Collective Remembrance and Private Choice: German-Greek Conflict and Consumer Behavior in Times of Crisis∗

Vasiliki Fouka† Hans-Joachim Voth‡

February 2020

Abstract

How is collective memory formed and when does it impact behavior? During World War II, German troops occupying Greece carried out numerous massacres, often as reprisals for partisan attacks. During the 2009 Greek sovereign debt crisis, political conflict erupted between the German and Greek governments; German car sales in Greece declined. Re- ductions were greatest in areas affected by German reprisals. Differential economic perfor- mance did not drive this divergence. Instead, survey evidence and an IV-strategy predict- ing the geography of German attacks suggest that current events reactivated past memo- ries, creating a backlash against German products in reprisal areas. Using quasi-random variation in official public recognition of a town’s victim status, we show that the insti- tutionalization of collective memory amplifies effects of time-varying political conflict on consumer behavior.

Keywords: collective memory, political conflict, consumer behavior. JEL Codes: D12, D74, F14, N14, N44.

∗For helpful suggestions we thank Alexander Apostolides, Leo Bursztyn, Bruno Caprettini, Ray Fisman, Nicola Gennaioli, Luigi Guiso, Yannis Ioannides, Emir Kamenica, Tim Leunig, Sara Lowes, John Marshall, Guy Michaels, Stelios Michalopoulos, Nathan Nunn, Sonal Pandya, Elias Papaioannou, Luigi Pascali, Giacomo Ponzetto, Dominic Rohner, Guido Tabellini, and Nico Voigtlaender. Seminar participants at Bocconi, CREI, Ente Einaudi, Harvard, UPF, Columbia, LSE, the Workshop on Political Conflict in Washington University in St. Louis, the 2017 Zurich Conference on Origins and Consequences of Group Identities, the EREH-London conference, and the 12th Conference on Research on Economic Theory and Econometrics in Naxos provided useful advice. The Hellenic Statistical Authority kindly provided car registration data. Evangelia Mathopoulou provided excellent research assistance. †Stanford University. Email: [email protected]. ‡University of Zurich. Email: [email protected].

1 1 Introduction

Every human being remembers events. In contrast, collective memory is not based on per- sonal recollections. The importance of distant events is typically passed from generation to generation through acts of public remembrance and the teaching of history, using symbolic “sites of memory”.1 Collective memory functions as “mythical glue” (Harari, 2014), help- ing humans to collaborate in large groups of genetically unrelated individuals by becoming members of “imagined communities” (Anderson, 1983). While collective memory is a constant of human society, it is unclear how collective memories are created, when they matter, and whether they can affect major economic de- cisions. Dessi (2008) presents a model of how collective memory binds nations together, emphasizing benefits for economic co-operation. Benabou´ and Tirole (2011) study the in- teraction between memory and identity, arguing that memories allow people to infer who they are and what to do in a given situation. Despite theory highlighting the importance of collective memory for individual and collective outcomes, there is no compelling empirical evidence that changes in economic behavior can be driven by collective memories, or how these memories come to matter. In this paper, we examine the creation and re-activation of collective memory and its effects in a high-stakes setting. During Greece’s military occupation by Germany in World War II, German armed forces committed numerous war crimes, including mass executions and the complete destruction of entire villages. This violence was typically carried out in retaliation for local partisan attacks (Mazower, 1995). Decades later, during the sovereign debt crisis of 2009–2014, political relations between the German and Greek government once again turned acrimonious. Under EU and German pressure, Greece had to implement stringent austerity measures. German newspapers were quick to blame “lazy Southern- ers” for the Euro debt crisis. As public discord erupted between the German and Greek governments, memories of Germany’s violent occupation of Greece during World War II resurfaced: Greek demonstrators waved placards of Angela Merkel in Nazi uniform, and consumer groups called for a boycott of German products. To examine how collective memory interacted with contemporary political conflict to influence major economic decisions, we focus on car sales. Cars are “big ticket items”, rep- resenting a major expense for consumers. They are also an iconic German product. We analyze variation in car sales over time and across prefectures in Greece and ask: Did Ger- man car sales during the debt crisis decline more in areas that suffered German reprisals during World War II – and especially in those that received official recognition as “mar- tyred” towns? Using original sources and secondary historiography, we construct a dataset of the uni-

1“lieux de m´emoire”, in the parlance of French historian Pierre Nora (1989). Examples include official holidays, mythical founders, monuments, sites of heroic struggle (for example, Masada) and symbolic figures such as Uncle Sam in the US or Marianne in France (Akerlof and Kranton, 2011).

2 verse of Greek towns that experienced German reprisals during World War II. We collect information on the level of destruction experienced by these towns from government docu- ments, and combine this information with data on new car registrations. Using a machine learning algorithm analyzing the online archive of Greece’s largest newspaper, we compile a time-varying measure of political conflict between Greece and Germany. In addition, we conduct a survey collecting individual-level data from a sample of over 900 households in towns that experienced reprisals, and a set of control locations. Figure 1 displays our main result. The dotted line tracks the level of animosity between Germany and Greece based on our newspaper-based measure. As the debt crisis wors- ened, conflict surged, with months of peak conflict in early 2010 and the summer of 2012. The solid line shows the difference in German market share between prefectures with and without reprisals. From the two time series, it is evident that heightened public conflict be- tween Germany and Greece coincided with lower German market shares in high- relative to low-suffering prefectures. Panel regressions confirm the pattern: the greater the share of reprisal towns in a prefecture, the greater the decline in market share for German producers in times of German-Greek conflict. Two factors may confound our results: the differential time-varying impact of the crisis, and the persistent effect of cultural characteristics of locations. To address the first concern, we show that our findings are robust to accounting for a potential differential effect of the crisis in reprisal-affected prefectures. We control for prefecture-level baseline economic controls interacted with our time-variant measure of German-Greek conflict, as well as for a country-level measure of monthly unemployment interacted with our prefecture-level measure of German atrocities. In addition to standard economic controls, we also use prefecture-level changes in nighttime luminosity, to proxy for GDP changes at the regional level. Results are robust to including these variables, as well as a measure of economic uncertainty for Greece (based on the method in Baker, Bloom, and Davis (2016)), which we allow to have differential effects across prefectures. Our results are also robust to excluding luxury cars from the analysis. Survey data provides additional evidence that results are not driven by a differential downturn in economic conditions. Residents in reprisal towns do not only buy fewer Ger- man cars – they also name German brands less often as their “ideal” car. In other words, Greeks in martyred towns do not buy fewer German cars because they cannot afford them; they fail to do so because they like them less. Strikingly, results are stronger for Greeks born in massacred towns; and they are strongest for those whose parents and grandparents already lived in the same location, making it more likely that family members were affected by German war crimes. Since the descendants of long-time inhabitants share the same eco- nomic environment as more recent arrivals, but dislike German cars more, it is likely that the past is shaping their attitudes – and not recent economic shocks. Also, in line with an interpretation of a differential backlash against German products, Greeks in reprisal towns view German policies during the crisis more negatively – they are more likely to blame Germany for the Greek sovereign debt crisis.

3 We next address the concern that German atrocities could simply reflect more deeply- rooted differences between Greek towns. For example, towns and villages with strong nationalist sentiment may have resisted the WWII occupation more vigorously – leading to reprisals. If this nationalist sentiment persisted after 1945, it could then map into a greater backlash against German products today. Such underlying characteristics – and not the memory of reprisals – could then drive differential reactions during the debt crisis. We provide three pieces of evidence against this argument. First, we analyze the geography of Greek armed resistance to isolate plausibly exogenous variation in the location of reprisals. Our instrumental variable strategy relies on the fact that resupply points are likely to be located in flat areas within rugged surroundings. Such locations allowed Allied aircraft to drop supplies, while still providing sufficient protection to partisans who picked them up. Terrain characteristics predict resupply, partisan activity, and the location of German massacres. When used as an instrument, the variation in reprisals driven by geographical features strongly predicts a backlash against German cars. Second, we conduct falsification tests using atrocities committed by Italians and Bul- garians – the other two major occupying forces of Greek territory during WWII. We find no differential adverse effect of German-Greek conflict for those locations. This rules out long-term persistence of nationalism as an alternative interpretation – the possibility that locations that resisted more in general, and were hence attacked by any occupier, also show more adverse reactions to outside pressure some 70 years later. Third, in our survey, we find that Greeks in towns affected by reprisals are not more nationalistic or more prone to activism than those in control towns. We next explore in more depth how the past comes to matter for the present. Since the 1990s, Greece awards communities “martyred” town status if they suffered severely during the occupation – a purely honorific designation with no material benefits. Applications of towns for martyr status are assessed by a committee composed of academic historians. Government data on town-level destruction combined with information from the commit- tee minutes allow us to identify an exogenous component in the assignment of martyr status. Higher wartime destruction in general led to higher acceptance rates, but with a pronounced discontinuity at around 50% of wartime destruction. This suggests that the committee followed a behavioral heuristic for deciding status assignment. Controlling for average levels of destruction, prefectures with a higher share of towns that received official martyr status saw market share declines for German cars above and beyond those expe- rienced by prefectures with a higher incidence of reprisals. We find a similar result after isolating the exogenous component of martyr status, as predicted by the discontinuity at 50% of wartime destruction. This finding indicates that public recognition of victim status can amplify collective memory. In support of this interpretation, based on a telephone survey of mayors’ offices in towns affected by reprisals, we show that places that continue to collectively commemorate their suffering during 1941-44 to the present day saw sharp declines in German car sales in periods of heightened disagreements between the German and Greek governments. These

