a hyper-presidential government system. Lastly, we add dummy variables standing for each day of the week to account for a weekly cycle that may be present in the news content.

Results displayed in Table 1 show that over time, in both newspapers, significantly, negativity as a percentage ratio of the content about the main opposition as well as vilification as a percentage ratio of the negative content are increasing. The opposition platform (neutral or positive content about the opposition) becomes an increasingly small part of the overall newspaper content, significantly for Sabah. The trends of change are stronger for Sabah, which was the newspaper with pro-government ownership all along the period under investigation. Data collapsed to the weekly level – without interpolation – produce similar results (Table A9 in the Supplementary Material).Footnote ⁴

Table 1.

Baseline Model

Table 1 Long description

Regression results on daily data estimate how key events relate to three outcomes for Hurriyet and Sabah: vilification, negativity, and platform share. The daily trend is positive for vilification and negativity in both outlets, while platform share trends slightly down for Hurriyet and down for Sabah. In the 60 days before the election, vilification drops strongly in both papers, negativity rises sharply in Sabah but shows little change in Hurriyet, and platform share increases in both. In the 30 days after the election, vilification falls even more in both outlets, while negativity increases in both, and platform share rises in both. During the coup attempt period, Sabah shows higher vilification and higher negativity, while Hurriyet shows lower negativity; platform share falls notably for Sabah and changes little for Hurriyet. Under the system change period, vilification declines in both outlets, negativity shows little change, and platform share decreases for Hurriyet but increases for Sabah. Model fit varies by outcome, with the strongest fit for Sabah negativity, and the transformed autocorrelation diagnostics are close to the expected value, suggesting the correction was effective.

Notes: Time series regression on daily data with Cochrane–Orcutt transformation. Durbin–Watson (DW) values approximate the (insignificant) 2 after transformation. Heteroscedasticity-robust standard errors are provided in parentheses.

* p < 0.1; **p < 0.05; *** p < 0.01.

One interesting result concerns the chief qualitative difference between the two media outlets: The military coup attempt of 2016, which was followed by the declaration of a State of Emergency, is associated with a significant reinforcement (here estimated as a one-time jump) of the long-term trends in data coming from Sabah: Negativity and vilification further increase, and the opposition platform becomes smaller still. These reinforcements are somewhat compensated by a shift (relative to the increasing daily trend) in the other direction with the adoption of the presidential system in June 2018, which replaced the State of Emergency with normality under an autocratic but stable new government system. This is in line with V-Dem’s assessment that autocratisation in Turkey progressed rapidly until 2018, when it finally stabilised at a high level (see Nord et al. Reference Nord2025: 19, footnote 3). The pattern is different for Hürriyet: There is no intensification with the coup attempt – if anything, negativity significantly declines. And the shift to the presidential system, which roughly corresponds for this newspaper to the timing of a shift of ownership by a government-connected business group, is associated not with more negativity or vilification but with a 7-percentage point shrinkage in the opposition platform as a ratio of the overall newspaper size (while newspaper size itself shrinks, see Table A10 in the Supplementary Material). We interpret this as being in line with an editorial strategy that combines partisan goals and market realities: For the conventionally centrist audience of this newspaper, an increasingly negative or vilifying tone towards the opposition could backfire, but the growing invisibility of the opposition could have the politically intended effect in favour of the incumbents.

Another interesting result concerns the election times: The 60 days before each election, on average, display 5–6 percentage points less vilification (as a ratio of negativity) in the two newspapers, compared to a regular day. For Sabah, this happens alongside a 10-percentage point increase in negativity (as a ratio of opposition-related content), meaning that negativity during this period is driven by rival rather than villain designations. The opposition also enjoys more of a positive or neutral platform. For Hürriyet, there is a 13-percentage point increase in the opposition platform measure, compared to 8 points in Sabah. The decrease in vilification is more dramatic during the 30 days after each election. To check the robustness of this finding, we replicated the analyses by including local and national elections as two separate variable sets (see Table A4 in the Supplementary Material). It turns out that, for all outcomes except one (negativity in Hürriyet during campaigns), general and local elections behave similarly in terms of coefficient signs, so below we continue to pool them together. All in all, within the context of Turkey’s competitive but increasingly authoritarian electoral regime, election periods appear as intense but relatively gentlemanly political episodes as far as the language captured in media content is concerned: Around elections, opposition-related content circulates more, and this is driven by positive and neutral content in addition to negative content characterised by rival rhetoric, while vilification rhetoric takes a back seat, compared to non-election season.

