WHY THE RADICAL RIGHT CLUSTERS IN SPECIFIC GEOGRAPHIC LOCATIONS?

A large amount of research on the radical right weighs the importance of alternative individual correlates of voter support of radical-right parties (Fitzgerald Reference Fitzgerald2018; Gidron and Hall Reference Gidron and Hall2017; Guiso et al. Reference Guiso, Herrera, Morelli and Sonno2020; Inglehart and Norris Reference Inglehart and Norris2016). But recently, there has also been attention to the fact that the populist radical-right parties’ electoral support clusters in specific geographic locations (Bolet Reference Bolet2021; Cremaschi et al. Reference Cremaschi, Rettl, Marco and De Vries2022; Eliasoph Reference Eliasoph2017; Fitzgerald and Lawrence Reference Fitzgerald and Lawrence2011; Munis Reference Munis2021). Scholars have observed, for example, that in Europe and North America, economically thriving cities are locations of “cosmopolitan” attitudes, and also show less support for the radical right. By contrast, the countryside and regions in economic decline appear to be locales where radical-right parties do well electorally and where the Brexit movement found a great deal of its support in 2016. One popular account portrays this as the divide between cosmopolitan “anywheres” who are not connected to any particular location and parochial “somewheres” who are deeply attached to their home communities (Goodhart Reference Goodhart2017). Whatever is driving these spatial patterns, a growing point of convergence across this literature is that important determinants operate at the level of geographically specific areas and not just individuals (Broz, Frieden, and Weymouth Reference Broz, Frieden and Weymouth2020).

Two sets of explanations have emerged to explain these geographic patterns. One strand of research highlights how economic shocks triggered by globalization have asymmetrically affected some regions within countries more than others. For example, Autor, Dorn, and Hanson (Reference Autor, Dorn and Hanson2013) and Autor et al. (Reference Autor, Dorn, Hanson and Majlesi2016) demonstrate that American local labor markets that experienced higher unemployment due to China-related trade shocks were also regions that showed greater support for Donald Trump in 2016. Similar patterns have been on display in Europe where radical-right support (Colantone and Stanig Reference Colantone and Stanig2018b) and Euroskepticism in the United Kingdom are disproportionately found in regions harder hit by China imports (Colantone and Stanig Reference Colantone and Stanig2018a).

Additional work by Broz, Frieden, and Weymouth (Reference Broz, Frieden and Weymouth2020) highlights how long-term de-industrialization—more advanced in some locations within countries than others—has had downstream effects on labor force participation rates, youth out-migration, declining property values, local tax revenue shortfalls, and even the prevalence of the opioid epidemic and public health outcomes. And, communities marked with these conditions show higher support for populist right-wing parties. Relatedly, Cremaschi et al. (Reference Cremaschi, Rettl, Marco and De Vries2022) show that the decline of public service in rural regions increases support for the radical right.

But because the economic drivers of radical-right voting are inherently intertwined with cultural developments, a second strand of literature has zeroed in on a different facet of this problem: the escalation of an urban–rural split in electoral politics. Rodden’s (Reference Rodden2019) evidence points to the sharp and growing partisan divide between rural regions and cities, suggesting that over the course of the twentieth century in many established democracies, economic development has not led to the decline of this geographic split but rather to its growth. Today, in the United States, population density is a stronger predictor of voting patterns than at the beginning of the twentieth century (Rodden Reference Rodden2019, 4). The drivers of this phenomenon are closely related to the economic developments described above. Technological change, the decline of manufacturing, and the rise of the knowledge economy all contribute.

But in these accounts focusing on the urban–rural divide, scholars have emphasized a variety of causal pathways linking geography to radical-right support. One channel emphasized by Cramer’s (Reference Cramer2016) important ethnographic work on rural communities in the Midwest of the United States is that residents in rural communities experience resentment over rural communities’ economic dislocation and political marginality vis-à-vis urban decision-makers. Other work emphasizes the importance of a strong sense of place-based identities in the countryside (though not confined there) that also make populist appeals attractive (Bolet Reference Bolet2021; Fitzgerald Reference Fitzgerald2018; Fitzgerald and Lawrence Reference Fitzgerald and Lawrence2011; Munis Reference Munis2020; Reference Munis2021). A third channel linking the urban–rural divide to radical-right voting is embedded in conflicting social values and associated political attitudes. According to Rodden (Reference Rodden2019), for example, in dynamic knowledge-based cities in which high-skilled workers increasingly live, distinctively “cosmopolitan” values, lifestyles, and political preferences are found that sharply diverge from the less secular and more traditional values prevalent in the rural areas inhabited by lower-skilled workers who feel “left behind” by the global economy (see also Iversen and Soskice Reference Iversen and Soskice2020).

