The Burden of a Violent Past: Formative Experiences of Repression and Support for Secession in Catalonia

Abstract This letter studies the impact of past violence and repression on current territorial preferences in a contemporary democracy. Does a violent past lay the grounds for pro-secessionist preferences, or does it lead individuals to cling on to the territorial status quo? We study whether exposure to the events of the Spanish Civil War and its immediate aftermath made people more or less likely to support Catalan secession from Spain. Our analysis employs a dataset that combines a large N of individual-level survey data with historical data about repression and violence in each Catalan municipality. Findings indicate that current preferences for secession tend to diminish among the oldest Catalan generation that was exposed to higher levels of violence in their municipality. Most crucially, we show that exposure to violence created a sense of apathy towards politics among the oldest cohort, which eventually leads to a lower predisposition to support secession, a feeling that was not transmitted to subsequent generations. Our findings qualify some of the existing knowledge on the effects of past political violence on present political attitudes.

hunger' is time varying and, in some circumstances, leads Ukrainians to behave more loyally towards Moscow. In a cross-national observational study, Torcal and Montero (2006) show that people living in countries with a recent authoritarian past, such as Spain, Portugal or Greece, are more likely to abstain from voting. Finally, and despite finding an anti-regime-bias logic in voting, Balcells (2012) concludes that younger generations are more influenced by past violent experiences than older generations, and that self-reported victimization does not have an effect on identities articulated around the centre-periphery cleavage.
As the previous summary illustrates, the question is far from settled. Thus, most previous research has focused on the long-term effect of violence and repression on people's political behaviour, mainly vote choice but also abstention (Lupu and Peisakhin 2017). Others look at attitudes, with a special focus on trust (Valencia and Tur-Prats 2020). Yet, previous works have largely overlooked an important outcome, namely, people's support for or rejection of the territorial status quo. Being more left-wing or right-wing, or having more or less trust, though undoubtedly important, might constitute a mild sign of rejection (or of bias). However, the willingness to change the territorial status quo is a clear indication of dissidence and can even constitute a better proxy of people's support for the current political system.
In this research note, we precisely complement this vivid debate by examining to what extentand in which directionviolence and repression experienced during the Spanish Civil War  and its immediate aftermath left an imprint on the Catalans' willingness to support the creation of a new state. We examine whether the effect of a violent past on current secessionist preferences, if any, is circumscribed to the old cohort or has travelled over time. In addition, we explore whether exposure to violence leads people to avoid involvement in new conflicts or a sense of political apathy that may trigger a lower willingness to support independence.
The coup d'état against the Spanish Second Republic  was designed, among other things, as a reaction against conceding power to the Catalans. As historians have documented (Solé i Sabaté and Villarroya i Font 1989), the centre-periphery cleavage was a salient topic before and during the war, and has remained salient ever sinceat varying degrees and together with the left-right divide. For instance, one of the first measures implemented by the Francoist army after occupying Catalonia was to ban the Catalan language and impose a unitary idea of the Spanish nationas illustrated in the indoctrination of children on Spanish nationalism in school curricula. If this created a reaction along the lines of the anti-regime-bias expectation, then we should observe that cohorts who were exposed to more violence during their formative years should have been more likely to embrace the secessionist claim.
Yet, one might need to consider that while repression from the Francoist regime was more intense, violence also came from the Republican side. Thus, on the one hand, it can be argued that violence from both the Francoists and the Republican supporters created the feeling, among cohorts exposed to it, that the situation was irreconcilable and that an alternative solution to the current state of affairs was Catalan independence. On the other hand, it can also be sustained that cohorts exposed to violence from both sides were more likely to develop a feeling of apathy or disaffection towards politics (Torcal and Montero 2006), which eventually made them less likely to support the secessionist claim. Experiencing violence and repression at first hand could have fostered feelings based on the awareness of the risks that political conflicts may entail. It might have increased the perception that a radical political confrontation and polarization can break social harmony and represent a risk for personal integrity. Supporting secession from Spain implies favouring a rupture of the status quo. Even within the framework of a democratic regime, it may recall the sort of hazardous political stances that cohorts with formative memories of political violence fear. Yet, to add to the mixed theoretical expectations, Spain is often portrayed as an example of a country that was capable of following a successful democratic transition and hence able to heal the wounds of the past (Higley and Gunther 1992). If this is true, we should not observe any difference in support for secession across different cohorts, as the scars of the conflict should have been healed.
