Previous Research

A growing body of research demonstrates that political parties frequently employ appeals to social groups in their campaign communication across countries, time periods, and communication channels. For instance, Thau (Reference Thau2018) analyses appeals to economic and non-economic groups by the Danish Social Democratic Party (1961–2004), while Thau (Reference Thau2019) examines the frequency, range, and concentration of group appeals by the British Labour and Conservative parties (1964–2015) to their traditional constituencies such as workers and businesses and newer, unaligned groups like parents and families. Horn et al. (Reference Horn, Kevins, Jensen and van Kersbergen2021) study welfare-related appeals in Denmark, Norway, and Sweden (2009–15), highlighting demographic and occupational groups. Huber (Reference Huber2022) investigates political parties’ appeals in Austria (2013–19) across various categories, including behaviour-based groups such as car drivers and smokers. Stuckelberger and Tresch (Reference Stuckelberger and Tresch2024) examine parties in Germany, Switzerland, and the Netherlands (2009–17), focusing on economic and identity-based groups, while Dolinsky (Reference Dolinsky2023) compares appeals by Israeli and Dutch parties to ethnic, religious, occupational, and demographic groups (1977–2015). Work by Howe et al. (Reference Howe, Szöcsik and Zuber2022) documents appeals by parties made to civic, class, national, religious, and professional groups in 1907 Imperial Austria.

Although these studies provide valuable insights, they remain limited in scope due to reliance on manual coding. Moreover, as different conceptual and measurement approaches were used across these studies and most of the resulting data are not publicly available, it is impossible to integrate them into a coherent dataset.

Recent work by Riethmüller and Franzmann (Reference Riethmüller and Franzmann2026) and Licht and Sczepanski (Reference Licht and Sczepanski2025) relies on automated methods (dictionary-based and supervised learning, respectively) to capture social groups in political texts, representing a significant step forward. Both, however, focus only on identifying mentions of groups and do not capture a party’s evaluative stance towards the group. The PSoGA Database addresses these issues by providing the first large-scale, cross-national data based on a coherent concept and measurement approach that captures not only which groups parties mention but also positive or negative stance, whether they include any policy commitments, and how they fit into meaningful group categories.

Constructing the PSoGA Database

Construction and Content

The PSoGA Database covers 791 election manifestos from 139 parties across Austria, Denmark, Germany, Ireland, the Netherlands, Spain, Switzerland, and the United Kingdom between 1970 and 2025.Footnote ⁴ The unit of analysis is a natural sentence, which provides a linguistically coherent and replicable unit across languages. Using quasi-sentences risks discarding information because the segmentation process focuses on policy topics, splitting sentences so that the social group is disconnected from the associated policy and stance. In the PSoGA Database, each group mentioned within a sentence is recorded separately with their related stance, policy linkage, and group category.Footnote ⁵ In practice, manifestos also contain other text formats such as bullet-point lists; these are treated as standalone sentences,Footnote ⁶ and our models perform well regardless of the specific text format.Footnote ⁷

To construct the Database, we follow the process described in Figure 1. Manifestos were obtained from the MARPOR corpus (Lehmann et al. Reference Lehmann, Franzmann, Al-Gaddooa, Burst, Ivanusch, Lewandowski, Regel, Riethmüller and Zehnter2025a) and the PoliDoc Archive (Benoit et al. Reference Benoit, Bräuninger and Debus2009) for all parties that gained a seat in the lower chamber of the national parliament between 1970 and 2025.Footnote ⁸ All manifestos were parsed into natural sentences and cross-checked against the original documents when available (step (1) in Figure 1). Next, we use four multilingual computational models, fine-tuned on hand-coded English and German materials by the authors, achieving Krippendorf’s Alpha scores of greater than 0.8 for all labels. After additional validation in Dutch, Spanish, and Danish, we analysed the corpus in two steps.Footnote ⁹ First, an mDeBERTaV3 (He et al. Reference He, Gao and Chen2023)Footnote ¹⁰ Bilingual Token Classifier identified all social group mentions (step (2) in Figure 1).Footnote ¹¹ Second, all sentences containing at least one group were passed through three models detecting (1) stance, (2) policy connection, and (3) meaningful group category (step (3) in Figure 1).