One example of work that clarifies how a clash in social values rooted in the urban–rural divide is driving politics in advanced democracies in Europe is offered in Maxwell (Reference Maxwell2019; Reference Maxwell2020), who highlights the contrast between positive attitudes toward immigration in large “cosmopolitan” cities and more negative attitudes in the “nationalist” countryside. Maxwell, more systematically than others, also attempts to sort out the sources of the attitudinal or “values” gap on this key issue that activates the radical right. He asks: Is the urban–rural divide on attitudes toward immigration contextual (i.e., something about living in cities makes people feel more positive about immigration)? Or is it compositional (i.e., something about the types of people who live in cities)? Maxwell’s evidence supports the latter position. Using panel data in which he traces voters’ residential mobility and social attitudes over time, he demonstrates that individuals moving to cities (i.e., context) do not directly become more cosmopolitan. Rather, preexisting attitude differences between urban and rural residents and self-sorting (i.e., composition) explain much of the urban–rural social values gap (see also Iversen and Soskice Reference Iversen and Soskice2020). As a result, Maxwell concludes that his “main takeaway message” is that the geographical clustering of attitudes toward immigration is usefully conceptualized as a “second-order manifestation of deeper demographic and cultural divides” (Maxwell Reference Maxwell2019, 473).

But, what exactly are these “deeper” cultural divides? And how do we study them? For example, are they best understood as the result of shared local identity and history, or rather are they the consequence of individual-level characteristics? Here is where we reach the limits of existing literature. Contemporary economic change and the growing urban–rural gap are clearly important drivers of the regional clustering of radical-right voting in economically “left behind” regions. But this literature itself suggests there are also longer-standing historical-cultural divides that preceded recent economic developments that may continue independently to shape where the radical right clusters. Even if research engages with these ideas (e.g., Fitzgerald Reference Fitzgerald2018; Fitzgerald and Lawrence Reference Fitzgerald and Lawrence2011), it is not fully clear whether these factors operate at the individual or community level. Thus, the precise content of these cultural and historical divides—and how to study them empirically—has remained elusive.

CULTURAL DIVIDES BETWEEN CENTER AND PERIPHERY

One answer to the question of the source of these regional divides is suggested in the pioneering work of Lipset and Rokkan (Reference Lipset and Rokkan1967), who argued that in the historical process of nation-state formation, a variety of political cleavages emerged to shape contemporary politics. One such important cleavage is the historical “center–periphery” division between the centralizing core that pursued the homogenizing project of nation-building and the culturally distinctive outlying areas of the nation-state that were peripheral to it.Footnote ¹

In this framework (Alonso and da Fonseca Reference Alonso and da Fonseca2012; Hooghe and Marks Reference Hooghe and Marks2016; Rokkan Reference Rokkan2009; Rokkan and Urwin Reference Rokkan and Urwin1983), during state formation, citizens in historically peripheral regions often carried lower status cultural markers—such as language, the local dialect, religion, dress, and customs. These community-level attributes provided the raw materials for a collective self-perception of marginality vis-à-vis the larger national political community. Lipset and Rokkan (Reference Lipset and Rokkan1967, 13) described these peripheral regions as “bastions of primordial local culture.” In Weber’s (Reference Weber1976, 67) classic work on French state-building, Peasants into Frenchmen, for example, Parisian primary school inspectors and tax authorities as late as the early twentieth century viewed the periphery as a land of “savages,” where civilization was absent and a “wealth of tongues” (i.e., regional dialects) hindered the operation of officialdom.

Indeed, the periphery in nation-states was usually defined by two features: their low social status within the nation-state and the prevalence of identifiable cultural markers such as dialect or nonstandard language that reinforced that status. In Scott’s (Reference Scott1998; see also Gellner Reference Gellner1993) account of state-building, for example, the state’s efforts to make society “legible” were hindered by peripheral regions’ linguistic diversity. Further, these regions were targets of what Laitin (Reference Laitin1992, 10–4) has called state-directed “language rationalization” efforts and that Hechter (Reference Hechter1972) has more pointedly described as “internal colonialism”—a process of subjugation and standardization of cultural differences that aimed at stamping out nonstandard regional languages (Reference Hechter1972, 191–205).