4 effects are smaller in magnitude than effects for towns that, in addition, also received official recognition of their victim status. This suggests that responses to time-varying conflict are stronger the more institutionalized collective memory is. Taken together, our findings show that collective and highly symbolic commemoration of the past can amplify adverse reactions to modern-day conflict, resulting in sharp changes of consumer behavior. Our paper relates to research in several areas. Conflict typically has negative effects on economic and social development (Collier et al., 2003). At the same time, it can enhance co- operation among survivors (Bauer et al., 2016) and leads to more complex societies (Collier, 2016). Countries with a longer history of military conflict typically levy more taxes, and have higher per capita GDP (Besley and Persson, 2010; Dincecco and Prado, 2012). Guiso, Sapienza, and Zingales (2009) show that a history of conflict predicts patterns of trade. Our study demonstrates that the effect of past conflict can be time-varying, and that the memory of conflict interacts with contemporary events to shape economic behavior. Similar in spirit to our study, Belmonte and Rochlitz (2019) and Ochsner and Roesel (2017) show that the salience of current events interacts with past history to affect political behavior. Several papers investigate the effects of consumer boycotts, in response to grievances against producer countries (Koku, Akhigbe, and Springer, 1997; Teoh, Welch, and Waz- zan, 1999; Ashenfelter, Ciccarella, and Shatz, 2007; Michaels and Zhi, 2010; Hong et al., 2011; Pandya and Venkatesan, 2016), with mixed conclusions on their efficacy. We study instead uncoordinated actions by individual buyers that collectively have a major impact on economic activity. In related work, Fisman, Hamao, and Yongxiang (2014) examine a diplomatic incident between China and Japan caused by how Japanese textbooks treated the 1930s invasion of China. They argue that inter-government conflict led to declining trade due to changes in the purchasing behavior of government-owned firms. A growing literature has shown that behavioral biases can have substantial effects on behavior (Kahneman and Tversky, 1982; Gabaix, Laibson et al., 2006; Schwartzstein, 2014; Mullainathan, 2002). When news arrives, models of limited attention typically predict under-reaction, and there is substantial empirical support for this prediction. Investors, for example, often react slowly to earnings news released on Fridays or on days when many other companies are reporting (DellaVigna and Pollet, 2009; Hirshleifer, Lim, and Teoh, 2009); similarly, consumers change their behavior only gradually in response to the rankings of hospitals and colleges (Pope, 2009). A closely related literature on memory limitations leads to the opposite prediction – namely, that news can effect sharp changes in behavior if it “brings back” memories of past events (Mullainathan, 2002).2 “Associative- ness” – similarities between the current situation and the past – can encourage recall, as several results in experimental psychology suggest (Serman and Kim, 2002). At the same

2Salience-based models such as those offered by Gennaioli and Shleifer (2010) and Bordalo, Gennaioli, and Shleifer (2015) combine limited attention with limited memory; these models can accommodate both under- and over-reactions.

5 time, there is as yet no paper that shows that memory effects can lead to sharp discontinu- ities in economic decision-making. This paper makes three main contributions. First, we provide evidence that location- specific historical experiences of conflict can be reactivated by current events and affect important consumer decisions. Second, we demonstrate the importance of collective mem- ory – through official recognition and rituals of collective remembrance, an adverse shock translates more sharply into attitudes that matter, especially when sanctioned from above. Our paper is one of the first to show that memory effects can indeed lead to abrupt, short- term changes. Third, we show that historical events do not have a linear effect on present attitudes and behaviors. There is a burgeoning literature documenting long-term persis- tence (Guiso, Sapienza, and Zingales, 2016; Becker et al., 2014; Voigtlander¨ and Voth, 2012). We show instead that past experiences have a time-varying effect, as they interact with present conditions. The rest of the paper is organized as follows. In the next section we review the historical background of German reprisals in Greece and of German-Greek conflict during the debt crisis. Section 3 describes our data. Sections 4 and 5 present our main results, address the main concerns with identification, and list additional robustness checks. We examine the creation and effects of collective memory through official recognition in Section 6, before concluding in Section 7.

2 Historical Background

2.1 Germany’s Wartime Occupation of Greece

In May 1941, Axis forces occupied Greece. The country was divided into three occupation zones, of which the largest was administered by Italy. Germany occupied less territory but controlled key locations including Athens, Thessaloniki, and Crete. Bulgaria was in charge of a relatively small part of the country close to its own borders. From the beginning, the civilian population suffered from expropriations and plunder. The German armed forces requisitioned foodstuffs, causing a major famine during the winter of 1941–1942, leading to an estimated 300,000 deaths (Hionidou, 2006). Reprisals against potentially uninvolved civilians in areas of armed resistance were first authorized by the German army in April 1941 in Yugoslavia (Mazower, 1995). The High Command of the German Armed Forces (OKW-Oberkommando der Wehrmacht) laid down quotas for reprisal killings: 100 civilians were to be shot for each German soldier killed in a partisan attack, 50 for each soldier wounded, and so forth. Such reprisals against civilians became standard practice of German anti-partisan operations in the Balkans and were later extensively used throughout Eastern Europe.

6 Crete saw the first German reprisals on Greek soil (Nessou, 2009).3 Partisan attacks were often followed by indiscriminate massacres of the civilian population and the destruction of every village near an attack. By 1944, an estimated 2,000-3,000 Greek civilians had been executed by German armed forces on Crete alone (Nessou, 2009). After Italy’s surrender in 1943, German forces occupied the zone held by its former ally. Fighting between guerrilla groups (andartes; mostly the Communist-led ELAS) and the Wehrmacht intensified. For example, during an anti-partisan sweep by the 117th Jager¨ Division in the mountains near Kalavryta (Peloponnese), Greek resistance fighters executed 78 German prisoners. In retaliation, German troops killed 1,200 of Kalavryta’s inhabitants, including women and children, on December 13, 1943 (Mazower, 1995). Some 28 towns and villages in the area were destroyed. Similarly savage reprisals occurred all over Greece, including in the famous cases of Doxato, Kommeno, and Distomo. After the war, the Greek Ministry of Reconstruction estimated that some 30,000 Greeks perished in reprisal attacks by German forces, with numerous villages and towns left destroyed (Doxiadis, 1947).4

2.2 German-Greek Conflict during the Eurozone Crisis

The Greek sovereign debt crisis began in late 2009, when revised budget deficit figures revealed the country’s dire financial situation. This discovery led to successive downgrades of its credit rating. Eventually, with debt markets closed to the Greek government, an EU bailout became inevitable. Table A.1 summarizes the key events of the debt crisis. From the beginning, the German government was sceptical of a financial rescue for Greece, emphasizing the scale of tax evasion and corruption as key obstacles to any per- manent improvement.5 It finally agreed to the bailout in exchange for harsh austerity mea- sures. Greek public opinion accordingly saw Germany as the instigator of foreign-imposed austerity. The reaction was immediate and intense: in February 2010, the Greek Consumers Association called for a boycott of German products – explicitly highlighting the importance of cars.6 Incendiary press coverage amplified the animosity. German newspapers portrayed Greeks as lazy cheaters, living it up at the expense of German taxpayers.7 The cover page of a German weekly featured Aphrodite making a rude gesture; a tabloid urged Greece to sell

3General Student, the German commander of Crete, instructed his forces to “leave aside all formalities and deliberately dispense with special courts.” Shortly thereafter, following the death of a German officer in Kondomari, Crete, German troops shot 19 civilians (Meyer, 2002). 4This figure does not include Jewish Greeks, shipped to death camps or murdered in Greece, nor victims of starvation. 5“German “no” to facilitating the repayment of the 110 billion euros”, Kathimerini, 13 October 2010. Research based on Greek bank lending data illustrates the scale of unreported income, concluding that as much as 30% of the country’s deficit may have been driven by tax evasion (Artavanis, Morse, and Tsoutsoura, 2016). 6Greek Consumers Association, 13 February 2012. 7“Die Griechenland-Pleite”, Focus Magazine, Nr.8, 2010.

7 some of its islands to repay its debts.8 As the Greek economy contracted and unemployment surged amid severe cutbacks of government services and support payments, anti-German sentiment in Greece deepened. In early 2012, Greek president Karolos Papoulias publicly complained that the German finance minister Wolfgang Schauble¨ was insulting the entire country.9 During the 2012 visit of German chancellor Angela Merkel to Athens, thousands of demonstrators filled the streets of Athens.10 Memories of Nazi massacres during the Occupation frequently resurfaced in Greece during that period. When a journalist from The Guardian newspaper interviewed Greeks during the Euro debt crisis about their country’s treatment by Germany, the massacre at Distomo immediately came up. A 45-year old bar owner contrasted this with the period immediately preceding the crisis: “Five years ago, no one had any problems with Ger- many.”11 In the past, family members of the victims of Distomo had sued for reparation payments, taking their case to the German courts and to the International Court of Hu- man Rights. Although Germany’s Constitutional Court dismissed the case in 2003, it was revived when an Italian court awarded victims’ descendants Italian property owned by a German non-governmental organization. The case reached the International Court in 2012, at the height of the Greek debt crisis, and featured prominently in the Greek press.12

3 Data

In this section, we describe the data on car registrations, German-Greek conflict, German reprisals and martyr status, and discuss data descriptives and issues of balancedness. In addition, we introduce the survey we conducted in Greece in the summer of 2017.

3.1 Car Registrations

Aggregate car sales in Greece slumped after the start of the financial crisis. Annual unit sales averaged around 180,000 before 2007. By 2011, with the Greek economy contracting rapidly, car sales fell to 60,000 per annum, a decline by two-thirds. Analyzing sales trends of cars in Greece is complicated by the fact that German car manufacturers performed strongly over the last decade. Worldwide, until 2015, the share of German brands was rising. This

8“Verkauft doch eure Inseln, ihr Pleite-Griechen”, Bild, 27 October 2010. 9“Greek president attacks German minister’s “insults””, Reuters, 15 February 2012. 10 “Athens protests amid Angela Merkel’s visit”, BBC, 9 October 2012. 11“Greece will pay its debts back, if you let us. But not with a German knife held to our throats”, Daily Telegraph, 11 February 2012. The piece profiles the life of Eleftherios Basdekis, who spent his “entire life beneath a German cloud”. A survivor of the Distomo massacre, he eventually built a trucking business that went bankrupt during the sovereign debt crisis. The article also cites a mother from Distomo saying that she “hated Germany”, that Angela Merkel was “a monster”, and that the Germans “killed Distomo; they stole our gold; they belittle Greece.” 12“The government in the Hague for Distomo”, Kathimerini, 13 January 2011.