Figure 1.

Text analysis process.

Note: the figure illustrates the three steps of the text analysis process. After document preparation (1) and identification of social group mentions (2), three models are applied to each sentence containing a social group reference: stance (positive/negative/neutral), policy connection (policy/no policy), and assignment to one or more meaningful group categories.

For stance and policy detection, we fine-tuned two multilingual natural language inference (NLI) models that combine BERTFootnote ¹² models’ well-known high performance on classification tasks with the ability to respond to natural language queries posed as true or false statements like ‘The text {is positive/contains a policy directed} towards {group}’ (Burnham Reference Burnham2025; Laurer et al. Reference Laurer, van Atteveldt, Casas and Welbers2024). The stance detection model directly considers whether each mention of a given social group is positive, negative, or neutral.Footnote ¹³ Following advances in text-as-data approaches, we focus on stance towards a specific target rather than sentiment (Bestvater and Monroe Reference Bestvater and Monroe2023), allowing us to differentiate within a single sentence between an appeal with a positive stance towards one group and an appeal with a negative stance towards another. For example, in this sentence, ‘The Conservative Party’s policy of tax cuts to large enterprises has favoured the rich at the expense of the struggling working class’, the stance NLI model identifies a negative stance towards the rich and a positive stance towards the working class. The policy NLI model similarly determines if a policy related to each of the mentioned groups is present in the sentence.Footnote ¹⁴

Finally, to address the variety of linguistic expressions used for social groups and facilitate substantive analyses, we use an mDeBERTaFootnote ¹⁵ group-category classifier to allocate the social group tokens into forty-three meaningful group categories.Footnote ¹⁶ The model accounts for intersectionality and compound tokens (for example, ‘retired workers’, ‘low- and middle-income families’, ‘young farmers’, ‘homeless veterans’) as well as single tokens like ‘girls’ that can be categorised both into ‘women’ and ‘children’ by allocating them into multiple relevant categories.Footnote ¹⁷

The final output yields two datasets: one on the sentence level and one on the manifesto level.Footnote ¹⁸ The sentence-level dataset provides the full text of each manifesto. Natural sentences with more than one social group appear repeatedly (once per social group), resulting in ∼770,000 observations derived from over 700,000 natural sentences. For each observation, the dataset provides the sentence text, the social group string (if detected), the stance towards the social group (positive/negative/neutral), whether a policy related to the social group was mentioned (policy/no policy), and the meaningful social group category into which the group string was allocated. In addition, the dataset provides the country, election date, party, party family, total number of natural sentences in the full manifesto, and the number of natural sentences in which no social group was detected. The full codebook is in Appendix B3.

Across all manifestos, 304,087 social groups are mentioned in 39 per cent of sentences. Parties mostly express positive stances towards these social groups (71.5 per cent), with negative (8.4 per cent) and neutral (20.2 per cent) stances in only a small number of cases. This may seem surprising, but manifestos are primarily focused on who parties want to appeal to – which is, by nature, a positive act – rather than who parties oppose. About one-third (33 per cent) of group mentions were accompanied by a related policy, and the vast majority of social group mentions (94 per cent) belong to one meaningful group category. Of these, 8.3 per cent are assigned to the residual category ‘Other’, capturing social group tokens that do not fall into any specified category.

The manifesto-level dataset provides the aggregate percentage of group appeals devoted to each meaningful social group category within each manifesto. Groups mentioned with a neutral stance are not considered group appeals and therefore excluded. The dataset contains the following content analytic information for each manifesto: (1) Percentage of appeals to a given meaningful social group category with a positive/negative stance, calculated as the share of the overall number of group appeals. (2) Percentage of appeals with a positive/negative stance that also included a related policy statement. The dataset also includes information on the country, election date, party, party family, total number of sentences analysed, total number of sentences without mentioned social groups, and total number of allocated codes. The full codebook is provided in Appendix B3.

