Shared war experiences and the formation of in-group/out-group preferences

The impact of political violence on societal beliefs, attitudes and behaviour is a well-established area of research in social sciences (Acharya, Blackwell, and Sen Reference Acharya, Blackwell and Sen2016; Rozenas, Schutte, and Zhukov Reference Rozenas, Schutte and Zhukov2017; Zhukov and Talibova Reference Zhukov and Talibova2018). An extreme form of political violence, war, is known to deeply scar communities by inflicting costly physical destruction, altering the demographic composition of the population and, above all, producing a grim legacy of wasted human life (Besley and Reynal-Querol Reference Besley and Reynal-Querol2014; Gay Reference Gay2023; Hadzic Reference Hadzic2018).

One robust pattern established in the literature is that a shared experience of war can heighten social capital within a community, increasing cooperative and altruistic behaviour (Bauer et al. Reference Bauer, Blattman, Chytilová, Henrich, Miguel and Mitts2016). However, such pro-social behaviour is confined to one’s own in-group, often defined in ethnic, racial or national terms. War renders individuals more attuned to the fragility of their own experience as a community, making them willing to protect the survival of that particular experience at the expense of others (De Juan et al. Reference De Juan, Haass, Koos, Riaz and Tichelbaecker2021). This is corroborated by experimental research in psychology, which shows that experiencing the death of a loved one increases one’s sense of belonging to a broader identification structure, such as kin or nation (Bauer et al. Reference Bauer, Cassar, Chytilová and Henrich2014). In societies where such a structure of identification extends beyond consanguinity, personal loss is not even a necessary condition for triggering heightened identification, as the indirect victimisation produced by learning the stories of others in the same community is sufficient.

Most often, an increase in in-group identification comes at the expense of a decrease in out-group affinity (Bauer et al. Reference Bauer, Cassar, Chytilová and Henrich2014; Canetti et al. Reference Canetti, Hall, Rapaport and Wayne2013). There are several reasons why this might happen. First, out-group members may be perceived as having a wholly different experience of war, shaped by alternative sources of communal trauma which, in an environment characterised by limited collective memory resources, would compete directly for salience with the experience of in-group members. Second, the experience of the out-group may be antithetical to that of the in-group and, although both are subjective in nature, could lead to members of the out-group being perceived as a threat to the validity of the in-group experience. Third, in more extreme cases, the out-group could be blamed for the trauma that produced the in-group experience if the two groups were on opposite sides of the conflict.

How does this in-group preference manifest itself in terms of political behaviour? In democracies, the most obvious form of potential manifestation is through voting, by supporting political parties whose message is similarly favourable to the in-group at the expense of the out-group. An obvious candidate for tracing the political effects of in-group preferences would be the far-right, which throughout modern European history has consistently displayed (ultra)nationalist tendencies, nativism, chauvinism and xenophobia, all elements of extreme support for the national in-group.

Empirical evidence supports this hypothesis, showing that communities most affected by war tend to disproportionately support far-right political parties or otherwise exhibit behavioural patterns congruent with far-right ideology. Acemoglu et al. (Reference Acemoglu, De Feo, De Luca and Russo2022) show that variation in the number of casualties inflicted on the Italian army during WWI was a major factor in the rise of fascism in some regions of the country. Similarly, Koenig (Reference Koenig2023) shows that German regions with more WWI veterans gradually moved to the right of the political spectrum, supporting conservative parties that enabled the erosion of democracy and the rise of the Nazi Party. Providing even more direct evidence of the link between WWI and Nazism, De Juan et al. (Reference De Juan, Haass, Koos, Riaz and Tichelbaecker2021) s show that in the interwar period, support for nationalist parties was significantly higher in places with above-average wartime human losses. Finally, Cagé et al. (Reference Cagé, Dagorret, Grosjean and Jha2023) find that French communes that were more strongly represented during the Battle of Verdun, led by WWI hero and future leader of Vichy France, General Pétain, also had higher levels of Nazi collaboration during the German occupation of World War II.