In this sense, “peripherality” is a community-level variable. Nonetheless, empirically its effects should play out and influence individuals living in the periphery. Citizens in such locations in the past and today believe they suffered distributive injustices in terms of power, wealth, and prestige. This in turn prompted a “politics of cultural defense” (Lipset and Rokkan Reference Lipset and Rokkan1967, 12), shaping residents’ perception of themselves, elites, and outsiders. Scholarship has contended that this center–periphery divide began to recede in salience in the middle of the twentieth century, but also that it can be activated—and some scholarship has shown—it has re-emerged in some settings in recent years as a salient political cleavage (Alonso and da Fonseca Reference Alonso and da Fonseca2012), in part explaining the emergence of regional political parties across Europe.

There are empirical and theoretical reasons to believe that the lower self-perceived status of citizens in these historically and culturally fringe regions—Rokkan’s “periphery”—can be enduring, which leads to support for anti-establishment and anti-immigrant radical-right political parties today. First, an extensive literature in sociolinguistics finds that regional dialects are more persistent than modernization theory might predict (Edwards Reference Edwards2013, 69) and that speakers of historically low-status dialects—again, especially low-status regional dialects—continue to suffer discrimination in housing, employment, wages, and negative evaluations in terms of perceived prestige, skill, trustworthiness, and education (Du Bois Reference Du Bois2019, 93; Edwards Reference Edwards2009; Purnell, Idsardi, and Baugh Reference Purnell, Idsardi and Baugh1999; Segrest Purkiss et al. Reference Segrest Purkiss, Perrewé, Gillespie, Mayes and Ferris2006). While dialect may not always connote lower status, it has been repeatedly demonstrated that in the German setting, speakers of nonstandard regional dialects still suffer prejudice in housing and labor markets (Rakić, Steffens, and Mummendey Reference Rakić, Steffens and Mummendey2011) and actually suffer a “wage penalty” when compared to speakers of standard Germany (Grogger, Steinmayr, and Winter Reference Grogger, Steinmayr and Winter2020). This status of peripherality or “marginality”—especially if is still subjectively felt by residents of these regions today—may make people susceptible to appeals from political parties that challenge “central” elites and the political establishment.

In addition, scholarship has demonstrated that at the individual level, lower social status generates support not only for anti-establishment parties but also for radical-right anti-immigrant political parties (Gidron and Hall Reference Gidron and Hall2017; Kurer Reference Kurer2020). As a literature in psychology clarifies, status threats, especially when coupled with even weak cultural markers (like dialect), can provoke strong out-group hostility (Enos Reference Enos2014; Kustov Reference Kustov2021; Tajfel Reference Tajfel1978; Tajfel, Billig, and Bundy Reference Tajfel, Billig and Bundy1971). Given these findings, we expect that residents in what amounts to low-status regions will have a propensity to vote for challenger and radical-right populist parties via two additional mechanisms: anti-elite sentiment (having experienced a history of exclusionary policies) and hostility toward outsiders, including refugees and immigrants.

Finally, we then also expect that this pattern becomes stronger in moments in which the perception of “cultural threats”—such as the large influx of refugees—emerges. The key mechanism behind this is that the perceived “uncontrolled” influx of refugees activates and further intensives feelings of lower status and outgroup hatred, as well as resentment of political elites—the individual-level mechanisms of our macro-level argument. Existing research demonstrates how “contact” can provoke anti-immigration reaction (Enos Reference Enos2014). Additionally, during the so-called refugee crisis, the public debate centered on elite failure (Harteveld et al. Reference Harteveld, Schaper, De Lange and Van Der Brug2018; Mushaben Reference Mushaben2020). Thus, if populist radical-right parties frame the influx of refugees as such a cultural threat—as the AfD did during the influx of refugees starting in 2015—one can expect this triggering effect to take hold. Similar arguments have been proposed and developed elsewhere for the German case during the so-called refugee crisis (Cantoni, Hagemeister, and Mark Reference Cantoni, Hagemeister and Mark2017; Choi, Poertner, and Sambanis Reference Choi, Poertner and Sambanis2019).