Figure 2 displays the ten most frequently appealed-to groups with either a positive or negative stance, providing a glimpse of the dataset’s content.

Figure 2.

Top ten social group categories receiving most appeals by stance.

Note: the figure shows the ten most frequently appealed-to social group categories across all manifestos in the database, separated by stance.

Validation

Unlike manually compiled datasets used in previous studies, the PSoGA Database is constructed using automated methods, enhancing scalability and ensuring consistency. To assess data quality, we check for face validity, cross-reference key variables with comparable measures from existing datasets, and benchmark our results against prior studies.

We begin by assessing the face validity of our descriptive results by examining keyness statistics for group appeals by party family in Figure 3. Specifically, keyness statistics (chi2) measure whether particular social groups appear significantly more often (overrepresentation) or less frequently (underrepresentation) in manifestos of given party families than would be expected if group mentions were distributed randomly across all party families. The baseline for comparison is the expected frequency of each group category under the assumption of independence between party family and group category. Positive keyness values (light blue) indicate overrepresentation, while negative keyness values (dark blue) indicate underrepresentation. The results align closely with the ideological profiles of different party families and their connections to specific social groups. Radical Left parties disproportionately mention LGBTQI, women, and employees and workers in their appeals, while Radical Right parties emphasise migrants and refugees, and criminals. Christian Democratic and Conservative parties overemphasise traditional constituencies like families and religious communities, whereas Ecological parties highlight women, migrants and refugees, and the LGBTQI community. Agrarian parties emphasise farmers, as expected, while Special Issue or Ethnic parties stand out for their focus on ethnic and national communities. These patterns reinforce the plausibility of the group coding and show that our method captures expected variation in party rhetoric across ideological lines.

Figure 3.

Keyness statistics by party family.

Note: Figure 3 displays keyness statistics (chi2) for the most distinctive social groups mentioned by nine party families. The bars indicate the strength of association between each group and party-family rhetoric, with positive associations shown in light blue and negative associations in dark blue.

To further assess the validity of the PSoGA Database, we contrast our results with the MARPOR dataset (Lehmann et al. Reference Lehmann, Franzmann, Al-Gaddooa, Burst, Ivanusch, Regel, Riethmüller, Volkens, Weßels and Zehnter2025b), which is one of the few publicly available datasets that permits systematic comparison. MARPOR was originally developed to measure issue emphasis and party positions. In addition, the MARPOR dataset includes six group-related categories: per701 (Labour Groups: Positive), per702 (Labour Groups: Negative), per703 (Agriculture and Farmers: Positive), per704 (Middle Class and Professional Groups), per705 (Underprivileged Minority Groups), and per706 (Non-economic Demographic Groups). Yet these are coded only when statements cannot be assigned to policy categories. As a result, MARPOR is expected to substantially understate both the frequency and the breadth of group appeals.

However, as these are the only categories available, we use them to evaluate convergent validity by calculating Pearson correlations between the PSoGA and MARPOR measures for the overlapping group categories in Table 1. The results show strong correspondence for farmers (r = 0.654, p < 0.001) and moderate correspondence for labour groups (r = 0.365, p < 0.001). Demographic groups (r = 0.296, p < 0.001) and minority groups (r = 0.156, p < 0.001) also show positive correlations. By contrast, appeals to the middle class only show a weak and statistically significant relationship (r = 0.087, p < 0.05), indicating conceptual differences in how the two schemes operationalise this category. Since MARPOR’s category of farmers is the closest conceptual match to our group categorisation scheme, the strong positive correlation illustrates that both datasets yield similar patterns when categories are conceptually equivalent. Graphical comparisons for all groups are reported in Appendix C.

Table 1.

Correlation between PSoGA and MARPOR group measures

Note: Pearson correlation coefficients between Parties’ Social Group Appeals (PSoGA) and MARPOR coding schemes applied to the same corpus of party manifestos. Correlations based on pairwise complete observations. *p < 0.05, **p < 0.01, ***p < 0.001.