Next, we analyse how these media discourse indicators respond to two sources of political distress – economic shocks and security challenges – to test for H6 and H7. To provide a measure of adverse economic change, we rely on currency depreciation. Compared to measures like inflation or unemployment, which are reported monthly, or GDP growth, which is published on a quarterly basis, the currency exchange rate has the advantage of having a frequent counting interval – it is available for every business day. In the case of Turkey, it is a particularly suitable measure because the country runs chronic current account deficits that require foreign financial inflows, while allowing for a floating exchange rate that varies on a daily basis in response to market demand and supply for foreign currencies; consequently, currency stability is widely considered a chief measure of the economy’s health. We focus on the US dollar (USD) daily opening exchange rate against the Turkish lira (TRY). We impute missing values for weekends and bank holidays by interpolating from the trend. Our measure is the percentage change in the USD exchange rate (theoretically ranging from −100 to 100) over a 30-day period that ends with the day before the day of observation. Greater values represent greater depreciation of TRY, hence an adverse economic development.

To provide a measure of adverse security developments, we rely on the number of casualties for Turkish security forces. Although Turkey has not been in a large-scale war during the period under investigation, the Turkish military has been fighting separatist Kurdish insurgents since the 1980s. Military casualties are therefore a regular part of the Turkish news cycle, and high numbers of casualties in a short period of time usually generate a national conversation about the government’s security policy performance. Our data for the daily count of Turkish military casualties come from Arzu Kibris (Reference Kibris2020) for the period that lasts through 2019; for the remaining period, we scrape the data from online announcements on the General Staff of the Turkish Armed Forces website. Our measure is the sum of military casualties over a 3-day period that ends with the day before the day of observation. Greater values represent adverse change in the security environment.

The full results of a multivariate estimation adding currency depreciation and military casualties to the regressions presented earlier can be found in Table A3 in the Supplementary Material. We display the coefficient estimates for the two new variables in Figure 4 and report the other relevant findings here: For all previously discussed variables, the direction and strength of the coefficient estimates remain essentially unchanged, and the results would be similar if we separated local and general elections (see Table A5 in the Supplementary Material). The coefficient for military casualties is significantly positive for vilification: every count addition to the 3-day running tally of the Turkish military casualties before the day of observation is associated with a significant increase in the daily vilification ratio of the negative content about the opposition by about 0.5 percentage points in Hürriyet and by a slightly higher amount in Sabah. This is matched by a smaller decrease (significant for Sabah) in negative content as a ratio of all opposition-related content; hence, rival rhetoric must be decreasing substantially following military casualties. Opposition platform (positive or neutral content relative to all newspaper size) also shrinks – again, significant for Sabah only.

Figure 4.

Multivariate Coefficient Estimates for Currency Depreciation and Military Casualties

Figure 4 Long description

The image contains three coefficient plots for vilification, negativity and platform. Each plot shows the effects of currency depreciation and military casualties on Hurriyet and Sabah. The x-axis represents coefficient estimates ranging from negative 0.5 to positive 0.5, with a vertical line at zero. The y-axis lists predictors: currency depreciation and military casualties. In the vilification plot, Hurriyet shows a positive effect for currency depreciation around 0.2, while Sabah shows a smaller positive effect. Military casualties have a positive effect for both, with Hurriyet around 0.3 and Sabah slightly lower. In the negativity plot, currency depreciation shows a negative effect for Hurriyet near negative 0.1 and a similar effect for Sabah. Military casualties have a negative effect for both, with Hurriyet around negative 0.2 and Sabah slightly more negative. In the platform plot, currency depreciation shows a positive effect for Hurriyet near 0.1, while Sabah shows a negative effect. Military casualties have a negative effect for both, with Hurriyet around negative 0.4 and Sabah slightly less negative. The plots use different colors to distinguish between Hurriyet and Sabah, with horizontal lines indicating 90 percent and 95 percent confidence intervals. The vertical zero line helps assess whether effects are statistically significant.

Note: The figure shows 90% and 95% confidence intervals.