But how do far-right parties manage to frame the experience of war in terms of national identity at the expense of a more localised imagined collective (e.g. the commune), given that local communities within the same country experience war very differently, but the national experience is homogeneous? Kocher, Lawrence, and Monteiro (Reference Kocher, Lawrence and Monteiro2018), building on the historical work of Keith (Reference Keith2013), suggest that the answer lies in the way modern states preserve the memory of war. When societies rely on a compulsory system of public education, children are exposed to a common national curriculum that includes the creation of a single image of collective trauma, thus producing the national in-group. This is not to say that the far-right manipulates the education system in order to weaponise the memory of the war, but that it indirectly benefits from the fact that the curriculum construes a shared experience at the national level, which reinforces the French people as the in-group towards which preferences are generated.

However, inter-community local differences are not erased by the school system, but become sources of variation in the perceived salience of the national trauma, rather than sources of variation in the form of the trauma. In short, French departments exposed to more wartime losses are more likely to have a strong attachment to the nationally defined in-group. All departments are affected by the memory of war, as the entire country was at war, but the intensity of this attachment is determined by local factors, such as the number of deaths or the extent of infrastructure destruction. This variation in the intensity of remembrance is also likely to persist because the practice of memory is endogenous to local institutions: political elites from departments more affected by the war are more likely to push for more stringent memorialisation and commemoration, thereby reinforcing the strength of the local narrative and thus the local preference for the in-group.

The lingering effects of shared war experiences

The remaining question is whether the in-group favouritism triggered by the shared experience of war within a community, which seem to translate into support for the far-right, persists over time, or whether the shock is transitory. At a more general level, the idea that significant historical conditions have long-term effects has gained traction in recent decades (Cirone and Pepinsky Reference Cirone and Pepinsky2022), but there is limited empirical evidence that wars constitute such a condition, most of it based on episodes of civil war rather than interstate violence. There is no a priori reason to believe that interstate violence would necessarily deviate from the patterns established in civil wars, but in the absence of explicit evidence we are faced with a gap, and need to rely on the literature on the persistent effects of civil wars.

In this sense, Costalli and Ruggeri (Reference Costalli and Ruggeri2019) argue that civil unrest changed the electoral geography of Italy by 1960, increasing the organisational capacity of the Italian Communist Party in the very same regions where Communists had fought against Nazi and Fascist forces during WW2. Drawing on the Spanish Civil War, Villamil (Reference Villamil2021) shows that regions where clandestine leftist networks operated during the war and were exposed to state-sponsored atrocities show higher levels of support for leftist parties in the long term. Finally, Rozenas, Schutte, and Zhukov (Reference Rozenas, Schutte and Zhukov2017), s who examine the long-term effects of Stalinist deportations in 1930s Ukraine, show that communities exposed to such atrocities are systematically less likely to support pro-Russian parties.

Theoretically, for war to be a critical juncture in the evolution of a society’s political equilibria, two necessary conditions must be met. First, there must be a causal mechanism, embedded in the socio-political context, through which the lasting effect of the war propagates. Second, there must be a political actor who becomes a mnemonic entrepreneur, willing to exploit the memory of war to activate the latent in-group preferences for electoral purposes. More specifically, we would need to identify an entity that directly benefits from the association between in-group preferences and nationalism, most likely by emphasising the latter as part of its political strategy. When both elements are present, the mechanism of persistence becomes clear: an interested political actor, who is a supplier of nationalism within the political system, leverages the infrastructure of collective memory to activate the community’s latent preference for in-group members, which translates into an increased demand for nationalism.

The first consideration, then, is what is the infrastructure through which the memory of war is propagated. A large body of work in sociology, anthropology and international relations explores precisely the politics of memory – how societies maintain, reproduce and instrumentalise shared experiences, such as wars (Malinova Reference Malinova2021; Maurantonio Reference Maurantonio2014; Müller Reference Müller2002; Uhl and Golsan Reference Uhl and Golsan2006). The politics of memory are the central link between war as a lived experience and war as a source of enduring nationalistic sentiments.