8 partly reflects the success of Volkswagen and the decline in Toyota’s market share.13 The Greek Ministry of Transport and Communications collects data on registrations of new passenger vehicles. These are disseminated by the Hellenic Statistical Authority (ELSTAT). We use monthly data on the number of new passenger vehicles registered in each prefecture for the period January 2008 to December 2014, by manufacturing plant, and exclude small manufacturers with less than 10 vehicles sold during the entire period. The raw data from ELSTAT does not contain information on the brand of registered vehicles. However, ELSTAT provides a correspondence list that allows us to match production plants to car manufacturers. This correspondence does not always distinguish between brands produced by the same manufacturer. This is true for the Daimler group, producer of both Smart and Mercedes vehicles, and for the Fiat group, which also produces Alfa Romeo and Lancia. We are nonetheless able to distinguish German from non-German brands in our sample; the former include Volkswagen, Opel, Audi, BMW, Porsche and the brands of the Daimler group. For our purposes, a car’s “nationality” is not determined by ultimate ownership of the company, but the place of manufacture of (most) cars. We count Seat as a Spanish carmaker despite the fact that it is owned by Volkswagen. To the extent that Seat is actually perceived as German, we will understate the shift away from German cars, biasing our results downwards. Table 1 summarizes our data. Toyota has the highest sales, followed by Opel and Volk- swagen. Many German cars are luxury products. These may suffer greater declines in sales as a result of the crisis. To address this concern, we will exclude luxury cars in part of our analysis. Unfortunately, there is no good data for the immediate post-war period on car registrations. We do not know whether German car sales were initially weak after 1945. The only available data disaggregated by car manufacturer was compiled by ELSTAT for 1961, and is available for 11 transport areas. Areas with above-median shares of population affected by massacres showed a three-percentage point gap in German market share (23.6 vs 26.5%). In other words, the gap in the immediate post-war period is in line with lingering resentment; however, prior to the outbreak of the economic crisis, the gap had closed.

3.2 Conflict Index

We compile an index of conflict by counting newspaper articles that refer to political ten- sions between Greece and Germany. To illustrate our approach, we use the LexisNexis news database. Figure A.1 shows the frequency of the joint occurrence of the words “anti- German” and “Greece” in articlesreen in international newspapers. This word pair is virtu- ally nonexistent prior to 2009, but thereafter the frequency count increases sharply before peaking in 2012 and 2013. To obtain a systematic measure of perceived German-Greek conflict within Greece, we

13“VW conquers the world,” The Economist, 7 July 2012.

9 compute the frequency of conflict-related articles in a leading Greek newspaper, Kathimerini. It is the largest daily newspaper by circulation during the period of our study, and its entire archive of articles is digitized and available electronically. The database comprises 101,889 articles published under the sections ‘Greece’, ‘Politics’, and ‘Economy’. We compute the monthly share of articles related to German-Greek conflict using an approach similar to the one described in Baker, Bloom, and Davis (2016). We first select all articles containing the stem ‘German’, and then manually classify a subset of these articles based on whether they refer to German-Greek conflict. With audited articles in hand, we examine the word frequencies in relevant articles compared to those classified as irrelevant. Out of the list of most distinctive words (i.e. those that are relatively more frequent in articles referring to German-Greek conflict compared with the rest), we select the combination of 5 words (haircut, summit, Merkel, troika, eurozone) that minimizes false negative and false positive classifications of articles in our training set. The resulting set of terms, which we then use to classify the remaining articles, performs closest to the “gold standard” of human readings. More details are provided in Section C.1 of the Appendix. This method yields a monthly count of conflict-related articles, which we normalize by the total number of articles published in Kathimerini during the same month. Figure 1 plots the share of conflict-related articles for the period 2009–2014. While there is a gradual rise in the overall conflict article share after 2010, there are also numerous short-term spikes, when public arguments between Greek and German politicians grew particularly heated. Several of them are concentrated in 2012, when unrest and protests against austerity in Greece coincided with dissatisfaction around the International Court’s ruling in favor of Germany on the issue of wartime reparations.

3.3 Reprisals and Martyr Status

We create a unique dataset of the universe of Greek towns that experienced reprisals by German occupying forces during World War II. With the help of a professional historian, we combine information from a variety of sources, including tables of destroyed localities compiled by the Greek ministries of Reconstruction and Social Welfare, primary archival sources in the Libraries of the Hellenic Parliament, and the archives of the committee put together by the Ministry of the Interior to deal with localities’ applications for official recog- nition of their reprisal status. We complement these primary sources with information from secondary historical literature. A list of sources is provided in Appendix Section C.2. The final list contains 396 towns that experienced reprisals in at least one of the sources we con- sulted. Because data on car registrations is available at the prefecture level, we construct a prefecture-level measure of exposure to reprisals, computed as the share of towns that ex- perienced reprisals divided by the total number of towns in existence in 1940. Information on the latter comes from the 1940 Greek census. In 1993, the Greek parliament debated recognition of Kalavryta, the site of a major civilian massacre, as a “martyr town”. During the discussion, the idea of launching a procedure to award similar status to other towns affected by reprisals during World War

10 II emerged. This was written into law in 1997. In accordance with Law 2503, a committee composed of professional historians and high-ranking senior officials was established in 1997. It adjudicated applications for martyr status. The Ministry of the Interior at the time stated that

“the ... Committee was established to examine all the proposals for the character- ization of towns and villages as ‘martyr’ towns and villages and submit relevant proposals to the Ministry. The characterization of ‘martyr’ was important for the great contribution of towns and villages in the national struggle against the occupied forces in the period 1941–1944, during which the toll in human lives, as well as the material damage, was tragic.”

The Presidential Decree that officially awarded martyr status laid down the following crite- ria for recognition of a town as martyred:

1. The total destruction of houses following a burning, explosion or bombardment.

2. The loss of 10% of the overall population due to collective or individual executions.

3. The destruction of 80% of the total number of dwellings due to burning, explosion or bombardment and the loss of 10% of the population ... also taking into consideration the absolute losses.

While the committee followed the rules mentioned above in principle, it exercised dis- cretion on various occasions and used additional sources and supporting evidence that are not recorded in the minutes. For example, on March 28, 2005, it handled the case of Vlacherna. The committee minutes state “...out of the pre-existing 186 houses, 50 were completely destroyed and 131 partly destroyed.” The committee however “unanimously decided to suggest the characterization of Vlacherna as a martyr village, since the afore- mentioned “partial” destruction of the houses is close to 80%, therefore we have a case of complete destruction.” The same day, it dealt with the case of Mouzakata, and argued that “33 out of the 106 pre-existing houses were completely destroyed...there is no adequate documentation that Mouzakata fulfils the requirements to be characterized as a martyr vil- lage. The Committee unanimously decided not to include Mouzakata...”. This example suggests a degree of randomness in allocating martyr status. The minutes of the committee discussions do not suggest that political considerations played a role in the ultimate deci- sion. Instead, in Section 6 we provide evidence that the committee followed a behavioral heuristic in assigning martyr designations to towns and exploit that to identify the effect of public recognition on collective memory and economic behavior. Figure 2 provides a map of German reprisals and of the towns designated as martyred. It shows a high concentration of reprisals in Central and Northern Greece, in the Pindus mountains, and on Crete. Though reprisals are correlated with rugged terrain, they are relatively widely dispersed across the territory.

11 3.4 Data Descriptives, Control Variables and Balancedness

Our main dataset contains information on 51 prefectures over the period January 2008 to December 2014. The main features of the data are summarized in Table 2. The average prefecture in our sample saw monthly sales of 211 cars during the period; sales were as low as zero in some prefectures, and could go as high as 16,365 cars per month in Athens. German cars had an average market share of 26%; especially in the smaller prefectures, the share fluctuates substantially from month to month. The mean proportion of articles in Kathimerini referring to Greek-German tensions during the 2008- 2014 period is 6.2%. Massacres during the German Occupation occurred in 36 out of 51 prefectures, or 70% of the sample. On average, a little more than five percent of Greek towns in 1940 were affected by reprisals; Iraklio in Crete is the worst-affected prefecture, with a share of 31%. Around 2 percent of Greek towns were designated as martyr. We collect a large number of prefecture-level control variables, including contemporary economic controls from the 2001 Greek Census and Eurostat, historical demographic and electoral controls, as well as geographic characteristics of prefectures. While our identifi- cation strategy does not require that prefectures with and without reprisals are balanced in terms of observables at baseline, Table 3 demonstrates that this is largely the case. Log population and the pre-crisis German market share are higher in prefectures affected by reprisals, but these differences disappear when we compare prefectures with martyred towns to the rest. The share of employment in agriculture is similar, as is the propor- tion of the labor force in industry. The share of civil servants, a group hit hard by the crisis, is almost identical. Education levels and unemployment rates are also comparable, and the income difference is small. Few meaningful differences exist also for geographic and historical controls. Prefectures with and without reprisals do not differ significantly in terms of the share of parliamen- tary seats assigned to the Communist Party in the last pre-war election in 1936. This is important because Communist-dominated ELAS was the main guerrilla group during the Occupation. Differential support for the communists could proxy for deeper ideological dif- ferences across prefectures likely to drive differential reactions to the German-Greek conflict during the debt crisis. As expected, prefectures with reprisal towns are on average more rugged, and closer to a 1940 road or railway line, but these differences are not significant, and disappear when comparing prefectures with and without martyr towns. More rugged terrain provided cover for the partisans, whose bases often lay in the mountains of Central and Northern Greece. Throughout our main analysis, we demonstrate robustness of results to the inclusion of interactions of these baseline controls with the monthly share of conflict articles. Importantly, the pre-crisis first differences of the German car share are orthogonal to reprisal status (Bruhn and McKenzie, 2009). This indicates that there are no differential pre-crisis trends in car sales, which is crucial for the validity of our difference in differences specification.

12 3.5 Survey

In the summer of 2017, we conducted a telephone survey in Greece. We sampled a to- tal of 928 individuals from 143 municipalities, distributed across 12 prefectures. Respon- dents were drawn from 30 martyred municipalities and 113 control towns that did not see reprisals. For each prefecture, we aimed for an approximately balanced sample of respon- dents between martyred and control towns. In addition to socio-economic characteristics, we collected information on respondents’ actual and ideal car, views of Germany and its role in the crisis, as well as proxies of activism and national identity. Table A.2 in the Appendix provides summary statistics for our survey sample. The sample is balanced on observables – there are no significant differences of age or gender, and the proportion of unemployed is actually higher in the non-martyred locations. The average respondent in a martyred town earned 66 euros less than a respondent in a non-martyred town, but the difference is not significant. Education levels are similar and the cars owned by survey participants are about the same age in both groups.