DATA

How can we study these cultural divisions, and how do we get an empirical grip on the historical center–periphery divide? We focus on one characteristic of localities we elaborated in our theoretical discussion above: language, more precisely dialects. A dialect is a subvariety of a language that differs from other subvarieties along three dimensions: vocabulary, grammar, and pronunciation (Edwards Reference Edwards2009, 63). Our strategy is to measure the strength of a dialect in a given region. As we have noted, the state-building and political development literature makes clear that a strong regional dialect that departs from the standard variant of a language is a key descriptive characteristic of the historical periphery (Rokkan and Urwin Reference Rokkan and Urwin1983; Weber Reference Weber1976). Based on further research in sociolinguistics, we know that nonstandard dialects are enduring and remain a source of a variety of social identities (including place-based identity) and can also provide cultural markers that citizens continue to use to evaluate each other’s origins and capabilities (Labov Reference Labov1963; Remlinger Reference Remlinger2009; Reference Remlinger2017).

We employ two data sources to measure how distant a given regional dialect is compared to standard German. One of them is based on contemporary data collected by the German magazine Der Spiegel (see Elspaß et al. Reference Elspaß, Leemann, Möller and Grossenbacher2018; Leemann, Derungs, and Elspaß Reference Leemann, Derungs and Elspaß2019), while the second relies on a linguistic survey of about 40,000 schools in the late nineteenth century (Lameli et al. Reference Lameli, Nitsch, Südekun and Wolf2014). We use the contemporary dialect data as a proxy for regions that were in the historical periphery in our main analyses. The historical data serve as evidence for the validity of the contemporary dialect data.

Contemporary Dialectal Data

The most comprehensive mapping of current regional German dialects is a unique online survey conducted by the German magazine Der Spiegel (see Elspaß et al. Reference Elspaß, Leemann, Möller and Grossenbacher2018). In 2015, Der Spiegel created a publicly accessible dialect quiz, where individuals answer a number of questions related to regional differences in dialect.Footnote ² Quiz respondents are shown a description of a verb, noun, or adjective. They are then asked to select the regional version of the word from a list of choices that they use. An example is a question on the informal version of the verb “to chat”: in East Germany, the majority of respondents use “quatschen,” speakers closer to the North-Western coast employ “schnacken,” and Bavarians use “ratschen.”Footnote ³ In total, respondents answer 24 questions, each aimed at the regional version of a specific noun, verb, or adjective. Importantly, these 24 questions relate to differences in pronunciation, grammar, and vocabulary, and therefore capture all relevant dimensions of what constitutes a dialect. While the quiz was not created for strictly social scientific purposes, the 24 questions are directly based on a prior linguistic research project aimed at describing regional differences in German dialects, the Atlas der Deutschen Alltagssprache (“Atlas of Colloquial German,” see Elspaß Reference Elspaß2005).

After completing the quiz, a predictive algorithm estimates the region where the person is from. Finally, respondents are asked to evaluate the accuracy of the prediction and can enter their hometown.Footnote ⁴ About two-thirds of all respondents enter their hometown, allowing us to trace responses to a specific location. In total, about 725,000 respondents indicate where they are from. We use this information to create a county-level measure of dialectal distance between a given county and region of Hannover, the area that is most strongly associated with standard German (for more elaboration on standard German and its relation to the Hannover dialect, see Mills Reference Mills1985, 142; Polenz Reference Polenz2009, 123). Given the large sample, the number of respondents in each county is high. As we show in Figure SI.1 in the Supplementary Material, the majority of counties have more than one thousand respondents. In a first step, we obtain the most common (modal) answer in each county for each of the 24 dialect questions that are part of the Spiegel quiz. For each county i, the modal answer to quiz item k takes on the value $ {X}_i^k $ . In the following, we will refer to the modal answer in each county as a county-specific dialect characteristic. Depending on the dialect characteristic, $ {X}_i^k $ can take between 2 and 24 different values. In standard German, the $ {k}^{\mathrm{th}} $ dialect characteristic takes on the value $ {X}_{\mathrm{Standard}\ \mathrm{German}}^k $ . We define the distance between a given regional dialect and standard German as follows:

(1) $$ \begin{array}{rl}{d}_i=24-{\displaystyle \sum_{k=1}^{24}}1\left({X}_i^k={X}_{\mathrm{Standard}\ \mathrm{German}}^k\right).& \end{array} $$

The sum on the right-hand side counts the number of times a region shares a dialect characteristic with standard German, which can be at most 24. We then reverse this measure such that $ {d}_i $ measures dialectal distance between a given county and standard German. It can range from 0 (a dialect that is equal to standard German) to 24 (shares no characteristics with standard German). We chose this measure of dialectal distance in accordance with prior work on the effect of dialects, chiefly Falck et al. (Reference Falck, Heblich, Lameli and Südekum2012), who use the same definition of distance in conjunction with the nineteenth-century data described in the section “Historical Dialectal Data.”

A potential drawback of our method is that it requires dialect characteristics to be exactly the same to count toward the distance measure. To ensure that the results are not driven by our choice of the dialectal distance measure, we also calculate the dialectal distance using the average Jaro–Winkler distance between the prototypical characteristics. The Jaro–Winkler distance accounts for cases when dialect characteristics are similar, but not exactly the same (for more details, see Cohen, Ravikumar, and Fienberg Reference Cohen, Ravikumar and Fienberg2003). The two measures are highly correlated and our main results are similar across the two dialectal distance measures.

In Figure 1, we present the county-level distance from standard German across Germany. Unsurprisingly, the counties surrounding the Hannover region (shaded in dark green) are most similar to standard German. We also observe a pronounced North–South divide: Southern German dialects are markedly more different from standard German than in the Northern part of the country. The two Southernmost states, Bavaria and Baden-Wuerttemberg, exhibit the greatest distance to standard German.

Figure 1. Dialectal Distance from Standard German by County (Landkreis)

Note: Greater values indicate that the local dialect is more distant from standard German. The region shaded in dark green in the Northern central part of the country is Hannover, the region that most closely approximates standard German. The map also includes locations of major German cities, including Hannover. State borders are indicated in black. Counties shaded in gray are counties for which data are missing.

Through the use of a supplementary dataset on individual-level dialect use, we corroborate that our aggregate measure broadly corresponds to individual patterns of dialect use (we describe these data in more detail in Section A.8 of the Supplementary Material). We find that self-reported dialect usage is much more common in areas that are more distant from standard German, using our aggregate measure. We can, therefore, confirm that individuals are likely aware of the aggregate-level dialectal distance, as individuals in distant regions (as measured by our aggregate variable) and are much more likely to know and use dialects.

Figure 1 suggests that physical distance to Hannover, as well as North–South or East–West divides, are correlated with dialectal distance. We address this directly by including state fixed effects in all our specifications. The fixed effects imply that we only compare differences in dialectal distance within federal states, rather than across states. This ensures that our results are not driven by East–West or North–South divides. What is more, we also control for the physical distance between counties and their respective state capitals. Finally, we present additional robustness checks (see the section “Robustness” and Table A.11 in the Supplementary Material) to show that controlling for physical distance to the German border or to Hannover itself does not change our substantive conclusions. We note that the correlation between aggregate-level dialectal distance and population density is close to zero (it is $ -0.06 $ ), indicating that the center–periphery divide is not simply reducible to the urban–rural divide. These steps ensure that our dialectal distance variable does not simply pick up on urban–rural divides or geographic peripherality.

In an additional supplementary analysis, we examine potential correlates of dialectal distance, based on survey- and aggregate-level data (see Section A.4 of the Supplementary Material). We observe that stronger dialects correlate with lower integration into surrounding areas, increased skepticism toward outsiders, and stronger suspicion of elites. Generally, these patterns corroborate that dialectal distance is indicative of center–periphery divides.

EMPIRICAL STRATEGY

Our empirical strategy proceeds in three steps, relying on aggregate- as well as individual-level data. First, we assess the relationship between dialectal distance and county-level AfD support in the 2017 general election.Footnote ⁶ These models include a number of relevant covariates as well as fixed effects to account for unexplained regional heterogeneity. We focus on potential confounders that have previously been shown to predict voting for radical-right parties, including population density, socioeconomic indicators, and distance to the border.