The starting point for making this connection is to recognise that nationalism, both as an ideology and as a form of political praxis, is rooted in the development of historical representations of shared narratives (Bell Reference Bell2003). But not all historical narratives are created equal. Central to the formation of the most enduring narratives of nationalism are episodes of collective trauma that trigger a social response of sufficient magnitude to construct and reinforce a common sense of belonging (Bell Reference Bell2006; Edkins Reference Edkins2003). Wars, especially the World Wars, are the epitome of such trauma, and there is empirical evidence of a link between wars, including WWI, and displays of nationalism (Cagé et al. Reference Cagé, Dagorret, Grosjean and Jha2023; De Juan et al. Reference De Juan, Haass, Koos, Riaz and Tichelbaecker2021).

This sense of nationalism, built on the foundations of shared historical experiences such as war, is reinforced and reproduced through the practice of memory. This practice is especially pronounced in Western democracies in relation to large-scale events such as the World Wars. To avoid repeating the mass violence that marked their history in the first half of the ${20^{th}}$ century, countries across Europe created multiple channels through which citizens could learn, mourn and remember what did not need to be repeated. Remembrance has become an integral part of the social fabric of politics, an essential cog in the nation-building machine (Turner Reference Turner2006). This phenomenon is particularly pronounced in France, where it has been at work since the interwar period (Sherman Reference Sherman1999). Visual reminders, including monuments, statues and cemeteries, are found throughout these countries and there are very few communities without any of them (Borghi Reference Borghi2021), ensuring access to political memory and, from the perspective of memory entrepreneurs, access to sources of identity (re)formation. Complementing these visual queues, school curricula try to explain the meaning of the World Wars and the connection between them in order to give children a sense of French historical completeness, a foundation of nationalism (Shapiro Reference Shapiro1997).

The politics of memory provide a plausible infrastructure for how the memory of war could have proliferated in direct relation to nationalist attitudes. The next consideration is which actors would be interested in exploiting this infrastructure. Theoretically, we would need to identify an agent who would win from this action, and who has the organisational power to perform it. The latter condition is more important because it limits the class of actors to be analysed to those most relevant in the political system, which for the time being remain the political parties. In fact, many studies show that political parties and their leaders are among the most important mnemonic actors, as they tend to take on the role of memory entrepreneurs whenever political remembrance suits them (Korycki Reference Korycki2023). Among them, far-right political parties, such as the FN in France, mobilise precisely that part of war trauma that resonates with their target electorate, constructing an out-group by othering another country, a domestic political elite, an ethnic group or immigrants from a distant nation (Couperus, Tortola, and Rensmann Reference Couperus, Tortola and Rensmann2022; Couperus, Rensmann, and Tortola Reference Couperus, Rensmann and Tortola2023; Soffer Reference Soffer2022; Zavatti Reference Zavatti2021). In particular, French patriotism and the homogeneity of national identity, to which the far-right has appealed over the years, are closely linked to the wartime mobilisation of WWI (Purseigle Reference Purseigle2023; Smith, Audoin-Rouzeau, and Becker Reference Smith, Audoin-Rouzeau and Becker2003). Nationalism as an all-encompassing category, sometimes in exclusionary forms, was consolidated in France during this period (Fuller Reference Fuller2014; Tombs Reference Tombs2003).

Simply put, the memory of war losses can trigger nationalist tendencies that increase the electoral standing of the far-right, the main proponent of exclusionary nationalism in contemporary politics.

In addition, the losses of WWI could be perceived as the French state sending its citizens to a certain death, creating an environment of mistrust. Mistrust is one of the most persistent narratives in societies, passed on from generation to generation (Becker et al. Reference Becker, Boeckh, Hainz and Woessmann2016; Lupu and Peisakhin Reference Lupu and Peisakhin2017). In the long run, accumulated mistrust benefits the far-right (Berning and Ziller Reference Berning and Ziller2017), which is able to portray itself as creating a rupture with the status quo and the remaining institutions against which people are hostile, weaponising not only the memory of war but the cynicism towards formal institutions that this memory engendersFootnote ² .

