Extend the context! Measuring explicit and implicit populism on three different textual levels

1 Introduction

Populism shapes political communication, as it defines dichotomous ideas and attracts adaptations from political agents (Bracciale and Martella, 2017): Most left- and right-wing political actors, and everyone in between, employ at least the stylistic elements of populism to some extent (Hawkins et al., 2019). Consequently, scholars focus on the communicational dimension of the phenomenon, and analyze the various forms it takes (Gründl, 2020). In recent years, a new content analysis perspective, based on a distinction between explicit and implicit populism, has emerged in the research field of political communication. The essential idea of this refined perspective is that messages containing one articulated “us versus them” dichotomy are classified as explicit populism (EP), while messages containing either people-centric or antagonistic expressions are classified as implicit populism (IP) (Tóth, 2020). The EP-IP distinction is useful, because it can help scholars reveal hidden tensions via content analysis: If the analysis relies on both deductive typologies and inductive observations, coders and scholars might detect unseen populist features in texts.

Contemporary research highlights that the EP-IP ratio varies between different political agents: Donald Trump mostly utilizes the implicit style in his tweets; HillaryClinton aims to balance her explicit and implicit messages via Twitter and the Hungarian Fidesz-KDNP’s Facebook posts are mostly explicit (Tóth, 2020, 2021a). Even though scholars aim to provide explanations for how specific time periods (campaigns and “off-seasons”), message types (live speeches, Facebook posts, tweets), and political positions (incumbent versus challenger) might influence EP-IP ratios, none of them comprehensively deal with a specific methodological aspect that might also affect the proportion of explicit and implicit messages. As seen in former studies, we focus on the possible changes between ratios of explicit and implicit styles, but we assume that different lengths of textual coding units could transform implicit messages into explicit ones.^[1] Therefore, our research revolves around the idea that narrow (e. g., sentences) and extended (e. g., three sentences in a row and paragraphs) content analysis units might contain significantly different proportions of explicit and implicit messages. In other words, we presume that the length of the coding unit might influence the ratio of EP and IP. If our assumptions are proved by the implemented quantitative content analysis and statistical processes, scholars should consider the EP-IP approach as a method that demands multilevel coding units of text segments with fixed diverging lengths. In practice, similarly to keyword-in-context methods, if the surrounding context of the core sentence is considered, hidden tensions might be reshaped into wider dichotomies. If this assumption is validated, content analysis that considers EP and IP could not be conducted on one single coding unit with a specific length, but instead on multilevel coding units with diverging fixed extensions, because the results might change significantly on each level.

According to our knowledge, exploring hidden dichotomies by conducting a multilevel, quantitative content analysis has not been attempted in the research field of populism. This paper’s contribution to communication studies is, first and foremost, a methodological one, as we aim to test the possible impact of diverging multilevel coding units on latent implicit dichotomies. Put differently, we assume that, in many cases, the hidden or fragmented nature of populist communication transforms into an explicit style. To validate this idea, we employ core-sentence-in-context methods.

Our research focuses on the explicit and implicit communication styles of the right-wing populist Viktor Orbán,^[2] known for his anti-refugee rhetoric, criticism of the European Union, and alliances with other right-wing populists such as Giorgia Meloni, Matteo Salvini, Marine Le Pen, and Recep Tayyip Erdoğan (de la Baume, 2021; Eatwell and Goodwin, 2018). Even though Viktor Orbán has ruled Hungary since 2010 and has become the cornerstone of the political establishment, he has utilized an anti-elitist style, depicted Brussels, the United Nations, and George Soros as hostile to the will of the Hungarian people, and argued this equals defending the homeland against immigration. Additionally, Viktor Orbán unambiguously employs populism combined with nationalism in his communication (Tóth, 2020). We selected speeches Orbán gave in 2018, when the parliamentary election campaign was due in Hungary, and the proliferation of populist style became extremely intensive (Tóth et al., 2019). We operationalized a hybrid method, comprising semi-automatic and manual quantitative content analysis (Guo et al., 2020), to (1) detect latent and manifest populist messages, and (2) observe the possible transformation of implicit suggestions (people-centrism or antagonism) into direct tension (“us versus them” narrative) when analyzing larger units than core sentences.