Methodologically, we test each of the previous expectations, utilizing fixed effects (FE) and random effects (RE) models applied to a large-N dataset that combines individual-and municipality-level data from Catalonia. The individual-level data come from repeated crosssectional surveys that cover 27 years . The municipality-level information has been obtained from historical records about repression and violence during and after the Spanish Civil War. Our main strategy consists of considering this combined multilevel dataset as a single cross-section in order to attain a representative sample of individuals within municipalities for Catalonia. Then, we perform multilevel random-effects models and specify cross-level interactions between cohort groups at the individual level and the degree of violence and repression at the municipality level, while including control variables at both levels to exclude potential observed confounders. We additionally rely on a fixed-effects dummy-variable approach and an extension to deal with cross-level interactions that adjusts for contextual unobserved heterogeneity (Giesselmann and Schmidt-Catran 2020), as well as on the more comprehensive hybrid framework of the 'within and between random effects' (REWB) models (Bell, Fairbrother and Jones 2019).
We find that cohorts that lived in contexts where repression and violence were more intense during their formative period are also less likely to support Catalan independence today. Therefore, our results show that when it comes to secessionist attitudes, the anti-regime-bias logic is not supported. In the second part of this letter, we explore why this is the case, finding support for two different logics: first, we show that older cohorts were more likely to develop a feeling of political apathy and fatalism, which might have alienated them from a high-stakes issue like Catalan independence; and, secondly, findings point to the idea that the bias against independence among older cohorts is circumscribed to them and was not transmitted to other generations.

Research Design
Our empirical analysis exploits a pooled dataset of repeated cross-sectional surveys that spans over 27 years . 1 All samples are representative of the Catalan voting-age population and have wide geographic coveragemunicipalities included in the pooled sample cover 96 per cent of the Catalan population. Our main outcome is the question: 'With regard to the Spanish state, do you think Catalonia should be…?' Respondents could choose one of the following options: 'a region', 'an autonomous community', 'a state within a federal Spain' or 'an independent state'. We created a binary indicator that captures whether the respondent wants Catalonia to become an independent state versus the rest. Although this question aims at measuring a respondent's preferred territorial option and only indirectly captures their position on an independent Catalonia, it has been consistently asked over a long period, and, most crucially, it correlates strongly with other direct questions (Guinjoan and Rodon 2016).
Our main explanatory variable is the percentage of people repressed during and after the Civil War by the Francoist regime (up until 1950) minus the percentage of people executed by the Left during the Civil War. The former comes from Solé i Sabaté and Villarroya i Font (1989); the latter comes from 'The list of juridical repair of victims of Franco's regime', published in 2015 by the Catalan government (Generalitat de Catalunya 2017). Percentages are based on the 1936 population as compiled in the census right before the Civil War. Both datasets were collected by historians, who thoroughly gathered information from several archives, cross-checking the data from different sources. Left-wing repression only took place during the Civil War, while Francoist repression continued after the conflict (mainly up until 1949). While left-wing repression essentially captures the number of killings during the conflict, right-wing repression includes the number of executions, people imprisoned during and after the war (up until 1945) and other 1 The Online Appendix includes the summary statistics, sources, additional analyses and several robustness checks.
repressive measures (that is, sanctions). This implies that the indicator is mostly positive and hence right-skewed. All in all, this indicator allows us to distinguish between both types of repression, as they are qualitatively different and might generate different effects (Balcells 2012). Francoist repression was perpetrated against the other ideological side (the Left), but it was also justified in stopping any national dissidence (Catalan identity or others) and Catalan secession. In contrast, left-wing repression mainly targeted wealthy, religious and conservative individuals in a pattern largely unrelated to their anti-or pro-secession stances. By taking the difference between both indicators, we are taking into consideration both dynamics and that one type of violence can potentially cancel out the other. For instance, if a municipality only experienced Francoist repression, the effect on secessionist preferences could arguably be higher compared to a municipality that experienced both Francoist repression and left-wing violence.