We therefore have both the infrastructure and the political actors who could use it. On the basis of these theoretical considerations, as applied to the case of France post-WWI, I propose the following argument. The deaths experienced by local communities during WWI increased in-group preferences which, in the case of France, manifest themselves as preferences for the French nation, i.e. nationalism. Moreover, because different death rates act as triggers of different intensity, they produce shifts in in-group preferences of different magnitude and hence differential demand for nationalism. Since far-right parties are the main suppliers of political nationalism, both in terms of discourse and policy, I expect that the demand for nationalism will in turn increase the electoral appeal of such political forces. Given that since the interwar period, far-right movements and parties have instrumentalised the memory of the war to consolidate their position and justify their (ultra)nationalist positions (Cagé et al. Reference Cagé, Dagorret, Grosjean and Jha2023), I hypothesise that support for the far-right is persistently higher in French departments where the death rate during WWI was higher. As the main far-right party in France, I expect the main beneficiary of this effect to be the FNFootnote ³ . A visual summary of the argument can be seen below, in Figure 1.

Figure 1.

Visual summary of the theoretical argument.

The argument is not to say that economic factors do not shape patterns of support for the far-right in France, or even to deny that they may be a central cause. Recent literature has shown that economic and cultural factors work in concert to favour the far-right when the circumstances are right for this type of actor (Margalit Reference Margalit2019; Rodrik Reference Rodrik2021; Scheiring et al. Reference Scheiring, Serrano-Alarcón, Moise, McNamara and Stuckler2024) Moreover, it could also be the case that the deaths of WWI shaped the deindustrialisation trajectory of the French departments, further reinforcing the effect of weaponising the memory of loss. This is especially likely given the strong correlation between death rates and the industrial profile of each French departments (Gay and Grosjean Reference Gay and Grosjean2023), as well as the long-term effect of WWI on the French labour market (Gay Reference Gay2023). My argument does not contradict claims of such additional reinforcing mechanisms; the more modest claim is that deaths during WWI are a contributing, persistent factor in inducing support for the FN. Future work should consider whether far-right political parties, which weaponise deeply rooted community norms about culture, are more successful in their efforts where there are appropriate economic circumstances. In other words, while cultural practices may be a sufficient condition for the success of the far-right, specific economic circumstances may be a necessary condition.

Data

My main independent variable is the department-level death rate in WWI, calculated as the ratio between the number of deceased French soldiers born in a department and that department’s total population according to the 1911 census, the last before the start of war. I get these from the Morts Pour la France database, compiled by Gay and Grosjean (Reference Gay and Grosjean2023), s which is based on over 1.5 million individual files from the archives of the Bureau of Archives of Victims of Contemporary Conflicts. The average department-level death rate in France during WWI was 3.89% of the population, with a standard deviation of 0.7. The department with the highest death rate was in Lozère, with 5.53%, while the department with the lowest death rate was Hauts-de-Seine with 1.25%. Alternatively, I also calculate the median death rate in each department as a function of the number of deaths in French communes, in order to deal with the risk of outliers. Figure 2 visually displays the distribution of deaths across the French territory. In the online Appendix, Figure B1 shows histograms of the distribution of death rates across the French departments, demonstrating that (a) there is significant spatial variation that is not visibly clustered, and (b) the two operationalisations of the independent variable exploit different variations and are thus complementary.

Figure 2.

Map of WWI death rates across French municipalities.

In terms of dependent variables, I leverage various measures of contemporary support for the far-right in France. For the main analysis, I measure the share of votes received by FN in national legislative elections since 1990Footnote ⁴ . I use data from the European NUTS-level dataset (EU-NED), which allows for a smooth matching based on the spatial characteristics of French departments (Schraff, Vergioglou, and Demirci Reference Schraff, Vergioglou and Demirci2023)Footnote ⁵ . In terms of scope conditions, I only look at post-1990 elections for three reasonsFootnote ⁶ .

First, while the FN’s consolidation began with the 1986 legislative elections (DeClair Reference DeClair1999; Shields Reference Shields2007), their remarkable results were due to an atypical electoral system for France, proportional representation, which had been implemented by François Mitterrand to cushion the expected defeat of his party (Knapp Reference Knapp1987). To understand the long-term electoral fortunes of the contemporary French far-right, it is necessary to exclude this particular outlier from the analysis. Second, French politics, traditionally more dominated by far-left parties than other West European party systems, changed markedly with the fall of the USSR (Bull Reference Bull1995). Looking for patterns of far-right support after 1990 thus allows to disentangle the complex effects of the Cold War as an external driver of political behaviour and to focus on domestic factors. Third, this period corresponds to a more modern version of the French party system, which is of primary interest in deciphering contemporary political events (Bornschier and Lachat Reference Bornschier and Lachat2009). In the online Appendix, I show that using all electoral results from 1986 onwards does not affect the results.