4 Main Results

In this section, we present our main result – that German car sales declined more in reprisal- affected prefectures during months of German-Greek conflict. We also show that this does not reflect the economic effect of the crisis, and is not driven by persistent underlying differences in nationalism across locations.

4.1 Baseline Estimates

We estimate the following equation for the share of German car sales in each prefecture i at time t:

Sit = ci + yt + mt + γ1Article Sharet + γ2Article sharet × Share townsi + eit (1) where S is the share of German cars, ci are prefecture fixed effects, yt are year fixed effects and mt are calendar month fixed effects that account for seasonality in the German market share. γ2 is the main coefficient of interest – the extent to which the German car market share declines differentially in prefectures with more towns that experienced reprisals by the Germans during World War II. We augment the above specification with interactions of baseline prefecture-level con- trols with the monthly share of conflict-related articles, as well as interactions of prefecture fixed effects with calendar month fixed effects, to account for prefecture-specific seasonality patterns in car sales. Our most parsimonious specification includes a full set of time (month) fixed effects and thus accounts for any unobservable factor that varies at the monthly level and affects all prefectures in the same way.

13 Table 4 gives the results. We find that overall, the higher the share of conflict articles in a given month, the lower the share of German cars – but the effect is not tightly estimated. However, in areas with a higher share of reprisal towns, there is a large and significant decline in German car sales in exactly those months when German-Greek conflict peaked. The effect is also economically large: Going from the median share of towns affected by reprisals (0.0167) to the 75th percentile (0.087) implies a reduction in German car sales by 0.6 percentage points in the median conflict month, a 2.5 percent relative decline. In column 2, we add an interaction of the conflict article share with a full set of pre-crisis control variables (share of agriculture, industry, population with secondary and higher ed- ucation, share of civil servants, average distances to road and rail in 1940, ruggedness, unemployment in 2008, log GDP per capita in 2008-2009 and per capita growth in GDP during the same period). This is intended to capture any differential effect that political ac- rimony between Germany and Greece after 2010 might have had in areas that were different before the crisis, either economically, or in terms of historical characteristics and political orientation. Adding these variables only increases the size of our coefficient of interest, which now implies a decline of German market share by an extra 1.2 percentage points for prefectures at the 75th percentile of exposure to reprisals, compared to the median exposure level. Similarly, adding an interaction of prefecture fixed effects and calendar month fixed effects (Column 3), thus accounting for seasonality patterns that potentially differ across prefectures, strengthens our results. In column 4 we control for a full set of time fixed effects. This leaves the interaction coefficient basically unchanged.

4.2 Accounting for the Effects of the Crisis

During the Eurozone crisis, Greece experienced a severe economic downturn. Poor current economic performance (and the prospect of future contraction) affected car purchases. Ger- man cars tend to be more expensive than those from other countries. While non-German brands sold in Europe were worth an average of e23,000 in 2012, the average German car sold for e36,600 – 58% more (European Commission, 2011). In the middle of the largest economic crisis to hit Greece in a generation, the taste for expensive cars may well have evaporated, providing an alternative interpretation of our main finding – if such effects differed by locality. We do not have direct measures of output at the town or prefecture level. To account for the effect of the crisis, we employ three strategies. We first show that our findings are unaffected after controlling for a number of variables that capture differential economic performance. Second, we repeat our analysis by dropping luxury cars, which were more likely to have been affected by the crisis. Third, we use survey evidence on buying inten- tions to show that distaste, and not empty pockets, stopped Greeks in martyred cities from purchasing German cars.

14 4.2.1 Controls

We first use the interaction between monthly unemployment in Greece as a whole and the share of reprisal towns in a prefecture to account for time-varying economic conditions that might have a differential impact across prefectures (Table 5, Panel A). There is no direct effect of unemployment on the share of German cars in the country as a whole – nor is there a differential effect in prefectures more affected by German massacres. The differential effect of time-varying conflict on prefectures with a reprisal history survives essentially unchanged. In Panel B, we use nightlight density as an indicator of economic activity. This is com- puted using luminosity data from NASA at the prefecture level. This data is only available until 2013, and is missing each year between the months of April and August. Due to these limitations, we lose about half of our sample. The basic pattern in the half of the sample with non-missing data is the same as in the rest, but coefficients on the interaction term are not significant at standard levels of confidence (Panel B1). The magnitude of the coefficient in the available sample is unaffected by the inclusion of a control for night-time luminosity (Panels B2 and B3). This indicates that the drop in German car market share we identify is not the result of high-frequency changes in economic activity correlated with the location of reprisals. As a third alternative, we compile an index of economic uncertainty at the national level. We use Google trends and extract the monthly volume of searches (in Greek) for “economy” or “economic”, “uncertainty” or “uncertain” and “policy” (Baker, Bloom, and Davis, 2016). Appendix Figure A.2 plots this index over time. We then control for the interaction of this measure with the share of a prefecture’s towns that experienced reprisals. Economic policy uncertainty has no differential effect on reprisal prefectures, but the differential effect of monthly German-Greek conflict remains large and significant. This is further evidence that our baseline measure captures German-Greek animosity specifically, and is not just a proxy for the debt crisis.

4.2.2 Sample Restriction

An alternative to controlling for differential economic performance is to exclude all expen- sive cars from our sample. This is done in Table A.3. Column 2 only uses sales of cars in the ‘Volkswagen category’, meaning non-luxury sedans (i.e. excluding Audi, BMW, and Mer- cedes). The coefficient on the interaction declines somewhat in size, but remains large and significant. This indicates that effects are not driven by differential declines in purchasing power across prefectures, manifesting as a curtail of purchases of luxury items.

4.2.3 Survey Evidence

Survey evidence from over 900 Greek respondents suggests there was no major additional economic hardship in martyred areas. As shown in Appendix Table A.2, the share of un- employed and the income of respondents are statistically indistinguishable across martyred

15 and non-martyred towns. In the summer of 2017, respondents were asked when they be- came unemployed; dates are not statistically different in both samples. Survey data also helps to further untangle whether distaste for Germany’s policies or economic distress drove the decline in German car sales. We do so by analyzing information on car ownership and buying intentions. Figure 3 summarizes our main findings. On average, 18% of Greeks in non-martyred towns drive German cars; this falls to 17.3% in martyr towns, and to 12.3% for those in martyr towns whose parents were also born in the same municipality. Importantly, not only actual purchasing behavior, but also aspirations differ by martyr status. German cars are an aspirational good for most respondents – many more people name a German brand as their ideal than there are respondents who actually own a Ger- man car. Respondents in martyr towns like German cars less. While on average, 32% of Greeks say that their ideal car is German, this share drops to 26% in martyr towns, and to 23% amongst people with parents native to the town. The differences with the non- martyred towns are significant at the 10% level. The difference in taste strongly suggests that differential economic performance is unlikely to drive our main result. The fact that Greeks in martyred towns whose parents were native to the town are more anti-German in their purchases and buying aspirations is an additional piece of evidence supportive of this interpretation – recent movers share the same economic environment, but do not show the same (additional) distaste for German products. Table 6 verifies that the pattern presented in Figure 3 holds more generally. We use town- level estimates of destruction provided in publications of the Ministries of Reconstruction and Social Welfare to assess how the degree of exposure to reprisals affected purchasing behavior and aspirations. Destruction is 0 for non-martyr towns. Respondents in towns that experienced a higher level of destruction are less likely to own a German car, though the effect is not significant. They are instead significantly less likely to report a German car as their ideal car, an estimate that increases in magnitude after controlling for a large number of individual socioeconomic characteristics and prefecture fixed effects. A final piece of evidence suggesting that results are not driven by differential economic conditions in martyr towns is blame attribution. The survey allows us to ask more directly to what extent Greeks in martyred towns hold different beliefs about Germany’s role during the Euro debt crisis. In the last two columns of Table 6, we analyze answers to the question “Do you blame Germany for the recent debt crisis?” While the majority of Greeks puts most of the blame on their own politicians, those in towns that experienced more destruction during the Occupation are markedly more likely to hold a negative view of Germany.

4.3 Accounting for Unobserved Culture

One concern is that prefectures with a higher incidence of reprisals were characterized by higher pre-war nationalism or a greater ability to organize collectively. If these traits persisted in the modern period, they could independently affect hostility against Germany and boycott activities. Our results would then reflect persistence of a cultural trait, and

16 not the effects of memory.14 Survey results and an IV strategy that identifies plausibly exogenous factors causing greater German repression allow us to demonstrate that we are capturing the repercussions of events in the 1940s and not latent cultural differences across prefectures. We can examine directly whether Greeks in martyred towns are more nationalistic today, or engage in collective action more readily. There is evidence for neither: In our survey, we asked how proud respondents were to be Greek. Table 7 shows no significant differences in national identity by exposure to destruction. There is also no evidence for higher activism or propensity for collective action: Respondents in towns with a higher share of destruction are participating in demonstrations and strikes at the same rate as the rest of the respondents. To provide additional evidence that effects are driven by the memory of reprisals and not latent propensity to engage in conflict, we use historical information on the Greek partisan war and the logic of guerrilla conflict to isolate exogenous variation in German reprisals. No military group can fight for any length of time without substantial supplies. Food, weapons, ammunition, fuel and clothing are everyday necessities. In this regard, irregular forces are no different from regular ones, even if their demands are lower (Van Creveld, 1982). Greek (and other European) partisans during World War II were typically resupplied by the Allies by air, and where feasible, by sea. In Greece, the British dropped supplies by parachute from aircraft flying from Egypt, or landed smaller planes on improvised airstrips; they also used submarines to bring agents and military supplies into occupied Greece. Supply points are important for guerrilla operations, but they are not easy to operate and defend. They need to be sufficiently remote for the occupiers to have difficulty in controlling the territory, but accessible enough for Allied aircraft and submarines to reach them. Based on this observation, we construct an instrument for partisan resupply – and, consequently, for reprisal attacks – based on the combination of two terrain characteristics. The first component of the instrument is ruggedness. Rugged areas generally provide safe havens to insurgents. Partisans and guerrillas fighting conventional forces are normally heavily outgunned, and need to use asymmetric warfare to even the odds. Che Guevara (1961) already emphasized the importance of inaccessible terrain for the initial stages of a guerrilla uprising. By retreating to the most inaccessible areas of a country, insurgents can typically deny their foe the use of his strongest assets – armor and heavy artillery (Asprey, 2002). This pattern held in 1940s Greece, where the largest concentrations of partisans were located in inner parts of Central Greece and the Peloponnese and around the Pindus mountains in Northern Greece. To calculate ruggedness, we use the method pioneered by Nunn and Puga (2012). The left panel of Figure 4 shows that ruggedness predicts more massacres. A one standard deviation increase in ruggedness increases the likelihood of a prefecture experiencing reprisals by 8%.