The coefficient for currency depreciation behaves in a way that could be considered surprising. For a given day, every 1-percentage-point increase in the USD exchange rate against TRY over the preceding 30-day period is associated with a significant decrease in daily vilification by about 0.2 percentage points in both newspapers (remember that a linear time trend remains as a control). In other words, on days preceded by greater depreciation of the Turkish currency, rival designations of the opposition overshadow villain designations. In Sabah, negative (villain + rival) content itself decreases as a ratio of all opposition-related content, and other (positive or neutral) opposition-related content increases as a ratio of both newspapers’ total size (while total size itself is increasing, see Table A10 in the Supplementary Material). The point is that opposition vilification seems to be negatively associated with (detrended) currency depreciation over the previous period. This may be because currency depreciation and related economic problems lead to a national conversation about management competence, which would call for rivalry-related discourse rather than vilification. Focusing on the economic crisis episode in 2022, Serkant Adiguzel (Reference Adiguzel2025) demonstrates that news about the economy increases in Sabah as economic hardships worsen in Turkey. Economic shocks also come with increased scrutiny from international financial institutions, investors and ratings agencies, necessitating a more restrained government discourse to maintain economic credibility.

In short, the results suggest that currency shocks put the incumbents on the discursive defensive in the sense that rival designations of the opposition dominate the content found in our data during the week following such a shock. Turkish military casualties, on the other hand, are followed by deplatforming of the main opposition, in the form of less coverage about them, and by an increase in the vilification ratio of the negative content about them. When more Turkish soldiers are killed, the opposition is vilified more in the national media.

Robustness checks

As a robustness check against our estimation techniques, which so far relied on the Cochrane–Orcutt procedure for correction of serial correlation in regression error terms, we replicated our analysis with a first-order autoregressive AR(1) model as applied by Stata, where maximum likelihood estimates are obtained via the Kalman filter. The previously reported findings, including those about election periods, currency depreciation and military casualties, remain essentially unchanged (see Tables A7 and A8 in the Supplementary Material).

As another check, we compare alternative measures for our economic and security variables. It is reasonable to expect political discourse to respond to these exogenous variables, but there is no a priori choice for what the time window could look like. So we inductively explore different options, presenting the results in Table A6 in the Supplementary Material and summarising here. For currency depreciation, if we replace the 30-day measurement with a 15-day period, the results are essentially unchanged. With a 7-day period instead, signs are preserved, but coefficients become weaker and lose significance for most of the dependent variables. In consideration of the possibility that political actors may react differently to marginal (even if detrended) changes in currency depreciation versus currency shocks of extraordinary magnitude, we also check for a binary measure. It turns out that in Sabah, vilification would be significantly reduced by about 3.5 percentage points after currency shocks if we define a shock as a binary variable indicating depreciation over the previous week that is 1 standard deviation above the mean. The reduction is 6.5 percentage points with a definition based on 2 standard deviations. Data from Hürriyet record similarly sizeable coefficient estimates but fail to obtain significance this time, since relatively few days record extraordinary currency shocks. In short, results about currency depreciation are consistent, but they may disappear as the time before observation shortens, except for extraordinary shocks. This is expected: Turkey has a floating exchange rate regime, so depreciation happens gradually. The finding about military casualties is also robust in terms of its general pattern. Replacing the 3-day count of military casualties with a 7-day window essentially reproduces the original findings. If we take 15- and 30-day periods, patterns are similarly significant, but most coefficients start to get smaller.

As one last check, we apply the analysis to a different data population. Remember that our original data were compiled by first picking opposition-related articles, then retaining only the sentences that pass a keyword filter. This involves a choice about whether we want to cast a conservative or wide net. Now we go for the latter, and dropping the sentence-level keyword, we analyse all sentences in all relevant articles. Applying our economic and security model to this larger data of more than 2 million sentences essentially reproduces the original results (see Table A11 in the Supplementary Material). Since our algorithm was not trained on this data, the precision of the estimates should be met with greater caution, but it bodes well that the patterns remain the same.

Before finishing, we should recognise the limitations of the present study. Our time frame begins in 2009 due to data access restrictions; however, the incumbent leadership in Turkey has been in power since 2002. The earlier period was known to be more liberal, and it would have been interesting to see how the elements of the dual discourse evolved during that time. Similarly, extending this study to other contexts like Hungary could test whether some of our specific findings, such as the relative civility of election periods, have parallels beyond the case of Turkey. We believe that our conceptual frame is original enough to make such extensions worthwhile.