To avoid any ecological fallacy, I also examine support for the FN at the individual level. For this, I use the Sub-national Context and Radical Right Support in Europe (SCORE) survey (Evans and Ivaldi Reference Evans and Ivaldi2021), which provides a nationally representative sample of 19,454 French respondents and a series of relevant questions, such as (1) whether they voted for Marine Le Pen in the first round of the 2017 presidential election, or (2) how they rate the FN on a Likert scaleFootnote ⁷ .

Finally, to verify my main argument on the mechanisms of support for the far-right, I construct additional dependent variables that capture not only the shift in revealed electoral preferences (i.e. explicit support for the FN), but also the electoral distribution around certain salient conflicts of the multidimensional French political space. Formally, $X = \left\{ {{X_1},{X_2},{X_3}, \ldots, {X_N}} \right\}$ is the set of political parties participating in each given election in France. Each party ${X_K} \in X$ has an ideological position ${I_{K,C}}$ for every political conflict $C$ (e.g. taxation, nationalism, agriculture, etc.). The positions are measured before the election, and subsequently we measure the electoral performance of all parties through the vote share ${{\rm{\pi }}_K}$ that they have obtained in the first election after revealing their ideological position ${I_{K,C}}$ . With this information, I compute ${E_C}$ , the political conflict equilibrium over each conflict $C$ :

(1) $${E_C} = \mathop \sum \limits_{K = 1}^N {I_{K,C}} \times {{\rm{\pi }}_K}$$

The political conflict equilibriumFootnote ⁸ , as a measure, satisfies a number of intuitive conditions that one would expect from any metric of the central tendency of electoral preferences. First, it aggregates individual party positions based on their electoral performance. This means that it takes into account the preferences of the majority of the electorate. If a party with a particular position ${I_{K,C}}$ wins a significant share of the vote, that position will have a greater impact on ${E_C}$ . However, smaller parties with extreme positions won’t disproportionately affect the equilibrium unless they win a significant share of the vote ${\pi _K}$ . In other words, it summarises the average preference of the voters for different ideological propositions supplied by the parties.

Second, the political conflict equilibrium varies as parties’ ideological positions or electoral performance change from election to election. This is important because it allows the measure to adapt and reflect changing political landscapes or evolving voter preferences over time. Furthermore, ${E_C}$ varies not only across elections but also across space, as different electoral constituencies produce different weights for the ideological positions of political parties due to the different vote shares obtained by the parties across electoral constituencies, making this measure also suitable for sub-national analysis, like in the case of this paper.

Third, ${E_C}$ does not inherently favour party systems with more or fewer parties, important in the case of the ever-changing French system. While the number of terms in the sum depends on the number of parties $N$ , the weight of each term is determined by the party’s vote share ${\pi _{{X_i}}}$ . This further ensures that ${E_C}$ can be used to compare political equilibria over time, regardless of the exact number of parties positioning themselves on different political conflicts.

Simply put, I construct such political conflict equilibria as the weighted average of the policy positions endorsed by political parties before an election, where the weights are given by the share of votes each party received in the immediate election after the formal policy endorsements were made public. While the main dependent variable used in this paper captures explicit support for the far-right, the political conflict equilibria provide corroborating evidence for my theory, as they measure whether the entire political space has gravitated towards the positions endorsed by the far-right, and whether these changes are (partly) driven by WWI.

I calculate political conflict equilibria for conflicts relevant to far-right parties: multiculturalism, nationalism, traditionalism, internationalism, Europeanism, and the unidimensional right-left index. In terms of data, I rely on EU-NED for department-level vote shares obtained by French political parties, and on the Manifesto Project dataset (Lehmann et al. Reference Lehmann, Burst, Matthieß, Regel, Volkens, Weßels and Zehnter2022) for measures of party positions on the issues of interest. Following established trends in the text-as-data literature, I use the log transformation proposed by Lowe et al. (Reference Lowe, Benoit, Mikhaylov and Laver2011) to scale political preferences from coded political texts.