2 Content analysis and populism

In communication studies, scholars pursue high quality content analysis by applying human coding to maximize measurement validity (Krippendorff, 2004), and computerized coding methods to take full advantage of their reliability (Aslanidis, 2018). Thus, to maximize accuracy and consistency in content analysis, scholars have developed hybrid models that aim to include both the objectivity of computerized analysis and the depth of human coding and interpretation (Su et al., 2017). Hybrid content analysis uses computational and manual strategies, since algorithms help researchers reduce and organize large-scale data, while the actual classification remains with the human coders (Baden, Kligler-Vilenchik, and Yarchi, 2020). In this case, unsupervised pattern-finder algorithms select coding units, that are then further analyzed by human coders using theory-driven categories (Baden et al. 2020). For pattern recognition, algorithms typically work with large dictionaries or, as in this case, so-called populist dictionaries, with a specific set of relevant words and expressions (Payá, 2019).

Content analysis has become an indispensable method to examine the communicative aspects of populism (Aslanidis, 2018). Therefore, scholars have analyzed many types of texts, including campaign speeches (Bonikowski and Gidron, 2016; Oliver and Rahn, 2016), press releases, conferences, party newspapers, political advertisements (Bernhard, Kriesi, and Weber, 2015), opinion articles (Rooduijn, 2014), party manifestos (Rooduijn, de Lange, and Van der Brug, 2014; Rooduijn and Pauwels, 2011), parliamentary speeches by party leaders (Vasilopoulou, Halikiopoulou, and Exadaktylos, 2014), party magazines (Pauwels, 2011), speeches by chief executives (Hawkins, 2009), excerpts (Jagers and Walgrave, 2007; Reungoat, 2010), tweets (Gonawela et al., 2018; Pal et al., 2017), and Facebook and Instagram posts (Bobba, 2019; Ernst, Engesser, Büchel, Blassnig, and Esser, 2017; Farkas and Bene, 2021; Heiss and Matthes, 2020; Stockemer, 2019).

Several approaches have attempted to characterize, compare, and understand the specific features of populist communication (Pauwels, 2017). Extant research has approached dictionary-based automatic content analysis in populism studies by (1) collecting words, phrases, and word combinations from samples, or (2) mixing deductive and inductive methodologies to create word lists (Bernhard et al., 2015; Bonikowski and Gidron, 2016; Bruter and Harrison, 2011; Oliver and Rahn, 2016; Reungoat, 2010; Rooduijn, 2014; Rooduijn et al., 2014; Rooduijn and Pauwels, 2011; Vasilopoulou et al., 2014). While this method acquires perfect reliability, it may result in low validity. In holistic gradings, trained coders evaluate degrees of populism by scaling whole texts as populist, non-populist (NP), or (alternatively) somewhat populist (Bernhard et al., 2015; Bonikowski and Gidron, 2016; Hawkins, 2009). The reliability of holistic grading could be harmed by the vast number of coders, small databases, and coders’ limited experience in reading political texts (Hawkins, 2009). Additionally, holistic grading does not provide results on the proportions of themes and issues. In thematic text analysis, researchers split samples up into units, and, after providing a typology of coding frames, coders analyze text segments (Jagers and Walgrave, 2007; Reungoat, 2010; Vasilopoulou et al., 2014). In this method, very accurate codebooks and several training sessions are needed to conduct this time- and often money-consuming content analysis approach. In clause-based semantic text analysis, semantic triplets are coding units in which the subject, the verb, and (possibly) the object appear and are used to analyze the fundamental syntactic elements of language (Aslanidis, 2018; Franzosi, 2010). Clause-based analysis suffers from a lack of rich context; thus, different coders might interpret the same units very differently.

There is no consensus among scholars on which and how many unit lengths would be ideal for exploring the nature of populist communication (Aslanidis, 2018). Individual sentences might be “minimal context units” (Krippendorff, 2004) since they provide the meaning of recording units, which are, in this case, populist-like words. However, sentences that are syntactical units (Anderson, Garrison, Archer, and Rourke, 2000), containing a universal populist-like word, might not be enough to explore latent and manifest dichotomies. Accordingly, and following Krippendorff’s (2004) suggestion, researchers may need to examine the preceding and following sentences to scrutinize the intended meaning or topic the specific words encompass. It is also possible that larger units, such as paragraphs, are sufficient to understand the meaning of the word in question (Krippendorff, 2004). Rooduijn, relying on Ji (2008) and Koen, Becker, and Young (1969), argues that paragraphs are elementary units for classical content analysis because they might highlight “thematic discontinuities” (Rooduijn, 2014), and they are not as limited for measuring a set of ideas as words or sentences are (Rooduijn, 2014). Therefore, we concluded that we should measure at the sentence level and two additional, extended context units. Our study contributes to the literature by showing that neither the narrow nor the extensive coding units should be eliminated from content analysis methods in populism studies because they might have different functions in the content. Narrow coding units such as core sentences might keep the populist style fractured, while extended contexts can upgrade and complete the meaning of the message, changing it to a direct-dichotomous style.