14This would be in line with a rich literature on long-term persistence (Guiso, Sapienza, and Zingales, 2016; Nunn and Wantchekon, 2011; Voigtlander¨ and Voth, 2012)

17 The second component of the instrument is variation in the ruggedness of the terrain. Conditional on a location’s ruggedness, which would have determined the presence of partisans, it was easier for the allies to drop supplies from the air in locations that were relatively flat. Small valleys ringed by rugged mountains allowed the partisans to locate supplies more easily, and at the same time be sufficiently protected from the German army while picking them up. To proxy for this determinant of supply points, we use the standard deviation of ruggedness. The middle panel of Figure 4 shows that this variable is strongly positively correlated with the share of destroyed towns in a prefecture. We construct a single instrument (Predicted resupply) out of the interaction of these two variables (Munshi and Rosenzweig, 2016). Based on German Army maps on partisan activ- ity and Allied supply points in Greece in 1944 from the German federal military archives (Figures A.3 and A.4), Table A.4 in the Appendix shows that this variable predicts actual proximity to supply points. It is also strongly predicts the share of reprisal towns in a prefecture, as can be seen in the right panel of Figure 4. In Table 8, we present the results of an IV analysis using predicted resupply as an instrument for a prefecture’s share of towns that experienced reprisals. Column 2 presents the coefficient from the parsimonious specification of Column 4 in Table 4. The first stage (column 2) is strong, with the F-statistic of 18.03 being well above the rule-of-thumb cutoff of 10. Columns 3 and 4 present reduced form and 2SLS coefficients, respectively. The estimated differential effect of German-Greek conflict is larger in size than the OLS estimate. The fact that the instrumented part of the variation in town destruction strongly predicts declines in German car purchasing in times of conflict strengthens the argument that the link between car buying and revived anti-German sentiment is indeed causal. Furthermore, the effect does not appear to be driven by outlier prefectures. Figure A.5 in the Appendix plots 2SLS estimates and confidence bands after dropping each prefecture from the sample. Effects are always significantly negative.

5 Additional Robustness and Placebo Exercises

In this section, we address potential issues with our analysis (Appendix Section B provides an overview of possible challenges and the way we resolve them). First, we conduct a set of placebo exercises to increase confidence in the fact that the changes in German car market share are directly driven by revived memories of German reprisals. Next, we demonstrate that our main results are not driven by a handful of “famous cases” or by the definition of our conflict article index. Finally, we show that our inference is not driven by patterns of spatial correlation.

5.1 Falsification Tests

We test the plausibility of our results by conducting two placebo exercises. It is essential for our argument that the changes in German car purchases are specific to German atrocities and time-varying conflict with Germany. First, we repeat our baseline analysis using as

18 source of cross-sectional variation the share of towns destroyed by Bulgarian and Italian occupying forces. Second, we compile an index of articles about disagreements with Italy for the same period, and examine whether it has explanatory power for changes in the German market share. Panel A of Table 9 shows that the market share of German cars is not more affected during conflict months in prefectures with a higher share of towns exposed to Italian and Bulgarian reprisals. This speaks against location-specific unobservable cultural differences driving our main results. Locations more prone to conflict or with a stronger national iden- tity might have provoked reprisal attacks from any occupying forc, regardless of nationality – and should have reacted to German ’bullying’ after 2009 whether the massacres in the early 1940s were committed by Italians, Bulgarians, or Germans. Yet it is atrocities commit- ted specifically by Germans that drive time-varying responses of consumer behavior. Next, we replicate our baseline analysis with an Italian news index. If our results pick up the effect of a Germany-specific news shock, we should not find effects if we use news coverage of another European country. We use the same keywords designed to capture German-Greek conflict during the Euro crisis, but substitute the word stem ‘German’ with ‘Italian’ and the German Chancellor Merkel with the Italian Prime Minister Renzi. De- spite substantial overlap in the terms composing the Italian placebo index, the estimated differential effect is far from significance, as shown in Panel B of Table 9.

5.2 Famous Massacres

Amongst the litany of bloodshed that marked the German occupation of Greece, a few cases stand out – Kalavrytra, Distomo, and Kommeno. In each of these towns, almost the entire civilian population was killed. We examine the robustness of our findings to dropping these famous cases. In Appendix Table A.5, we first drop the three towns when we compile indices of the share of towns affected by reprisals (Panel A), and then exclude the entire prefectures in which they are located (Panel B). Results are almost entirely unaffected. The coefficient on the interaction between the share of destroyed towns and news coverage of German-Greek conflict remains large and highly significant, suggesting that effects are not driven by a handful of famous cases.

5.3 Alternative News Index

As an alternative measure of German-Greek conflict, we use data from Google Trends dur- ing the period 2008–2014. The alternative conflict index is based on the following terms: “Germans”, “German reparations” and “Distomo” (only searches that were conducted in the Greek language, in Greece). The index value from Google Trends is a normalization of the share of total searches represented by a term in a given time and region. For each of the terms above we compile a monthly search index from Google for the period 2008–2014 and aggregate them into a single measure, ranging from 0 to 100. This index has two advantages: it represents a measure of demand, rather than supply of

19 news on German-Greek conflict, and it consists of terms more directly related to the history of the German occupation of Greece during WWII, rather than the debt crisis. Table A.6 shows results using this index as a time-varying measure of conflict. We consistently find negative, highly significant interactions between the Google index and German massacres. The effect is almost identical in size to the main specification – a one standard deviation increase in German-Greek conflict interacted with the share of towns that suffered reprisals leads to a 0.8% decline in the German car share in both our baseline and in Column 1 of Table A.6.

5.4 Spatial Correlation

Geographic data usually display a high degree of spatial correlation, and regressions in- volving geographic data on the left and right hand side may produce artificially inflated t-statistics. Conley standard errors adjusted for spatial correlation may not be enough to remedy this problem. Following Kelly (2019), we take a two-step approach to addressing the issue of spatial correlation. First, we assess the degree to which spatial correlation may be a problem in our data. We compute Moran’s I for the residuals resulting from Column 4 of Table 4 for every month in our data. Figure A.6 plots z-scores by month. Only in about 6% of cases is the z-score larger than 1.96 (the suggestive cutoff for significance at the 5% level). The average value of Moran’s I in our data is 0.075, lower than any of the persistence studies analyzed in Kelly (2019). The test based on Moran’s statistic thus indicates that our data is not characterized by a high degree of spatial correlation. We further confirm this by replacing our main explanatory geographic variable (Share towns) with spatial noise. We run 1000 replications of the regression in Column 4 of Ta- ble 4, each time drawing a vector of spatially correlated random values for Share towns.15 Figure A.7 plots p-values resulting from this simulation exercise, alongside the p-value in Column 4 of Table 4 (vertical line). The explanatory power of noise is higher than that of Share towns less than 5% of the time. Overall, we conclude that spatial correlation is not an important threat to inference in our context.

6 Institutionalizing Collective Memory

In this section, we first show that official recognition of martyr status affected consumer behavior. In other words, a government-awarded label that publicly recognizes the suffering of a town is associated with behavioral changes in times of conflict. Importantly, we isolate an exogenous component of the recognition decision, using RDD-style variation. We also demonstrate that other forms of public remembrance at the local level have similar effects.

15Specifically, we draw from the standard normal distribution, but impose a variance covariance matrix based on the Matern function, as in Kelly (2019). We report results for a correlation range of 30 degrees, but the pattern is similar with different ranges.

20 6.1 The Effect of State Recognition

We start by providing evidence that the ministerial committee assigned with the task of awarding martyr status to towns affected by reprisals did so by following a relatively sim- ple heuristic. Figure 5 is a binned scatterplot of reprisal towns granted martyr status by percentage of destruction. The likelihood of receiving martyr status is increasing in the extent of destruction recorded in ministerial sources. Importantly, the probability of re- ceiving it jumps around a destruction level of 50%. This is examined more systematically in Figure 6, which shows how the probability of martyr status being granted changes for alternative cut-offs. 50% is the only level with a significant positive value. As the reading of the committee minutes suggests, there was substantial leeway in making a final determi- nation; the distinct jump at the 50% cut-off suggests that the committee de facto followed the rule of assigning martyr status to borderline cases whenever half or more of the town was registered as destroyed in official records. We verify this jump in the probability of receiving martyr status, by implementing a regression discontinuity design in the sample of towns that experienced reprisals. The result is shown in column 1 of Table 10. The estimator uses a bandwidth of 0.14, with 17 and 22 municipalities to the left/right of the threshold. Results are robust to using a quadratic polynomial on either side of the cutoff and to specifying different optimal bandwidths.16 The same jump is not found when we use population in 1940 or in 2010 as dependent variables in columns 4 to 9, suggesting that no discontinuity in other characteristics of towns exists around the 50% destruction threshold. The evidence from the RDD and a close reading of the discussions of the committee minutes suggests that the awarding of martyr status – while correlated with destruction and population loss – also contains an important accidental element. We therefore ask whether the receipt of martyr status influences German car purchases above and beyond the effect of destruction itself. To this end, we first show that martyr status itself is strongly associated with declining German car sales. Next, we demonstrate that when we use both variables, martyr status and destruction, the effect of martyr status dominates – underlining the importance of institutionalized collective memory and officially recognized victimhood. Panel A of Table 11 replicates results from Table 4 for purposes of comparison. In Panel B, we estimate equation 1 using the share of martyr towns in a prefecture as the source of cross-sectional variation. The differential effect of conflict in prefectures with a higher share of martyr towns is negative and significant. Magnitudes are consistent across specifications and estimated effects are large: a one standard deviation increase in the anti-German news share, evaluated at the mean of the respective measure of reprisal exposure, leads to a

16These results are based on the rdrobust routine, as described in Calonico, Cattaneo, and Titiunik (2014). We also used the alternative technique implemented in the STATA routine rd, using a range of bandwidths to show robustness as described in Nichols (2016). For the specifications in columns 1-3 of Table 10 we find treatment effects of 0.35, 0.38 and 0.28 with p-values of 0.01, 0.02, and 0.013.