We focus on studying the generation likely to have experienced the violent events at first hand (those turning 18 years old between 1917 and 1949 in our sample) and consider the rest of the age groups altogether. 2 Our approach implies a within-cohort analysis, that is, a comparison between senior citizens with formative experiences dating back to the Civil War and the first years of the Francoist Regime who experienced dire violence and repression in their municipalities, and equivalent elders who were less exposed to such dreadful events in their villages. We also perform a between-cohort comparison, looking at the difference in the effect of violence at the municipal level in the oldest generation compared to its impact in the rest of the cohort groups. The latter allows us to detect an eventual intergenerational transmission of the effects of violence. Our modelling strategy mainly relies on RE models where individuals are nested within municipalities, as well as on FE models with municipality dummies. We hypothesize that the effect of past violence on current political attitudes is channelled through the cohort's formative experiences of being closely exposed to that violence. 3 We performed two types of RE models. In our first RE approach, we implement two-level models where individuals are nested within municipalities. In doing so, we consider our dataset as one large cross-sectional sample from a single time point. By pooling different datasets, we can obtain a good sample for most Catalan municipalities. 4 It is also important to bear in mind that our period of analysis crucially captures changes in the saliency of the independence debate by covering several contemporary contextual circumstances, which implies various governments, both at the regional and the national levels, and different economic periods.
In these RE models, we include several important controls that adjust for potential observed confounders, helping to rule out that other observed factors related to the individual characteristics of respondents or the composition of municipalities confound the relationship between violence and support for secession through the formative experiences of older citizens. At the individual level, we include a respondent's Left-Right ideology, as Catalonia's political 2 Separate results for each cohort are also provided in Sections E.F of the Online Appendix, confirming that the effect of violence is circumscribed to the oldest generation. 3 With such a design, it is not necessary to deal with the age-period-cohort identification problem as we do not need to identify each of these separate effects. We are interested in estimating not current period effects (from 1991 to 2016), but the period effects of the past in the form of the formative experiences of the oldest cohort. We neither need to account for age effects, as they are adjusted for when comparing individuals from the oldest cohort with different degrees of exposure to violence. Yet, we provide an age-period-cohort analysis as a robustness check in Section E.C of the Online Appendix. 4 Our data can be viewed as neither a panel of individuals, nor a panel of municipalities. We are able to reach an adequate representative sample of individuals across municipalities thanks to pooling the repeated cross-section surveys over time (following the procedure also employed in Rodon and Guinjoan [2018]). With such a design, we attain a satisfactory crosssection dataset at two levels. In this case, we control for time by dividing the sample into two periods in the main models, as well as including year dummies or defining a third level for year in age-period-cohort (APC) RE models (see Sections E.C and E.F in the Online Appendix). Furthermore, current period effects are not of much interest here since we are investigating a period effect of the long distant past that does not change in our dataset over time. It can be considered constant, as linked to the formative experiences of the oldest generation. competition follows a bidimensional structure (Left-Right and pro-/anti-independence dimensions) and the weight of the repressive past might be different between individuals ideologically aligned with the Francoist dictatorship and the rest (Galais and Serrano 2020). As a proxy for an individual's political identity, an important factor shaping political attitudes and vote choice in the Catalan case (Rodon and Guinjoan 2018;Serrano 2013), we control for the language a respondent speaks at home.
We also include several indicators at the municipality level that help us tackle different contextual circumstances related to independence support (Rodon and Guinjoan 2018). We incorporate: (1) the percentage of people born in Catalonia; (2) the percentage of older people (64 years old or more); (3) the percentage of men; (4) the population of the municipality; and (5) the electoral district (province) a respondent belongs to. Finally, all models are based on individuals that were born in Catalonia. We do not know where respondents born outside Catalonia lived before; therefore, it is not possible to capture the intensity of repression (if any) that they were exposed to when they were young. 5 In our second RE approach, we explicitly include the time dimension in the analysis. We perform a REWB model, a hybrid approach explained by Bell, Fairbrother and Jones (2019; see also Fairbrother 2014; Tormos 2019). In this three-level model, individuals are nested within municipality-year units, which, in turn, are considered observations of each municipality (a panel of municipalities). This hybrid modelling strategy allows estimating the effects of both time-invariant and time-varying contextual covariates. It combines the between-effects estimator of RE models with the within-effects estimator of FE models.