Model specification

I estimate the following regression:

(2) $${Y_{t,d}} = {\rm{yea}}{{\rm{r}}_t} + {\rm{grid}} {\hbox -} {\rm{cel}}{{\rm{l}}_d} + \gamma \cdot {\rm{death}}\_{\rm{rat}}{{\rm{e}}_d} + \phi ' \cdot {G_{\bf{d}}} + \rho \cdot X + {\varepsilon _{t,d}},$$

${Y_{t,d}}$ is the vote share obtained by the FN in department $d$ during election $t$ ; ${\rm{yea}}{{\rm{r}}_t}$ and ${\rm{grid}} - {\rm{cel}}{{\rm{l}}_s}$ are year-fixed and grid-cell-fixed effects, respectively, which absorb the impact of time-invariant and region-invariant features of departments; $\phi '{G_d}$ describes a set of spatial covariates, including the degree of urbanisation of a department, as well as whether the department is predominantly coastal or mountainous; $X$ are the respective Moran eigenvectors of the grid; the main coefficient of interest is $\gamma $ , which measures the difference in vote share between departments induced by differential WWI death rates; ${\varepsilon _{t,d}}$ is the error term.

The necessary condition for $\gamma $ to represent a causal effect is that WWI death rates are not determined by factors that also determine contemporary support for the FNFootnote ⁹ . In the language of natural experiments, the “assignment mechanism” of the WWI death rates needs to be as-if-random, which would ensure that the departments are, in expectation, homogeneous in terms of pre-WWI characteristics (Kocher and Monteiro Reference Kocher and Monteiro2016, p. 954). However, validating the as-if-random assumption is beyond the scope of statistical testing, as it requires knowledge not only of the distribution of observable characteristics of French departments at the onset of WWI, but also of unobservable characteristics. Instead, we need to reconstruct the process through which departments were ’assigned’ a particular death rate during WWI, and if certain factors are found to systematically influence the assignment, which might also shape political preferences, we need to control for them.

Fortunately, the question of what factors best explain the differences in rates between departments is not a new one for historians. There is a large body of literature on how the French army conducted its military operations between 1914 and 1919 and on the factors that determined battlefield mortality (Greenhalgh Reference Greenhalgh2014; Porch Reference Porch1988; Prete Reference Prete1985). In a nutshell, death rates are a function of the shifting territorial priorities of the war effort, which can be broken down into three elements: recruitment patterns, troop deployment and battle strategy. The latter is plausibly unrelated to the political preferences of the citizens of the French departments, so only the first two must be examined.

In the beginning, citizens recruited from the same department were usually sent to the same front, ensuring a fairly homogeneous distribution of deaths as combat situations became more dispersed. As the war of attrition progressed, people from all corners of France were sent precisely when and where they were most needed, regardless of their cohort, socio-demographic characteristics or department of origin. From that point on, death rate heterogeneity was determined by how the French authorities decided to respond to the industrial needs of their war machine, to the disruptions caused by the German occupation of the north-east, and to the shortages of artillery weapons in the early stages of the conflict (Bostrom Reference Bostrom2016). Naturally, the more urban departments, where citizens were equipped with skills more suited to industrial work, became the primary sources for compensating for production shortfalls. Accordingly, as demonstrated by Gay and Grosjean (Reference Gay and Grosjean2023), the strongest predictor of death rates among the population of a department is the degree of urbanisation of that department: the more urbanised, the lower the death rate. In fact, the degree of urbanisation explains over 75% of the variation in death rates between French departments, and once urbanisation is taken into account, other socio-demographic or geographical characteristics do not significantly improve the predictive power. This is crucial because it implies that pre-WW1 political and demographic differences between departments cannot explain the main independent variable; even if they are related to the dependent variable, they are not confounders by construction and their inclusion in the regression is not necessary, but should be avoided (Cinelli, Forney, and Pearl Reference Cinelli, Forney and Pearl2022).