21 market share decline of 1.5% in both the baseline specification and the martyr analysis (equivalent to 5.7% of German market share). Next, we examine the effect of martyr status, conditional on the share of towns affected by reprisals in a prefecture. In Panel C of Table 11, we conduct a horserace between the (time-varying) effect of reprisals and public recognition. We always control for the average extent of destruction in a prefecture, to account for the fact that martyr status is more likely to be awarded, on average, to towns with higher estimated levels of destruction. Prefectures with a higher share of martyr towns display large and significant differential drops in the German car market share in months of conflict. Prefectures more exposed to reprisals also register a differential drop, though this is not significant. That means that in a set of two prefectures, with the same average exposure to reprisals, the one with a higher share of officially recognized martyr towns saw a much sharper fall in German market share – officially approved and publicly recognized status as a major victim of Nazi aggression matters over and above the memory of conflict itself. Finally, we exploit the discontinuity in the awarding of martyr status at the 50% cutoff of destruction to isolate the exogenous component of public recognition. In Panel D of Table 11, we use the predicted share of martyred towns in a prefecture based on the 50% criterion to explain changes in market share. We create a dummy for towns that are assigned a predicted value > 0.5 in a regression of martyr status on a dummy for destruction higher than 50% and a quadratic polynomial in destruction. We then aggregate towns predicted to receive martyr status at the prefecture level, as a share of all towns in 1940. Estimated coefficients are again large and significant. As before, we control for the average level of destruction of towns in a prefecture, so that estimated effects are driven by the proportion of towns that are above the 50% threshold used by the committee to assign martyr status. Since we do not have town-level data on car sales, we cannot implement an exact ana- logue to the RDD specification – showing that a narrow band of destruction driving as- signment to martyr status is responsible for declining German market share. Instead, we show that the coefficient on the interaction between martyr status and article share remains negative and significant as the share of towns in a prefecture with destruction in the range of 25-75% increases (Table A.7).

6.2 Public Memorials

Official recognition as a martyred town is not the only way in which memories can be passed on from generation to generation. Many towns in Greece commemorate World War II atrocities, often by public festivities or through memorials. We conducted a telephone survey of mayors’ offices in all towns that experienced reprisals to ascertain whether a community organizes a commemorative event or whether a monument exists. There are approximately twice as many towns with a monument or public event as there are places officially recognized as martyred. Table 12 uses memorials to WWII massacres to measure the strength of collective mem- ory. Panel A shows that the drop in German market share in months of conflict is signif-

22 icantly larger in prefectures with a higher share of towns that commemorate reprisals. In Panel B, we separately examine the effect of memorials and of official government recogni- tion. All towns officially recognized as martyr have a public commemoration or memorial, so we construct a separate measure for the prefecture share of towns with a memorial that do not have official martyr status. Memorials in the absence of official martyr status strongly predict differentially lower German market shares after 2010. Martyr status itself also has a strong and, in most cases, larger effect on the German market share than memorials. These results suggest that the differential response of consumer behavior to time-varying conflict is increasing in the degree of institutionalization of collective memory.

7 Conclusion

How is collective memory created? And when does it matter? We argue that collective memory matters the most when the present is reminiscent of the past, for example by pitting the same groups of “insiders” and “outsiders” against each other. We also show that collective memory can be created exogenously through government intervention, by recognizing the particular status of a group or location. As such, public acts of recogni- tion amplify the effects of other “places of memory” – from local plaques and statues to commemorations. We analyze these questions in a setting with large economic stakes - the purchase of new cars. During the Eurozone debt crisis, Greece suffered from a major loss of international confidence. As the crisis worsened, relations between the Greek and German governments became increasingly acrimonious. The more conflict there was between German and Greek politicians, the lower the market share of German cars in Greece became. Much of the aggregate decline occurred in areas affected by German atrocities during the occupation. This was not driven by differential economic effects of the debt crisis or by underlying unobservable characteristics of locations that drove both German reprisals during WWII and contemporary consumer behavior. Survey respondents in reprisal towns were not more likely to show nationalist sentiment than those in non-reprisal towns – but they did blame Germany more for the Greek economic crisis and are less likely to name German cars as their ideal car choices. An IV strategy based on the logic of guerrilla warfare allows us to exploit terrain characteristics in Greece and isolate a plausibly exogenous component in German atrocities. The backlash against German cars was greatest in towns and cities that were awarded official “martyr” status from the Greek government. These patterns did not only reflect greater wartime suffering. Towns and cities above the destruction threshold value for the “martyr” designation were more likely to receive public recognition, creating arguably ex- ogenous variation in victimhood status. Towns thus recognized show greater declines in German market share. We find similarly large interactions with public remembrance rituals that are organized (bottom-up) locally. These findings could either reflect personal preferences, i.e. a greater dislike of Ger-

23 many, or concerns over the social acceptability of purchasing a German product. That survey respondents are more likely to name a German brand as their ideal choice suggests a muted role for social image considerations and points in the direction of deeply-held views of Germany that are revived by current conflict. The larger decline in German car purchases after 2009 in prefectures that were more affected by German reprisals during World War II suggest that collective memory can be a powerful driver of economic behavior, and that its effects are non-linear. Memories can be revived if the present resembles the past. Such a time-varying response is more pronounced when public remembrance is institutionalized through memorials, and through the official recognition of victim status. Our paper is the first to provide causal evidence for the role of institutionalized collective memory on economic decision-making.

24 References

Akerlof, George A. and Rachel E. Kranton. 2011. Identity Economics: How our Identities Shape our Work, Wages, and Well-being. Princeton: Princeton University Press.

Anderson, Benedict. 1983. Imagined Communities: Reflections on the Origin and Spread of Na- tionalism. New York and London: Verso Books.

Artavanis, Nikolaos, Adair Morse, and Margarita Tsoutsoura. 2016. “Measuring Income Tax Evasion Using Bank Credit: Evidence from Greece.” The Quarterly Journal of Economics .

Ashenfelter, Orley, Stephen Ciccarella, and Howard J. Shatz. 2007. “French Wine and the U.S. Boycott of 2003: Does Politics Really Affect Commerce?” NBER Working Paper 13258.

Asprey, Robert B. 2002. War in the Shadows: The Guerrilla in History, vol. 2. iUniverse.

Baker, Scott R, Nicholas Bloom, and Steven J Davis. 2016. “Measuring Economic Policy Uncertainty.” The Quarterly Journal of Economics 131 (4):1593–1636.

Bauer, Michal, Christopher Blattman, Julie Chytilova,´ Joseph Henrich, Edward Miguel, and Tamar Mitts. 2016. “Can War Foster Cooperation?” Journal of Economic Perspectives 30 (3):249–74.

Becker, Sascha O, Katrin Boeckh, Christa Hainz, and Ludger Woessmann. 2014. “The Em- pire Is Dead, Long Live the Empire! Long-Run Persistence of Trust and Corruption in the Bureaucracy.” The Economic Journal .

Belmonte, Alessandro and Michael Rochlitz. 2019. “The Political Economy of Collective Memories: Evidence from Russian Politics.” Journal of Economic Behavior & Organization (168):229–250.

Benabou,´ Roland and Jean Tirole. 2011. “Identity, Morals, and Taboos: Beliefs as Assets.” The Quarterly Journal of Economics 126 (2):805–855.

Besley, Timothy and Torsten Persson. 2010. “State Capacity, Conflict, and Development.” Econometrica 78 (1):1–34.

Bordalo, Pedro, Nicola Gennaioli, and Andrei Shleifer. 2015. “Memory, Attention and Choice.” Tech. Rep. w23256, NBER Working Paper.

Bruhn, Miriam and David McKenzie. 2009. “In Pursuit of Balance: Randomization in Prac- tice in Development Field Experiments.” American Economic Journal: Applied Economics 1 (4):200–232.

Calonico, Sebastian, Matias D Cattaneo, and Rocio Titiunik. 2014. “Robust Nonparametric Confidence Intervals for Regression-Discontinuity Designs.” Econometrica 82 (6):2295– 2326.

Che Guevara. 1961. Guerilla Warfare. New York: Praeger.

Collier, Paul, V. Elliott, H. Hegre, A. Hoeffler, Marta Reynal-Querol, and Nicholas Sambanis. 2003. Breaking the Conflict Trap: Civil War and Development Policy. Washington DC: World Bank and Oxford University Press.

25 Collier, V. Elliott H. Hegre A. Hoeffler Marta Reynal-Querol Nicholas Sambanis, Paul. 2016. Ultrasociety: How 10,000 Years of War Made Humans the Greatest Cooperators on Earth. Chap- lin, CT: Beresta Books.

DellaVigna, Stefano and Joshua M Pollet. 2009. “Investor Inattention and Friday Earnings Announcements.” The Journal of Finance 64 (2):709–749.

Dessi, Roberta. 2008. “Collective Memory, Cultural Transmission, and Investments.” Amer- ican Economic Review 98 (1):534–60.

Dincecco, Mark and Mauricio Prado. 2012. “Warfare, Fiscal Capacity, and Performance.” Journal of Economic Growth 17 (3):171–203.

Doxiadis, Constantinos. 1947. “Thysies tis Ellados: Aitimata kai Epanorthoseis ston B’ Pagosmio Polemo.” Report 19, Hellenic Ministry of Reconstruction, Athens.