3 Protocolizing populism: Explicit andimplicit populism

Although prior scholarship has provided laudable methodological protocols for implementing content analysis, examining multilevel coding units is surprisingly uncommon in content analysis in populism studies. We argue that there is an inextricable need to focus on both narrow and extended coding units to discover the (hidden) presence of the two fundamental features of populism in the EP-IP approach to content analysis: The “good people” and the “culprit others.” In other words, scholars might be able to detect manifest and latent contents using content analysis (Lind, Gruber, and Boomgaarden, 2017) in populism studies too. Scholars argue that only manifest contents could be coded with scientific objectivity (Holsti, 1969), and because of their observable nature, computer-assisted methods might easily capture them (Anderson et al., 2000). In contrast, latent content is a more interesting and challenging element to investigate.

Scholars might face similar challenges in the proliferating research field of populism. However, when EP and IP emerge, scholars typically examine the manifest, “people versus elite” dichotomy, or the latent, inarticulate tension by highlighting either corruptness or relative deprivation (Tóth, 2020). There is ample evidence to suggest that populism often relies on fragmented communication: Praising the people or blaming the common enemy define populism as a fragmented ideology, implying fractured elements of dichotomies to gain votes and keep the message plain (Engesser, Ernst, Esser, and Büchel, 2017; Müller et al., 2017; Reveilhac and Morselli, 2021). This tactic makes scholarly analysis difficult. The fragmented-implicit style raises methodological problems (how to detect and measure the phenomenon), and paves the way for endless conceptual struggles: The more extensive a definition, the more challenging it is to create well-elaborated context-specific typologies of populism.

In the following, we will propose that the EP-IP approach inherently demands both hybrid and multilevel content analysis methods. Concerning the object of the study, hybrid content analysis might help differentiate between, and analyze manifest and latent variables (Neuendorf, 2017). Automatic or semi-automatic quantitative content analysis detects manifest variables, while in manual quantitative content analysis the objects are variables that cannot be directly observed (Strijbos, Martens, Prins, and Jochems, 2006). Arguably, hybrid content analysis is sufficientto provide valid and reliable results in the measurement, as much as possible. On the other hand, the analysis of multilevel units in the EP-IP approach anticipates possible deviations in results when extended contexts tend to attract the missing agent of the dichotomy. In other words, implementing hybrid content analysis helps us overcome some shortcomings of the content analysis methods introduced in the former section. While perfect reliability can be reached via automated methods, acceptable reliability and high validity are achievable through manual coding. The “explicit” and “implicit” categorizations might reveal to what extent people-centrist and antagonist themes appear in the database.

Scholars suggest (Tóth, 2020) that EP can be perceived if any direct, Manichean dichotomy appears between the homogenous masses and the “culprit others” in the analyzed coding unit (Mudde, 2004). In our research protocol, if the communicator mentions an enemy, either by emphasizing a person’s name, a specific group, or providing a universal picture of the foe (e. g., “globalist elite” and “immigrants”), and refers directly to “the people” in the same unit, EP occurs.

In the conceptualization of explicit and implicit populist styles, researchers argue (Tóth, 2020) that IP has two fundamental subcategories. The first one relies on people-centrism (Franzmann, 2016); therefore, if the people are addressed or the general will is emphasized (Laclau, 2005) but the foe remains inarticulate, the coding unit falls under the category of IP. As the literature suggests (Heiss and Matthes, 2020), people-centrism in IP is challenging to measure through quantitative methods, because people can be addressed as “we” or “us”, but these terms sometimes refer to the citizens and, in other instances, to the political agents themselves (Rooduijn and Pauwels, 2011). As our research considers the challenge above, it relies on hybrid content analysis to overcome this difficulty to some extent. The second implicit category focuses on antagonism (Hameleers, 2018): If the communicator mentions an enemy – regardless of whether it is a specific person, different elite groups (Maurer and Diehl, 2020), or minorities such as immigrants (Hughes, 2019) – that represents a common threat to the people or the general will, but does not evoke the masses, the coding unit is coded as IP (Tóth, 2020).

In line with Jagers and Walgrave’s (2007) perspective, the refinement above considers populism as a political communication style (Blassnig, Ernst, Büchel, Engesser, and Esser, 2019) that avoids complexity in communication and praises common sense politics supported by emotional appeals (Meijers and Zaslove, 2021). Besides emotionalization and simplification, the EP-IP approach connects primarily to the stylistic approach because the direct, articulated dichotomies and suggested tension might also be part of this political communication style. The IP category allows us to measure the fragmented and latent elements of populism (Tóth, 2021b): Political agents might focus explicitly on either the people or the enemy in IP messages; however, the other entity is still part of the coding unit in a concealed way. IP supports the minimal features of populism: appealing to the “good people” (Jagers and Walgrave, 2007) or criticizing the “culprit others” (Hameleers, 2018) by searching for them in messages where they seem to be missing at first glance. Considering this, IP has an “inherent incompleteness” (Taggart, 2004). EP is stricter and more rigid, and easier to perceive than IP. Therefore, identifying IP requires both deductive and inductive typologies for detecting hidden populist features by capturing people-centrism or antagonism as a more flexible refinement.