Thus, while mortality is not distributed as-if-randomly, once the effect of urbanisation is taken into account, we could treat the distribution of World War I mortality rates as a quasi-natural experiment (DiNardo Reference DiNardo2010). However, as the relationship I am investigating occurs in the long run, with several other events taking place, including another World War, we need to be wary of spatial noise that might capture alternative causal channels (Kelly Reference Kelly2020). I carry out the adjustment protocol that alleviates concerns over alternative explanations in two steps.

First, I control for the degree of urbanisation ${G_d}$ of each French department using spatial data collected by the European Commission’s Geographic Information System (GISCO). In addition, to improve robustness and decrease uncertainty in the estimates, I include Moran eigenvectors in the regression model; including these eigenvectors is a statistical technique that explicitly incorporates any spatial patterns predictive of the outcome into the regression model, thus avoiding the estimation being sensitive to spatial noise (Dray, Legendre, and Peres-Neto Reference Dray, Legendre and Peres-Neto2006; Griffith and Peres-Neto Reference Griffith and Peres-Neto2006). Thus, even if the effect of post-WWI events that took place in France is correlated with the effect of WWI on electoral options, the Moran eigenvectors remove these associations, leaving only the direct effect of interest. Then, I include polynomials in the latitude and longitude of the centroid of each department, which deal with potential spatial gradients. As suggested by Kelly (Reference Kelly2020), I control at least for ${2^{nd}}$ degree polynomials, with and without interaction terms between coordinates, but not any higher degree to avoid over-fitting (Gelman and Imbens Reference Gelman and Imbens2019). I do not control for economic characteristics of the departments after WWI that might explain support for the FN, as these are post-treatment variables that potentially capture mechanisms through which WWI death rates induce patterns of support for the far-right. In this regard, I follow the advice of methodologists to avoid over-controlling just for the sake of it (Achen Reference Achen2005; Clarke Reference Clarke2005, Reference Clarke2009) and instead follow a theorised data-generating process (Hünermund, Louw, and Caspi Reference Hünermund, Louw and Caspi2023). In this paper, over-control would at best remove some of the effect, and at worst introduce collider bias (Elwert and Winship Reference Elwert and Winship2014).

Second, I reweight the observations in my dataset based on the degree of urbanisation and the Moran eigenvectors using entropy balancing (Hainmueller Reference Hainmueller2012). Entropy balancing is a matching algorithm that calibrates the unit weights for the control group so that the specified sample moments of the covariate distribution (e.g. mean, variance, skewness) of the reweighted control group approximate those of the treatment group (Watson and Elliot Reference Watson and Elliot2016). This algorithm is doubly robust to linear regression, meaning that the causal effect of WWI death rates on contemporary support for the far-right is well identified if either one of the matching or regression models I use is correctly specified, even if the other is not (Zhao and Percival Reference Zhao and Percival2016).

To further reduce the threat of selection bias by restricting identifying variation to a more plausible subset of within-period and within-region variation, I add two sets of fixed effects (Mummolo and Peterson Reference Mummolo and Peterson2018). First, year-fixed effects, which I include because of election-specific political dynamics that should not be pooled during estimation. Second, grid-cell fixed effects, which, following the approach of Doucette (Reference Doucette2024), divide observations into spatial units along latitude and longitude and capture unobserved confounding related to the natural distribution of French departments in physical space. Grid-cell fixed effects divide the French map into dozens of rectangles of equal size, capturing all the structural elements specific to each region that have remained stable over time, including geography and baseline population characteristics. I avoid including administrative unit fixed effects, such as those based on NUTS2 units, because these were created after WWI based on decisions that are credibly related to the post-war economic situation, and could lead to over-control for post-treatment variables, a major threat in historical political economy (Homola, Pereira, and Tavits Reference Homola, Pereira and Tavits2024; Pepinsky, Goodman, and Ziller Reference Pepinsky, Goodman and Ziller2024a, Reference Pepinsky, Wallace Goodman and Ziller2024b). In the online Appendix, I show nevertheless that replacing grid-cell fixed effects with NUTS2 regional effects does not alter the results.

I report Conley standard errors, which are heteroskedastic and two-dimensional autocorrelation consistent (Conley and Molinari Reference Conley and Molinari2007)Footnote ¹⁰ .