European Commission. 2011. Car Prices within the European Union.

Fisman, Raymond, Yasushi Hamao, and Wang Yongxiang. 2014. “The Impact of Interstate Tensions on Economic Exchange: Evidence from Shocks to Sino-Japanese Relations.” Review of Financial Studies 27 (9):2626–2660.

Gabaix, Xavier, David Laibson et al. 2006. “Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive Markets.” The Quarterly Journal of Economics 121 (2):505–540.

Gennaioli, Nicola and Andrei Shleifer. 2010. “What Comes to Mind.” The Quarterly Journal of Economics 125 (4):1399–1433.

Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2009. “Cultural Biases in Economic Ex- change?” The Quarterly Journal of Economics 124 (3):1095–1131.

———. 2016. “Long-Term Persistence.” Journal of the European Economic Association 14 (6):1401–1436.

Harari, Yuval Noah. 2014. Sapiens: A Brief History of Humankind. Random House.

Hionidou, Violetta. 2006. Famine and Death in Occupied Greece, 1941-1944. Cambridge: Cam- bridge University Press.

Hirshleifer, David, Sonya Seongyeon Lim, and Siew Hong Teoh. 2009. “Driven to Distrac- tion: Extraneous Events and Underreaction to Earnings News.” The Journal of Finance 64 (5):2289–2325.

Hong, Canhui, Wei-Min Hu, James E. Prieger, and Dongming Zhu. 2011. “French Automo- biles and the Chinese Boycotts of 2008: Politics Really Does Affect Commerce.” The B.E. Journal of Economic Analysis & Policy 11 (1):26.

Kahneman, Daniel and Amos Tversky. 1982. Judgment Under Uncertainty: Heuristics and Biases. New York: Cambridge University Press.

Kelly, Morgan. 2019. “The Standard Errors of Persistence.” Discussion paper DP13783, CEPR.

Koku, Paul Sergius, Aigbe Akhigbe, and Thomas M. Springer. 1997. “The Financial Impact of Boycotts and Threats of Boycott.” Journal of Business Research 40 (1):15–20.

26 Mazower, Mark. 1995. Inside Hitler’s Greece: The Experience of Occupation, 1941-1944. New Haven and London: Yale University Press.

Meyer, Hermann Frank. 2002. Von Wien nach Kalavryta. Die blutige Spur der 117. J¨ager-Division durch Serbien und Griechenland. Mannheim: Peleus.

Michaels, Guy and Xiaojia Zhi. 2010. “Freedom Fries.” American Economic Journal: Applied Economics 2 (3):256–81.

Mullainathan, Sendhil. 2002. “A Memory-Based Model Of Bounded Rationality.” The Quar- terly Journal of Economics 117 (3):735–774.

Munshi, Kaivan and Mark Rosenzweig. 2016. “Networks and Misallocation: Insurance, Migration, and the Rural-Urban Wage Gap.” American Economic Review 106 (1):46–98.

Nessou, Anestis. 2009. Deutsche Besatzungspolitik und Verbrechen gegen die Zivilbev¨olkerung- eine Beurteilung nach dem Volkerrecht.Gottingen:¨ Vandenhoeck & Ruprecht.

Nichols, Austin. 2016. “Rd: Stata Module for Regression Discontinuity Estimation.” Work- ing paper, Boston College Department of Economics.

Nora, Pierre. 1989. “Between Memory and History: Les Lieux de Memoire.”´ Representations 26:7–24.

Nunn, Nathan and Diego Puga. 2012. “Ruggedness: The Blessing of Bad Geography in Africa.” Review of Economics and Statistics 94 (1):20–36.

Nunn, Nathan and Leonard Wantchekon. 2011. “The Slave Trade and the Origins of Mis- trust in Africa.” American Economic Review 101 (7):3221–52.

Ochsner, C. and P. Roesel. 2017. “Activated history – The case of the Turkish sieges of Vienna.” Working paper 6586, CESifo.

Pandya, Sonal and Rajkumar Venkatesan. 2016. “French Roast: International Conflicts and Consumer Boycotts – Evidence from Supermarket Scanner Data.” Review of Economics and Statistics 98 (1):42–56.

Schwartzstein, Joshua. 2014. “Selective Attention and Learning.” Journal of the European Economic Association 12 (6):1423–1452.

Serman, D.H. and H.S. Kim. 2002. “Affective Perseverance: The Resistance of Affect to Cognitive Invalidation.” Personality and Social Psychology Bulletin 28 (2):224–237.

Teoh, Siew Hong, Ivo Welch, and C Paul Wazzan. 1999. “The Effect of Socially Activist Investment Policies on the Financial Markets: Evidence from the South African Boycott.” The Journal of Business 72 (1):35–89.

Van Creveld, Martin. 1982. Fighting Power: German and US Army Performance, 1939-1945. Westport, Connecticut: Greenwood Press.

Voigtlander,¨ Nico and Hans-Joachim Voth. 2012. “Persecution Perpetuated: The Medieval Origins of Anti-Semitic Violence in Nazi Germany.” The Quarterly Journal of Economics 127 (3):1339–1392.

27 Figures and Tables

Figure 1: German-Greek conflict and German market share in prefectures with and without reprisals 4 2 0 −2 −4 Jan 2009 Jan 2010 Jan 2011 Jan 2012 Jan 2013 Jan 2014 Date

German share, reprisal−non reprisal Article share

Notes: The solid line is the difference in the seasonally adjusted (expressed as difference of month t from month t − 12) share of German car registrations in reprisal vs non-reprisal prefectures. The dotted line is the monthly share of Kathimerini articles related to German–Greek conflict. Both series are normalized by their standard deviation.

Figure 2: The geography of German reprisals in wartime Greece

Notes: See Section 3.3 and Appendix Section C.2 for sources on reprisals and martyr status.

28 Figure 3: German market share and ideal car shares by martyr status

Notes: The figure plots the share of survey respondents in each group that own a German car (German actual) or report a German car as their ideal car (German ideal).

Figure 4: First stage .2 .2 .15 .15 .15 .1 .1 .1 Share towns Share towns Share towns .05 .05 .05 0 0 0 100 200 300 400 500 0 50 100 150 200 0 20 40 60 80 Ruggedness Standard deviation of ruggedness Predicted resupply Notes: Each figure represents a binned scatterplot of the share of reprisal towns in a prefecture against the instrument denoted on the x-axis. Predicted resupply is the product of ruggedness and the standard deviation of ruggedness.

29 Figure 5: Martyr status and destruction .6 .4 Granted martyr status .2 0

0 .2 .4 .6 .8 1 % Destroyed

Notes: Binned scatterplot of towns granted martyr status, among towns that experienced reprisals, by percent- age of destruction.

Figure 6: Probability of being assigned martyr status, by cutoffs of destruction

Destruction > 10 %

Destruction > 20 %

Destruction > 30 %

Destruction > 40 %

Destruction > 50 %

Destruction > 60 %

Destruction > 70 %

Destruction > 80 %

Destruction > 90 %

−.5 0 .5 1 Change in probability of being assigned martyr status

Notes: The figure plots coefficient estimates and 95% confidence intervals from a regression of an indicator for martyr status on dummies for destruction above a given cutoff, as indicated on the y-axis. The dataset consists of towns that experienced reprisals in WWII.

30 Table 1: Monthly car sales by brand

Variables Mean S.D. Min Max

Aston Martin 0 1 0 7 Audi 315 251 68 1543 Bentley 0 1 0 5 BMW 383 288 90 1334 Chang’an 1 2 0 13 Chevrolet 170 176 0 698 Chongqi 1 1 0 4 Chrysler 77 81 0 361 Citroen 421 266 87 1358 Dacia 37 27 0 99 Daihatsu 164 220 0 783 Ferrari 1 1 0 7 Fiat 759 487 223 2513 Ford 684 513 143 2087 General Motors 19 29 0 129 Honda 204 219 7 1025 Hongxing 1 3 0 16 Hyundai 662 575 93 2524 Jaguar 15 25 0 99 Jeep 1 2 0 12 Jiangling Landwind 0 0 0 2 Kia 253 259 33 1636 Lada 6 11 0 52 Lamborghini 0 1 0 3 Lexus 1 2 0 10 Lotus 0 1 0 2 Maserati 0 1 0 4 Mazda 194 285 0 1081 Mercedes 494 439 51 1780 Mitsubishi 157 165 8 675 Nissan 576 413 86 1998 Opel 969 604 312 2806 Peugeot 385 325 19 1319 Porsche 13 18 0 67 Renault 191 123 39 723 Saab 24 36 0 194 Seat 333 288 45 1227 Shuanghuan Auto 2 4 0 18 Skoda 417 278 99 1439 Smart 5 10 0 63 Ssnagyong 7 12 0 57 Subaru 48 63 0 251 Suzuki 535 493 65 1896 Toyota 1162 837 235 3909 Volkswagen 967 596 220 2436 Volvo 135 63 30 319

Notes: Source is ELSTAT. All car producers with an average of at least 10 car registrations per month. Data for the period January 2008 to December 2014.