A specific instance shows the problematic nature of hidden-inarticulate tensions, direct dichotomies, and coding unit lengths in populist political communication style. Sinn Féin published the following sentence on Facebook on 22 January 2020: “Pearse Doherty says it’s time for a government for the people!” The sentence above implies people-centrism but disregards antagonism. However, the entire post appears as follows:

Pearse Doherty says it’s time for a government for the people! [Irish flag emoji] Fianna Fáil and Fine Gael have governed for almost 100 years in the interests of the elites. It’s now time for a Sinn Féin government that puts the interests of ordinary people first!

The first sentence denotes people-centrism, while in the remaining sentences the message is completed with anti-elitism, transforming the latent, inarticulate tension into an explicit dichotomy.

4 Why does multilevel analysis matter?

If one considers the chameleonic nature of the populist style, capable of operationalizing appeals and blame distinctively or by contrasting them in the samemessage, there might be no precise answer related to what the appropriate coding unit is. However, a multilevel analysis may help scholars detect IP, which has a context-specific nature, in smaller coding units, while the conceptual rigor of EP could be easier to perceive on extended levels. Taking this into consideration, we argue that choosing three levels – micro-contextual,^[3] macro-contextual,^[4] and decontextualized^[5] coding units – might be a reasonable decision in our research.

Previous content analyses, with very few exceptions (Rooduijn and Pauwels, 2011), tended to measure populism by focusing on single coding units such as words (Bonikowski and Gidron, 2016; Oliver and Rahn, 2016; Pauwels, 2011), semantic triplets (Aslanidis, 2018; Wirz et al., 2018), core sentences (Vasilopoulou et al., 2014), excerpts (Jagers and Walgrave, 2007; Reungoat, 2010), paragraphs (Rooduijn, 2014; Rooduijn et al., 2014), and whole texts (Ruth-Lovell, Doyle, and Hawkins, 2019). However, few scholars have considered using at least two different textual coding units in tandem for populism studies. Rooduijn and Pauwels (2011) analyzed paragraphs and words during their dual-concept analysis that relied on classic content analysis and dictionary-based measurement. They split the text up into paragraphs, that were then analyzed by trained coders. Along with this, and based on their dictionaries, the scholars searched for populist-like words in manifestos. Measuring populist communication on one specific coding unit is an appropriate method, but the case is different if we consider EP and IP. Even though researchers have analyzed these concepts on single units with a fixed length (Tóth, 2020), they admitted that an in-depth analysis of EP and IP supported by multilevel coding units could be a feasible endeavor.

This paper relies on Krippendorff’s argument (2004) that coding units are not axiomatic items in content analysis. In other words, the choice of coding units “depends on the analyst’s ability to see meaningful conceptual breaks … on the purposes of the chosen research project” (Krippendorff, 2004, p. 98). We aim to conduct a content analysis where “each largest unit consists of all units it contains, and each smallest unit is specified by all higher-order units of which it is a part” (Krippendorff, 2004, p. 145).

Accordingly, our research contribution aims to analyze core sentences and paragraphs, as scholars tend to choose these coding units in measuring populism as micro-contextual and decontextualized coding units (Aslanidis, 2018). The former is the smallest, while the latter is the largest unit. However, we also outline the macro level, which is between the former two units, and consists of three sentences: one before and another after the core sentence plus the core sentence itself. If the analyzed coding unit is the text’s first or last sentence at the macro level, we consider either the subsequent or the preceding two sentences. We implemented the macro-level unit because, in specific cases, a single sentence might not be enough, while a paragraph might be too extensive, as to “identify the referent of a personal pronoun, for instance, an analyst may need to examine a few sentences preceding the noun” (Krippendorff, 2004, p. 101). In Viktor Orbán’s speeches, “we,” for instance, might refer to the prime minister’s party, or to the people as well. Therefore, we implemented the macro level coding unit into our analysis. Finally, Krippendorff stresses (2004) that researchers must focus on if and how the categories applied on the first level relate to the categories utilized on the subsequent level(s). As Anderson and colleagues argue (2000), when the size of the coding unit expands, for instance, from a sentence to a paragraph, so does the likelihood that the unit will magnetize more variables. With regard to our study, the extended levels might encompass antagonism and people-centrism; consequently, IP might “change into” EP. First, to conduct our research in aiming to explore this presumption, we need to find possible connections between the styles and coding units:

RQ1: Is there an association between the micro, macro, and decontextualized units, and the explicit, implicit, and non-populist communication styles in Viktor Orbán’s speeches?