In addition, we also run FE models to control for unobserved heterogeneity at the context level by specifying municipality dummies. We extend this approach in two different ways. On the one hand, we perform further models that include year dummies in addition to municipality dummies. On the other hand, we follow Giesselmann and Schmidt-Catran's (2020) advice to properly specify cross-level interactions in FE models by including dummies of the interaction of the group variable with the variables of the interaction. In our case, this means including dummies of the interaction of municipalities with municipal violence and with an individual's generation. This type of model is called 'country fixed effects and slopes' (cFES) and it is aimed at controlling for heterogeneity in the two variables interacted. Table 1 shows the results of several logistic regressions with the aforementioned specifications. Models 1 to 5 are RE models with individuals nested within municipalities while considering the whole dataset as a large single cross-section. Models 6 and 7 are hybrid REWB models with individuals nested within municipality-year units and then within municipalities. Finally, Models 8 to 10 are FE models with municipality dummies (8), municipality and year dummies (9), and a cFES model for the slopes of the interacted variables (10). Most of the models only include the independent variables of interest (1, 2, 6, 8 and 9), while the remaining ones also add controls at both the individual and the context levels (3, 4, 5 and 7). The coefficients of interest come from the interaction between the cohort dummy (old versus the rest) and the municipality-level repression indicator (the difference in violence/repression by the two ideological sides). The RE models are specified with random slopes for the lower-level variable 5 Sections E.A and F.F in the Online Appendix discuss the null influence of an individual's language and their Left-Right position, which can be considered post-treatment, on our estimates. Section C in the Online Appendix deals with population movements and selective sorting. Sections E and F in the Online Appendix include, respectively, a control for time that divides the sample into two periods in the main models, and year dummies or defining a third level for year in APC RE models. Notes: Effects of the difference in the percentage of the population repressed by both sides in each municipality during and after the Civil War in the formative experiences of senior citizens. * p < 0.1; ** p < 0.05; *** p < 0.01. involved in the cross-level interaction (cohort), as advised by Heisig and Schaffer (2019) to obtain robust estimates. 6 In all the models in Table 1, the interaction is statistically significant, negative and similar in strength across specifications. 7 This means that support for secession is lower among the old cohort that lives in places that experienced a relatively high level of repression by the Francoist regime during and after the Civil War. The rest of the cohorts who could not have experienced such violence directly do not seem to have become affected. Interestingly, the old cohort is not per se less pro-independence when the cross-level interaction is specified (see Model 2). Old cohorts are only less prone to independence if they live in municipalities where repression by the Francoist dictatorship was more intense. As Model 4 shows, the effect is particularly pronounced when we restrict the sample to the 1991-99 period. During these years, when secession was not very popular and the older cohort was still demographically relevant, Catalans who were exposed to the Civil War and subsequent repressive period and lived in a more repressed locality were particularly anti-secession. 8 Figure 1 graphically portrays the cross-level interaction of the RE models. In Figure D1 in Section D of the Online Appendix, we show that the magnitude of the interaction effects across RE and FE specifications is similar.

Results
We performed an equivalent analysis but using Left-Right ideology as a dependent variable (see Section E.A in the Online Appendix). In contrast to studies examining vote choice (Balcells 2012;Villamil 2020), our results indicate that being exposed to higher levels of violence did not affect an individual's Left-Right position, the other relevant dimension of competition in Catalonia.
Finally, in additional models, we examine the effect of the percentage of people that experienced violence/repression at the municipality level, regardless of the perpetrators, and we unpack the effect of both types (see Section E.B in the Online Appendix). Results show that total violence is associated with a lower support for independence among cohorts that experienced it, essentially because these were places where the Francoist repression was more intense. In contrast, the effect of left-wing violence on secessionist support, adjusted for the intensity of Francoist repression, does not have an effect on secessionist attitudes.

The Persistence of the Gap
Our empirical analysis shows that cohorts exposed to violence during their formative years are less likely to embrace the secessionist project. Findings, therefore, go in a different direction than the anti-regime-bias logic. When it comes to being opposed to or in favour of the creation of a new state, our results show that people exposed to more violence are not more likely to develop a bias against one of the ideological pillars of the perpetrators of violence. Why do we observe such effects? We explore two different mechanisms.
First, our intuition, aligned with previous work (Torcal and Montero, 2006), is that people who were exposed to a higher intensity of violence developed an anti-politics feeling, which also made them more likely to take a stance against secessionism. The political issue of Catalan independence has the capacity to polarize society, and this might not sit well among the old cohorts that 6 An alternative modelling strategy (see Section F.E in the Online Appendix) uses a linear probability model with pooled individual-and aggregate-level variables together and clustered standard errors at the aggregate level of municipalities (Wang, 2021). 7 The exception is Model 5, run on the subsample of the period 2000-16. During this period, the majority of individuals in the oldest cohort had died. The remaining individuals might be unrepresentative of the original cohort due to possible biases in its composition (e.g., the differential survival of those well-off). 8 In order to avoid an eventual model dependency of our analysis, we replicated the interactions using linear RE models instead of logistic ones on support for secession (see Section F.E in the Online Appendix), as well as with territorial preferences as the dependent variable (an ordinal scale with four response options) separately (see Section F.C in the Online Appendix).