Second, we complement the aggregate-level election results with individual-level data from the GLES. In contrast to the aggregate-level analysis, the GLES data allow us to control for a set of key alternative explanations on the individual level, such as socioeconomic background, education, nationalism, and local attachment. The outcomes we use are a binary measure of AfD vote intentions as well as an 11-point AfD likability scale.Footnote ⁷ Given the high frequency at which the data are observed—18 waves between 2013 and 2018—we can further use these data to study how the relationship between dialects and radical-right support varies over time.

Finally, we shed light on the mechanisms that underlie our proposed argument through additional analysis of survey data that measure perceived status, anti-elitism, and anti-immigrant attitudes. The goal of this final exercise is to provide support for the individual-level mechanisms that animate our macro-level argument.

Throughout the “Results” section, we estimate a set of linear models that can be described as follows:

(2) $$ \begin{array}{rl}{y}_{i,j}=\alpha +{\gamma}_j+\tau {d}_{i,j}+{\beta}^{\prime }{X}_{i,j}+{\varepsilon}_{i,j},& \end{array} $$

Here, $ {y}_{i,j} $ is the outcome of interest for unit i in state j, which is either radical-right voting or individual attitudes. Our main independent variable is $ {d}_{i,j} $ , the dialectal distance between a given county i and standard German. We always standardize $ {d}_{i,j} $ such that coefficients can be interpreted as the effect of a one-standard-deviation increase in dialectal distance. We also include a vector of covariates $ {X}_{i,j} $ as well as state fixed effects $ {\gamma}_j $ .Footnote ⁸ For all county-level models, we control for regional GDP/capita, average wages, population density, unemployment rates, total population, % Catholic, the share of commuters, the physical distance to the state capital, and the CDU/CSU vote share in the 2013 general election. We provide summary statistics for all outcomes and explanatory variables in Table A.1 in the Supplementary Material.

Causal identification is difficult in the context of our research question. We rely on the assumption that dialectal distance is independent of the potential electoral outcomes, conditional on county covariates and fixed effects. The strongest threat to identification is omitted variable bias, namely that unobserved factors drive our findings. In a series of robustness checks, we address confounding through additional control variables, more fine-grained geographical fixed effects, a design-based weighting approach, as well as a sensitivity analysis. We particularly highlight the results of the sensitivity analysis. We demonstrate, even if there was an unobserved confounder with a partial correlation that is twice as strong as the one between the unemployment rate and AfD support, its inclusion would not change our substantive conclusions. Given that we control for a host of relevant explanatory factors, the existence of an unobserved confounder of this magnitude appears extremely unlikely. As we elaborate in the section “Robustness,” the additional analyses thus suggest that our main results are not merely due to unobserved confounding.

ROBUSTNESS

In addition to the main results, we conduct several additional checks to ensure that our results are not driven by unobserved confounding, idiosyncrasies in the sample, the choice of covariates, the model specification, or the operationalization of the outcome.

In a first step, we show that the results are robust to including two additional controls. To ensure that our measure of dialectal distance is not confounded by a history of nationalism or racism, we add an indicator for a history of pogroms in the 1920s (taken from Voigtländer and Voth Reference Voigtländer and Voth2012) and the NSDAP (Nazi party) vote share in 1933 as additional controls. In columns 2 and 3 of Table A.8 in the Supplementary Material, we show that the effect size and significance remain unchanged when we include these controls.

Second, we demonstrate that our results are robust to an alternative operationalization of the outcome. As shown in the section “Contemporary Dialectal Data,” our dialectal distance measure requires exact matches between characteristics of regional dialects and standard German. We relax this requirement in Table A.9 in the Supplementary Material. Here, we instead use the Jaro–Winkler distance, which takes into account words that are similar, but not exactly the same.Footnote ¹⁸ We show that the choice of dialectal distance measure does not change our conclusions. Substantively, the effect sizes are similar to what we show in our main specification in Table 1.

Third, we use additional controls and fixed effects to alleviate concerns of unobserved spatial confounders. First, we add fixed effects for Regierungsbezirke (“administrative districts”), the administrative level below federal states.Footnote ¹⁹ While state fixed effects likely already account for a large degree of unobserved regional heterogeneity, the smaller administrative districts allow us to account for an even greater amount of spatial differences. As we show in Table A.10 in the Supplementary Material, adding administrative district fixed effects does not change our substantive conclusions. Second, we present results from alternative specifications that control for (i) the physical distance to Hannover and (ii) the physical distance to the German border (see Table A.11 in the Supplementary Material). We again find that the main result remains unchanged.