Moreover, we aimed to scrutinize whether the co-occurrences of the three coding units and styles appear with significant frequencies. Additionally, we analyzed which specific levels and styles might have strong correlations.

RQ2: Which intersections of context unit types and populist styles are statistically significant in the multilevel analysis of Viktor Orbán’s speeches?

We assumed that core sentences imply either people-centrism or antagonism rather than direct dichotomies; thus, implicit style might dominate the smallest unit. Besides this, we hypothesized that extended context units (macro and decontextualized levels) might contain the complementary feature of the populist style. Therefore, the style might “change” from IP to EP. From a methodological perspective, analyzing EP and IP in multiple contexts is vital because the more extensive the analyzed unit, the higher the chance that EP occurs if a communicator intends to utilize dichotomies. Therefore, our first hypothesis is:

H1: Macro context will contain a significantly higher portion of EP than micro context in Viktor Orbán’s speeches.

As communicators may be inclined to follow the logic of the fragmented ideology (Engesser et al., 2017) in their populist styles, they might employ IP rather than EP in the micro context and possibly on other units. In contrast, former research (Tóth, 2020) suggested that Viktor Orbán’s public speeches contained slightly more EP than IP when four speeches made by the Hungarian PM were analyzed. However, we assumed that a robust database, which implies 18 speeches in our research, might supply the dominance of implicit-fragmented messages over explicitness at the micro level.

H2: Implicit populism has a statistically significant, higher portion than explicit populism in micro-context units in Viktor Orbán’s speeches.

5 Materials and methods

Speeches from politicians are popular research samples in content analysis (Bonikowski and Gidron, 2016). However, types can still vary. As the Global Populism Database and The New Populism Project at the Guardian suggest (Hawkins et al., 2019), measuring populism might be appropriate not only in campaign speeches, but ribbon-cutting speeches, international speeches, and famous speeches too. Therefore, we selected all of Viktor Orbán’s campaign speeches from 2018 for analysis in this study. Moreover, we added two speeches that were given a few days before the official campaign period, and another speech held in Serbia. The database consists of 9,177 specific terms and 34,465 words. The first speech was held on February 16, 2018, while the last one was held on April 8, 2018. All speeches were given during the campaign stage, except the Speech at the Christian Democrat International’s Conference and the Year Assessment Speech, which still fit the “important speech” category, and occurred close to the beginning of the campaign. We decided to analyze the period above because anti-immigrant and anti-elitist styles had never been as influential during the parliamentary elections in Hungary as they were in Viktor Orbán’s campaign speeches in 2018 (Tóth, 2020). For appropriate comparison of EP and IP on comprehensive data, we included the four speeches investigated in former research (Tóth, 2020).

Following Krippendorff’s (2004) instructions, each context unit type was coded separately, and we kept “a complete but redundant record of all variables by all smallest recording units identified in the body of texts, similar to the data structure” (Krippendorff, 2004, p. 146). First, we focused on the semi-automatic, quantitative approach of our analysis by operationalizing the populist dictionary validated by prior research (Tóth, 2020). Following scholars’ recommendations related to hybrid content analysis (Baden et al., 2020; Pauwels, 2017; Rooduijn and Pauwels, 2011), we operationalized populist-like words from the dictionary to reduce the large sample size. Next, trained coders coded at the micro, macro, and decontextualized levels. Now, we introduce our analysis step by step. First, to provide results with perfect reliability for our hybrid content analysis, we separated the dictionary into two sets of words following the “us versus them” logic of the populist political communication style, to help us find explicit dichotomies using the automated method. These categories are “friend” (n = 60) and “foe” (n = 26), based on the Manichean “good versus evil” approach (Mudde and Rovira Kaltwasser, 2013), which characterizes the populist style. In our data, we aimed to include words in the dictionary for three different coding units: (1) core sentences, (2) three sentences in a row, and (3) paragraphs. We created the three different coding levels and searched dictionary words using Maxqda 2022. The program broke the text into micro, macro, and decontextualized units, and searched words from the dictionary on every level separately. Then, the software transformed the words to either the codes of “friend” or “foe.” Note that the two codes might have appeared in the same coding unit. At this point, we utilized the so-called code relations browser, a function that crawls the intersections of codes on a specific level to show how many context units any two codes are attached to. The program listed the results on each level based on the co-occurrences of “friends” and “foes.” If both categories appeared within the same context unit, the program listed it as co-occurrence or, in other words, as EP.