As a final validation step, we replicate earlier findings and compare our results to those of Thau (Reference Thau2019), who manually coded group appeals by British parties between 1964 and 2015. Since Thau’s data are not publicly available, we reproduce his key findings at the aggregate level using our dataset and a comparable set of categories. Figure 4 shows the evolution of group appeals by major UK parties towards workers, the lower class, tenants, and employers. The resulting trends closely mirror those reported by Thau (Reference Thau2019): Labour’s appeals to workers and the lower class decline over time, consistent with its rhetorical shift away from traditional working-class constituencies (Evans and Tilley Reference Evans and Tilley2017), while references to employers rise steadily for both parties. Mentions of tenants also decrease over time in both datasets. Although some variation in magnitude and timing is evident (reflecting variations in coding schemes and group definitions), the overall consistency reinforces the validity of our automated approach and its ability to reproduce established patterns in party rhetoric from manually coded datasets.Footnote ¹⁹

Figure 4.

Replication and extension of Thau (Reference Thau2019).

Note: the figure shows the percentage of appeals to four group categories: (a) employees and workers, (b) lower class, (c) tenants, and (d) employers and business owners. Circles represent observed values for each party-year observation. Solid lines show fitted trends from linear regression (ordinary least squares), while dashed lines show smoothed trends from kernel-weighted local polynomial regressions.

Applications

The PSoGA Database offers numerous avenues for future research in various subfields of political science, including comparative politics, political communication, representation, party competition, populism, and electoral studies. Its structure – comprising manifesto-level and sentence-level datasets – enables both macro-level comparative analysis and more fine-grained qualitative research.

A central application of the database is to compare the frequency and diversity of social group appeals in party manifestos. The categorisation of group appeals by demographic, economic, ethnic, religious, and other dimensions facilitates cross-national and longitudinal studies of how political parties articulate their relationships with different segments of society. For example, it allows researchers to track how different party families and party types construct their group appeals and whether these patterns evolve depending on their position in government or opposition.

Moreover, the database provides information on how often parties’ symbolic appeals to specific groups coincide with substantive policy content. This enables studies of how parties translate symbolic rhetorical representation into programmatic commitments. The database can also be combined with existing datasets such as MARPOR or CHES, enabling exploration of how parties’ group appeals relate to their issue priorities or ideological positions. Similarly, integrating datasets on public opinion, such as the European Social Survey (ESS) or the Comparative Study of Election Surveys (CSES), can help assess whether group appeals become more targeted or expansive during competitive elections and how these strategies evolve in response to shifts in public opinion. Finally, pairing the database with socio-economic indicators from sources like the Organisation for Economic Co-operation and Development (OECD) or the World Bank provides an empirical basis for testing whether political rhetoric reflects real-world challenges for specific groups.

Verbatim mentions in the sentence-level dataset provide a rich resource for in-depth analyses, enabling researchers to examine the content, tone, and framing of appeals to social groups. Importantly, the sentence-level data allows for fine-grained analysis of specific groups that may be subsumed under broader categories at the manifesto level or fall outside the predefined classification scheme altogether. This includes references to subgroups like single mothers, indigenous communities, or specific occupational groups, making the dataset a valuable resource for capturing the full diversity of group-related discourse. The dataset captures nuanced rhetorical distinctions that are often lost in higher-level coding schemes and enables researchers to study not only how frequently parties refer to certain groups but also how they differentiate between subgroups, emphasise particular categories over others, or respond to evolving political contexts.

With the sentence-level dataset, it is also possible to examine the context and framing of appeals to specific groups, such as pensioners, immigrants, women, or young people, as deserving, vulnerable, empowered, or threatening. This opens the door for studies of political framing, such as how exclusionary narratives are constructed around certain outgroups or elites, and how parties draw symbolic boundaries between ‘the people’ and perceived adversaries. Similarly, the dataset also supports research into evolving discursive norms around historically marginalised or underrepresented groups, such as national or ethnic minorities. This makes the sentence-level dataset valuable for scholars working on political discourse, identity politics, populism, and political communication.