Fourth, we use a bootstrap approach to examine whether our results are affected by uncertainty in the measurement of our dialectal distance variable. As mentioned before, the county-specific dialectal distance measure is based on a sample of dialect quiz respondents who are from the county in question. To incorporate measurement uncertainty, we re-sample from all quiz respondents and then calculate all county-specific dialectal distances using the resulting five hundred bootstrap samples. We then re-estimate our main models for each bootstrap sample. In Section A.7.1 of the Supplementary Material, we show that incorporating measurement uncertainty in this manner does not change our conclusions

Fifth, we address post-treatment bias through the use of the sequential g-estimator (see, e.g., Homola, Pereira, and Tavits Reference Homola, Pereira and Tavits2020, in a similar setting). As stated before, the majority of our control variables are likely measured after “treatment,” that is, after the development of local dialects. In Section A.7.2 of the Supplementary Material, we demonstrate that our conclusions remain unchanged when accounting for post-treatment bias.

Finally, we use an alternative, design-based approach to estimate treatment effects. In doing so, we rely on the covariate balancing propensity score (CBPS) for continuous treatments (see Imai and Ratkovic Reference Imai and Ratkovic2014) to (1) estimate a propensity score model for treatment assignment and (2) obtain covariate balancing weights. The propensity score uses all covariates and state fixed effects that we include in our main model. While the dialectal distance treatment remains correlated with some of the covariates, CBPS weighting greatly improves balance. In Section A.12 of the Supplementary Material, we present the results. In the weighted models, the estimated coefficient for the relationship between dialectal distance and AfD voting is comparable in magnitude and significance to our base models.

Sensitivity to Unobserved Confounding

As an alternative approach to address confounding, we implement a sensitivity analysis (Imbens Reference Imbens2003). Although we already control for several social, demographic, and economic variables, it is impossible to account for all possible confounders. One alternative approach would be to find a suitable instrument for our dialectal measures. Yet finding an instrument that fulfills the exclusion restriction for dialectal distance—which has deep cultural and historical roots—seems unlikely.

Instead, we implement a sensitivity analysis to gauge how strong an unobserved confounder would have to be to invalidate our findings (for a more in-depth discussion of such methods, see Imbens Reference Imbens2003). We implement the sensitivity analysis using the method and package developed by Cinelli and Hazlett (Reference Cinelli and Hazlett2020). We use the main results from the second model in Table 1 as the baseline model for the sensitivity analysis.

In Figure 4, we present the results of the analysis. A point in the plot represents a hypothetical unobserved confounder. The x-coordinate represents the partial $ {R}^2 $ of the confounder with respect to the treatment (dialectal distance) and its y-coordinate represents the partial $ {R}^2 $ with respect to the outcome (radical-right voting). For reference, we have included the partial $ {R}^2 $ values for the unemployment rate as well as a hypothetical confounder that is twice as strong as the unemployment rate (see also Table A.13 in the Supplementary Material, where we show the same quantities for all covariates). The numbers shown next to the variable names in the plot indicate the effect size if an unobserved confounder with the same strength was included in the model. If we had failed to include a confounder that is as twice as strong as the unemployment rate—by far the strongest county-level predictor in our models—the estimated effect size would drop to 0.73. To put it differently, even if there was an unmeasured confounder as twice as strong as the unemployment rate, adding it would not change the substantive conclusions of our article.

Figure 4. Sensitivity Analysis

Notes: Results from the sensitivity analysis proposed by Cinelli and Hazlett (Reference Cinelli and Hazlett2020). The plot indicates how strongly confounders would have to be correlated with the treatment and the outcome to reduce the estimated effect size to zero (dashed red line). The original effect size from Table 1 is shown in the bottom-left corner—the scenario where there is no unmeasured confounding. The red diamond shape indicates partial correlations for unemployment rates and a hypothetical confounder that is twice as strong as the unemployment rate. The interpretation of the figure is as follows: if we were to include a confounder as strong as the unemployment rate, the estimated effect size would drop to 0.95. For a confounder that is twice as strong as the unemployment rate, the estimated effect of dialectal distance on radical-right voting would drop to 0.73.