Afterward, to prepare for the manual quantitative coding process, we searched every coding unit in which a single “friend” or “foe” code could appear, and we re-ran the dictionary-based analysis on the same three levels. In this part of the analysis, we did not only focus on co-occurrences; we also considered results in which either the enemy or the people appeared alone. We followed this protocol because examining only co-occurrences might skew the results. In other words, co-occurrences relying on deductive dictionaries cannot detect every EP message because the context might complete the validity of the “friend” or “foe” categories by exploring unique words and phrases. The same logic is applicable in IP units: Deductive dictionaries are imperfect in detecting extra, specific, vague, and actual elements that a communicator uses in a specific situation, like Donald Trump’s operationalization of both the words “crooked” and “she” to refer to Hillary Clinton in his tweets (Enli, 2017).

The next step was manual coding. To label false positive hits in the database, we provided three codes for the trained coders: 0 = NP, 1 = EP, and 2 = IP. When the manual coding was finished, we calculated Cohen’s Kappa (Freelon, 2013). Finally, we conducted a cross-table Chi-square test of independence to analyze the possible ties between styles and context units.

6 Results

Our semi-automatic analysis shows that “friends” codes were present in a larger share (4.01 %, n = 1,382,) of the sample than “foes” codes (0.35 %, n = 121). The shares of the former code are larger in every individual speech than those of the latter.^[6] Five speeches did not contain any references to antagonist entities, while “friends” appeared in every speech.

The code relations browser showed the diverging portion of code co-occurrences within the micro (n = 113), macro (n = 439), and the decontextualized levels (n = 45). To prepare for manual coding, the program listed 1,503–1,503–257 individual and co-occurred hits based on the dictionary, including duplications for the micro, macro, and decontextualized levels. After cleaning the hits for any redundant inclusions, coders analyzed the micro (n = 923), macro (n = 904), and decontextualized (n = 209) units.^[7] The results of manual coding suggest that coders’ decisions were sufficiently reliable (see Table 1).

Table 2 shows that IP style dominates every level; however, its shares decrease at the second and third levels. NP has a slightly higher frequency than EP does at the micro level, but its share decreases at the macro level, and almost disappears at the decontextualized level. EP has the smallest share at the micro level, but on the next level, it more than doubles its portion, pushing NP back to the third place, and keeps increasing on the last level.

A Chi-square test of independence was conducted to explore the association between the context unit types and populist styles. All expected cell frequencies were greater than five. There was a statistically significant association between context and populist style, χ²(4) = 174.30, p < .001 (RQ1). The association was moderately strong (Cohen, 1988), Cramer’s V = .207. To follow up on this significant result, we implemented a cell-by-cell comparison or analysis of residuals. Cells with large absolute adjusted standardized residuals (regardless of the sign: positive or negative) indicate where a lack of independence occurs. Extant research has suggested that residuals greater than either 2 or 3 indicate where a cell deviates significantly from independence (Agresti, 2007). Taking 3 as the baseline, our results indicate that only the association between the macro level and IP, along with the decontextualized level and IP, were not greater than 3. In the rest of the cell-by-cell comparison, there is strong evidence to consider them significant (RQ2).

The two most significant adjusted standardized residuals were micro context and EP, and micro context and NP (see Table 3). For the micro-level analysis using the EP style, 40.89 % less explicit populist style was used compared to what would be expected if the null hypothesis was true, with an adjusted standardized residual of –10.8. In contrast, compared to the preceding coding unit, macro context attracted a significant increase in EP because its standardized residual reached 7.9. Consequently, these results support H1. For the micro-level analysis, using NP style, there was an almost 67 %-increase in NP style compared to expectation, with an adjusted standardized residual of 9.7. Finally, the standardized residual for IP reached 3.8 at the micro level, while, as mentioned above, EP attracted a strong negative correlation at the same level; thus, H2 was supported.

7 Discussion and conclusion

Our results show that the latent version of populism, namely IP, dominates every level in Viktor Orbán’s speeches, as its shares do not fall under 50 % in either of the three contexts. Consequently, our findings affirm researchers’ claims that the nature of the populist style is fragmented (Engesser et al., 2017; Müller et al., 2017). In other words, people-centrism appears most frequently in Viktor Orbán’s speeches on each level, fracturing the “us versus them” dichotomy into one of the features above. This outcome contrasts with former findings claiming that the explicit style represents a larger share of Viktor Orbán’s speeches than the implicit one (Tóth, 2020). An explanation for this difference might be that the database used in previous research consisted of only four, perhaps EP-heavy, speeches (Tóth, 2020), while our database utilizes 18 speeches, including speeches given on occasions when the Hungarian PM had business meetings and other ceremonies where less explicit dichotomies needed to be stressed.