were exposed to violent events. Attitudes against politics can be expressed through a variety of forms, including apathy, apoliticism, cynicism and distrust towards politics. If the effects of violence operate through the development of anti-politics feelings, we would spot a spread of such political attitudes among the older generation that was exposed to violence in their municipality during their formative years. To that end, we exploit a series of indicators measuring this type of attitude, as well as feelings of political efficacy, using agree-disagree scales that have been consistently included in the dataset. Figure 2 shows that there are indeed differences in two of those indicators. Senior citizens in municipalities that experienced higher levels of violence are more likely to believe that 'elections are not really useful because the same people always rule' than the same cohort group in municipalities with less violencea feeling of fatalism towards the outcome of democracy. The cohort group exposed to higher violence in their formative years is also more inclined to think that 'it is better not to get involved in politics'an indication of political apathy. These people may connect their traumatic experiences of violence during and after the Spanish Civil War to the political polarization and ideological confrontation of those times. As a result, they might think that getting involved in politics and taking sides can be dangerous, even in peaceful times. All in all, dictator's Franco cynical saying, 'Do like me, don't get involved in politics', constitutes a metaphor of how this cohort may feel towards the independence project.
Secondly, part of the transmission of traumatic memory hinges upon the idea that new cohorts take up the values of older ones. If the relationship between violence and support for secession is circumscribed to older cohorts, and intergenerational transmission does not occur, the automatic process of generational replacement would progressively dilute the attitudinal legacy of the past (Tormos 2019). Was the intergenerational transmission of the memory of violence more likely to happen in places that experienced more of this violence? The absence of such a main contextual effect of violence in support for secession in most models of Table 1 (Models 2-7 and 10) constitutes an indication that memories live as long as those who hold them. There are no signs of intergenerational transmission or attitudinal transmission related to the municipal level. The significance and shape of the interaction indicate that younger cohorts who were not directly exposed to violence, but lived in municipalities that experienced it, do not become affected. 9 Perhaps the lack of such contextual effects is due to an overt will to forget the past. The Francoist dictatorship silenced those memories. Forgetfulness might have endured well into the present. Explicit policies of collective memory were implemented in Catalonia just a decade ago.

Conclusions
Findings in this letter stand in contrast with the most recent empirical evidence on the long-term effects of violence and repression on political attitudes and political behaviour. Contrary to the anti-regime-bias hypothesis, which would have predicted higher support for independence among the generation that experienced more violence perpetrated by the Francoist regime, but in line with other recent works (Wang 2021;Zhukov and Talibova 2018), we show that cohorts  More detailed results where the rest of the cohorts are decomposed into ten-year groups yield equivalent results, as can be seen in Section E.D in the Online Appendix. Only the oldest cohort with formative experiences potentially linked to violence is affected. exposed to violence during their formative years are less likely to embrace the secessionist project today. In line with the disaffection argument sustained by some previous works (Torcal and Montero 2006;Zhukov and Talibova 2018), we find that the oldest cohort was more likely to develop a feeling of apathy towards politics and that the generational transmission of the antiindependence feeling did not occur. Our results are also in line with previous works showing the heterogeneous effects of violence (Rozenas and Zhukov 2019) and how certain conditions mediate the long-term effects of violence on behaviour and attitudes (Villamil 2020).
A first potential explanation of these different findings is that polarization in the pro-or antiindependence debate activates the memory differently than polarization along the left-right dimension. It might be that supporting secession implies a more disruptive event than conflict over Left and Right, and the former makes cohorts with formative memories of political violence more concerned (or even afraid) about the outcome than the latter. Another explanation may be that the long-term effect of violence travels differently over time in democracies or even in contexts with two dimensions of political competition.
Overall, we show that cohorts exposed to violence are less likely to support a disruptive event, such as the independence project, or, in other words, they develop a bias in favour of the status quo, in this case, being supportive of territorial stability. Thus, our findings qualify those of Balcells (2012, 311), who concluded that 'civil war victimization experiences did not have a major influence on identities articulated around the centre-periphery cleavage'. In addition, results point to the lack of intergenerational transmission, while other works do (Lupu and Peisakhin 2017). One reason may be that, compared to other studies, certain political events in Spain, such as the transition to democracy or economic growth, might have broken the transmission of these values.
All in all, this letter helps us understand how the burden of the past travels to the present day and, especially, shows that violence does not always create an attitudinal 'bonus' against the perpetrators. On some occasions, like the one we have analysed here, it entails a negative burden that the cohort that experienced it is not able to overcome.