Our quantitative findings suggest that people-centrism might be more intensive than antagonism in Viktor Orbán’s speeches. Possibly, the Hungarian PM intends to strengthen cohesion among his supporters (Anderson, 2006) by operationalizing populism mixed with nationalism and nativism for the most part (Korstenbroek, 2021; Waterbury, 2020). A subtype of IP, namely people-centrism, is the dominant feature of Viktor Orbán’s speeches in the analyzed period; the Hungarian prime minister intentionally addresses the people and emphasizes their will to confirm that he is the true political leader who cares for the ordinary citizens. In line with the populist logic, Viktor Orbán suggests through an overwhelmingly people-centrist implicit style that he adjusts his policies to the serve the will of the Hungarian people. The people-centrist style of the Hungarian PM fuels implicit populism, as he unambiguously claims that he listens to the Hungarian people’s voice, while also suggesting that the oppositional, foreign, and globalist elite work not only against him, but against the entire nation.

Even though our findings support that “friends” emerge over eleven times more frequently than “foes”, occurrences of EP increase with statistical significance at the macro and decontextualized levels. A possible explanation for this outcome is that the macro and decontextualized units magnetize many words referring to “friends”, and only a few connecting to “foes.” In this light, emphasizing the core elements of populism by overrepresenting either “the people” or the dangerous “others” might not exclude an increase in direct dichotomies on extended context units. In other words, the people-centrist implicit style in Viktor Orbán’s speeches, especially at the micro level, still attracts antagonism at the macro or decontextualized levels. We demonstrate the transformation of the implicit populist style into an explicit one through the following coding unit:

In other words, my dear friends, we want to win not just an election, but our future. Europe and we, Hungarians, have reached a turning point in world history. National and globalist forces have never fought each other in such an open way.

The quote above is from a speech held at a large public gathering that was organized to celebrate the anniversary of the 1848–49 Hungarian revolution on March 15, 2018. The Hungarian prime minister addresses both the Hungarian and European people in the first two sentences: Viktor Orbán suggests that a new era begins in the history of the Hungarian and other European nations. In the last sentence of the macro-contextual unit, the nationalist Viktor Orbán interjects the hostile globalist forces into his speech, transforming the implicit style into an explicit one. ThePM possibly aims to remind people that recent economic “achievements” and “development” are under risk because foreign interests are against the will of the people and their well-deserved prosperity. The upgrading function of the macro context is aligned with the populist logic: The people have a better life and welfare, but the corrupt, self-interested elite aims to compromise that through financial speculation and by supporting immigration, cheap labor, and “drastic” changes in the national culture and religion in favor of immigrants. The extended, dichotomous messages might foster anxiety about both job and welfare loss, which happened in Hungary after the collapse of state socialism at the beginning of the 1990s. To avoid a similar deprivation, Viktor Orbán assures the people, through dominant people-centrism, that his number-one priority is nothing but fulfilling the general will. Even though the “general will” is a very broad concept, Viktor Orbán understands it in this context as sustaining (relative) welfare, avoiding high unemployment, and preserving security to protect recent economic achievements. Viktor Orbán consciously evokes an enemy that threatens the populist vision, and legitimizes the conservative, right-wing PM as the voice and protector of the unheard masses.

Interestingly, besides our assumption based on transformations in several implicit messages, there are remarkable shifts towards IP and EP within NP messages. In contrast to the increase in explicit style, the proportion of NP messages decreases at the macro and the decontextualized levels, suggesting that (1) the style might drastically change in broader contexts, and, (2) as Strijbos and colleagues (2006) highlight, overlapping coding units might modify coders’ decisions in content analysis. In Viktor Orbán’s speeches, therefore, extended coding units might not maintain NP contexts but transform them by implementing the populist style’s features. Consequently, the macro context provides the greatest increase in explicit populism, and the most significant decline in the NP style. Therefore, populist-like words, embedded into NP micro contexts, might upgrade NP styles to explicit ones and attract latent tensions in the macro and decontextualized units. There are possible explanations for the decline of NP on the macro and decontextualized levels. First, core sentences might not be extensive enough to express the unit of meaning that represents arguments, discussion topics, and ideas (Chi, 1997) when adjusted to the populist style. Examples could be an emphasis on the people’s demands, relative deprivation, popular will, anti-elitism, sovereignty, and the harmful role of immigrants (Hameleers, 2018). However, the macro context might implement either or both “friends” and “foes.” Second, events such as ribbon-cutting ceremonies, where Viktor Orbán could avoid the populist style, might still magnetize “friends” and, occasionally, “foes” too, thus NP coding units might change to implicit ones. Finally, Viktor Orbán stresses the government’s achievements, which might be labelled as NP style at the micro level. However, he also emphasizes in the extended coding units that the vital achievement, which provides prosperity for the people, vanishes if the corrupt opposition wins the elections and obeys “Brussels.” The Hungarian PM highlights that achievements must be “protected”, but often avoids mentioning the source of threat at the micro level, which might create latent populist dichotomies not in the core sentences, but in the extended units.

Our contribution to the research field is the following: Multilevel content analysis is preferred if one aims to utilize the EP-IP approach. Based on the results above, the aforementioned method demands the utilization of multilevel coding units in content analysis if scholars aim to detect direct and hidden populist dichotomies. We argue that researchers should not use a single fixed-length text unit when applying the EP-IP approach to perform content analysis. At least two text units, a narrow and an extended one, are needed, as significant differences between the proportions of latent and manifest populist dichotomies can be observed when considering both the level of core sentences and a larger unit in the same analysis. Moreover, our second contribution is that EP could be captured with a significantly higher chance in the extended units, especially in the macro context, where thematic discontinuities start decreasing. Therefore, if researchers aim to examine populist dichotomies in live speeches, they should analyze at least three subsequent sentences as extended coding units, because the largest shifts towards the explicit style were observed at the macro level.

It is important to note that there might be specific instances where the operationalization of multilevel coding units might become insufficient. For example, suppose political agents or their communication team deliberately decide to use a specific populist style, for instance, implementing dichotomies in every sentence, especially in shorter messages such as tweets. Other strategies might also appear that could discredit the multilevel analysis: A mainstream politician who implements the populist style to gain extra votes might only focus on people-centrism, and disregard antagonism in an entire speech. In such cases, choosing multiple coding units is not a reasonable decision because the ratio of explicit and implicit styles will not change either on the narrow or the extended units.

Moreover, we propose that explicit and implicit styles do not occur only in populism, but in several (thin) ideologies such as racism, nationalism, nativism, ethnicism, xenophobia, islamophobia, homophobia, and so on (Aslanidis, 2016). Consequently, hybrid content analysis combined with the EP-IP approach might allow scholars to detect latent and manifest tensions within (political) agents’ communication when they aim to conduct content analysis in other research fields.

Finally, even though this paper examines the right-wing Viktor Orbán’s speeches as segments of a case study, the presented and tested method is suitable for analyzing left-wing populists or mainstream, (neo)liberal politicians’ styles as well. As we noted above, former research already investigated whether, and to what extent, EP and IP appear in Hillary Clinton’s tweets from the 2016 presidential elections. It found that explicit- and implicit-style messages prevailed over NP messages (Tóth, 2021). Additionally, and interestingly, coders have perceived in our research that Viktor Orbán also utilizes some antagonistic features of the left-wingpopulist style when he blames some harmful businessmen, banks, and austerity policies (e. g., extra taxes, sustaining overhead reduction). Therefore, some features of the left-wing populist blame-game were also tested in this research. Consequently, the EP-IP approach is suitable for comprehensive analysis of populist styles.

Our study has specific limitations. First, refined dictionaries using more synonyms, other authors’ word sets, and considering as many validated inductive phrases as possible, might extend the research horizon. Second, considering other coding unit levels, for instance, semantic triplets, whole texts, meanings, utterances, and propositions, might also offer insights into the proportions of the aforementioned styles, and possible statistical deviations. Third, our study analyzed Viktor Orbán’s speeches; therefore, scholars should research other political agents, including parties and activists, considering implicit and explicit styles for the sake of comparative aspects. Fourth, as this research is language-specific, only Hungarian messages were scrutinized; thus, other languages and regions should be investigated. Fifth, other types of messages, for example, Facebook posts and manifestos, might provide opportunities for extensive comparisons to support the analysis of the nature of diverging populist styles. Finally, we admit that manual analysis of EPand IP could be time-consuming. However, the recently emerging crowdcoding content analysis methods help scholars overcome this difficulty (Lind et al., 2017).

Our last remark suggests a possible opportunity to research the explicit and implicit populist styles from another angle. We assume that online experiments might be feasible methods to measure whether, and to what extent, the use of NP, implicit, and explicit styles fuels fear, anger, enthusiasm, engagement, or other attitudes in respondents, in order to acquire knowledge not only on communication strategies on the supply side, but also on their effects on the targeted audiences.