31 Table 2: Summary statistics of main variables

Variables Mean S.D. Min Max N

Cross-section

Reprisals in prefecture (0/1) 0.706 0.460 0 1 51 Share towns with reprisals 0.052 0.071 0 0.313 51 Share martyr towns 0.017 0.028 0 0.126 51

Monthly series

Article share 0.062 0.039 0.008 0.182 84

Panel

Total car sales 211 973 0 16365 4284 Share German cars 0.259 0.135 0 1 4243

32 Table 3: Balancedness

Variable All Non-reprisal Reprisal Difference Non-martyr Martyr Difference

Log population 12.336 11.989 12.480 -0.491∗∗ 12.562 12.273 0.289 (0.881) (0.723) (0.909) (0.241) (1.256) (0.757) (0.397) Share agriculture 0.264 0.256 0.267 -0.0118 0.276 0.261 0.0155 (0.107) (0.109) (0.108) (0.0334) (0.123) (0.104) (0.0404) Share industry 0.219 0.208 0.223 -0.0150 0.218 0.219 -0.001 (0.058) (0.046) (0.063) (0.0158) (0.050) (0.061) (0.018) Share civil servants 0.014 0.015 0.013 0.00181 0.013 0.014 -0.001 (0.004) (0.005) (0.003) (0.00133) (0.003) (0.004) (0.001) Share secondary education 0.179 0.170 0.183 -0.0122 0.173 0.181 -0.0079 (0.031) (0.023) (0.034) (0.00816) (0.040) (0.029) (0.0129) Share higher education 0.110 0.105 0.112 -0.00654 0.110 0.110 -0.0001 (0.024) (0.019) (0.026) (0.00659) (0.033) (0.021) (0.0105) Unemployment rate 0.122 0.127 0.119 0.00805 0.116 0.123 -0.0069 (0.029) (0.033) (0.027) (0.00977) (0.017) (0.032) (0.0073) Income p.c. (2008-2009) 18,378 17,993 18,538 -513.3 18,604 18,316 258.2 (4,515) (3,609) (4,836) (314.3) (5,965) (4,028) (541.1) Income p.c. growth (2008-2009) -0.004 -0.004 -0.005 0.0010 -0.006 -0.004 -0.002 (0.0189) (0.0153) (0.020) (0.0015) (0.025) (0.017) (0.0023) Share German cars 0.234 0.199 0.249 -0.0497∗∗∗ 0.240 0.233 0.0073 (0.085) (0.085) (0.080) (0.005) (0.088) (0.084) (0.006) First diff. share German cars pre-2010 0.003 0.0001 0.0041 -0.0040 0.0053 0.0024 0.0029 (0.099) (0.101) (0.0985) (0.0090) (0.0945) (0.1004) (0.0094) Log population in 1940 11.645 11.437 11.726 -0.289 11.680 11.635 0.0447 (0.619) (0.525) (0.641) (0.176) (0.934) (0.514) (0.293) Share communist seats 1936 0.028 0.037 0.025 0.0125 0.010 0.033 -0.023 (0.057) (0.065) (0.055) (0.0190) (0.001) (0.062) (0.0143) Ruggedness 248.92 227.03 258.05 -31,02 247.73 249.25 -1.529 (77.25) (95.75) (67.58) (27.17) (54.00) (83.09) (20.92) Average distance 1940 road 15.31 33.24 7.84 25.40 14.11 15.64 -1.532 (35.42) (57.81) (16.13) (15.17) (28.88) (37.34) (10.52) Average distance 1940 railway 78.17 96.88 70.37 26.51 59.06 83.42 -24.36 (92.13) (103.94) (87.13) (30.51) (52.58) (100.21) (22.41) Observations 51 15 36 11 40

33 Table 4: Baseline results

Dep. Variable Share German cars (1) (2) (3) (4)

Article share -0.021 -2.298 -1.222 (0.109) (5.847) (6.375) Article share × Share towns -1.491∗∗ -2.996∗∗ -3.016∗∗ -2.992∗∗ (0.734) (1.288) (1.306) (1.312)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Year FE !!!! Prefecture FE !!!! Calendar month FE !!!! Pre-controls × Article share !!! Prefecture × Calendar month FE !! Time FE × Article share !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Share towns is the share of a prefecture’s towns that experienced reprisals by German occupying forces. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

34 Table 5: Accounting for the crisis

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: Monthly unemployment

Article share -0.033 -2.307 -1.226 (0.106) (5.857) (6.385) Article share × Share towns -1.559 -3.057∗∗ -3.127∗∗ -3.136∗∗ (0.932) (1.474) (1.500) (1.508) Monthly unemp. rate 0.140 0.142 0.139 (0.270) (0.270) (0.290) Monthly unemp. rate × Share towns 0.048 0.043 0.076 0.099 (0.587) (0.591) (0.653) (0.653)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Panel B: Nightlight density

Panel B1: Baseline regressions for sample with non-missing luminosity data

Article share -0.169 -6.725 -6.248 (0.140) (6.653) (7.129) Article share × Share towns -0.760 -2.537 -2.384 -2.369 (0.883) (1.593) (1.630) (1.641)

Observations 1,972 1,972 1,972 1,972 R-squared 0.247 0.259 0.383 0.430

Panel B2: Controlling for monthly prefecture-level luminosity (log)

Article share -0.163 -6.614 -6.150 (0.140) (6.704) (7.180) Article share × Share towns -0.782 -2.559 -2.398 -2.368 (0.884) (1.591) (1.628) (1.638) Log(Nighttime light density) -0.004∗∗ -0.004∗∗ -0.003 0.000 (0.002) (0.002) (0.002) (0.003)

Observations 1,972 1,972 1,972 1,972 R-squared 0.248 0.260 0.383 0.430

Panel B3: Controlling for monthly prefecture-level luminosity (level)

Article share -0.140 -7.360 -7.124 (0.139) (6.699) (7.170) Article share × Share towns -0.784 -2.576 -2.399 -2.376 (0.868) (1.597) (1.635) (1.643) Nighttime light density -0.004∗∗ -0.004∗∗ -0.005∗∗∗ -0.002 (0.002) (0.002) (0.002) (0.002)

Observations 1,972 1,972 1,972 1,972 R-squared 0.250 0.261 0.386 0.430

Panel C: Placebo – Economic policy uncertainty index from Google trends

EPA Index 0.033∗∗ 0.033∗∗ 0.031∗∗ (0.014) (0.014) (0.016) EPA Index × Share towns -0.187 -0.180 -0.118 -0.125 (0.187) (0.187) (0.243) (0.246) Article share × Share towns -1.713∗∗ -3.082∗∗ -3.099∗∗ -3.125∗∗ (0.704) (1.286) (1.312) (1.306)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Pre-controls × Article share !!! Time FE !!

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008?2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Share towns is the share of a prefecture’s towns that experienced reprisals by German occupying forces. EPA Index is the average monthly Google trends index for the terms ‘economy’ or ‘economic’, ‘policy’ and ‘uncertain’ or ‘uncertainty’. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

35 Table 6: Survey – Main outcomes

Dep. Variable German car German ideal car Blame Germany (1) (2) (3) (4) (5) (6)

%Destruction -0.030 -0.033 -0.077∗∗ -0.127∗∗∗ 0.033 0.041∗∗ (0.027) (0.027) (0.030) (0.027) (0.022) (0.016)

Observations 910 851 910 851 844 797 R-squared 0.001 0.067 0.005 0.087 0.004 0.042

Controls !!!

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. % Destruction is the share of the town destroyed during WWII. German car is an indicator for respondents who own a German car. German ideal car is an indicator for respondents who report their ideal car is German. Blame Germany is an indicator for respondents who name Germany as the main actor responsible for the Greek crisis. Controls include indicators for gender, 6 age categories, 4 educational categories, 7 income categories and prefecture fixed effects. Standard errors clustered at the town level (N= 142) in parentheses.

Table 7: Survey – Placebo outcomes

Dep. Variable National Pride Lost job Demonstration Strike (1) (2) (3) (4) (5) (6) (7) (8)

%Destruction -0.050 0.014 -0.012 -0.001 -0.014 -0.026 -0.015 -0.059 (0.054) (0.053) (0.025) (0.024) (0.048) (0.029) (0.044) (0.054)

Observations 905 848 904 848 907 849 906 848 R-squared 0.001 0.033 0.000 0.137 0.000 0.117 0.000 0.142

Controls !!!!

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. % Destruction is the share of the town destroyed during WWII. National Pride is an indicator for respondents who agree with the statement “I am proud to be Greek.” Lost job is an indicator for respondents who report having lost their job within the last five years. Demonstration and strike are indicators for respondents who report ever having participated in a demonstration and a strike, respectively. Controls include indicators for gender, 6 age categories, 4 educational categories, 7 income categories and prefecture fixed effects. Standard errors clustered at the town level (N= 142) in parentheses.

36 Table 8: Instrumental variable analysis

Dep. Variable Share German cars Article share × Share towns Share German cars Share German cars (1) (2) (3) (4) OLS First stage Reduced form 2SLS

Article share × Share towns -2.992∗∗ -6.094∗∗ (1.312) (2.636) Article share × Predicted resupply 0.002∗∗∗ -0.015∗∗∗ (0.001) (0.006)

Observations 4,243 4,284 4,243 4,243 R-squared 0.391 0.778 0.233 0.230 F-stat 18.03 18.03

Prefecture × Calendar month FE !!!! Pre-controls × Article share !!!! Time FE !!!!

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Share towns is the share of a prefecture’s towns that experienced reprisals by German occupying forces. See text for the construction of Predicted resupply. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

37 Table 9: Falsification tests

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: Italian and Bulgarian reprisals

German article share -0.118 -0.360 0.935 (0.108) (6.656) (7.239) German article share × Share Italian/Bulgarian 0.363 0.046 0.070 0.078 (0.529) (0.694) (0.792) (0.798)

Observations 4,159 4,159 4,159 4,159 R-squared 0.258 0.266 0.352 0.390

Panel B: Conflict with Italy

Italian article share -0.083 -0.002 0.014 (0.189) (0.226) (0.241) Italian article share × Share towns -2.175 -2.759 -3.038 -2.946 (1.697) (2.061) (2.346) (2.353)

Observations 4,243 4,243 4,243 4,243 R-squared 0.257 0.265 0.352 0.390

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Italian article share is constructed out of articles containing the stem “Italian-” and the words in the German news index, substituting “Merkel” with “Renzi”. Share towns is the share of a prefecture’s towns that experienced reprisals by German occupying forces. Share Italian/Bulgarian is the share of a prefecture’s towns that experienced reprisals by Italian or Bulgarian occupying forces. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

Table 10: Regression discontinuity design – Martyr status

Dep. Variable Granted martyr status Population 1940 Population 2010 (1) (2) (3) (4) (5) (6) (7) (8) (9) Coef. 0.433∗∗ 0.533∗∗ 0.442∗∗∗ 608.182 628.380 116.278 340.609 667.123 -57.265 P-value 0.029 0.034 0.027 0.580 0.822 0.899 0.543 0.959 0.975

Observations 39 67 135 32 60 115 32 60 112 Polynomial Linear Quadratic Quadratic Linear Quadratic Quadratic Linear Quadratic Quadratic Bandwidth CCT CCT IK CCT CCT IK CCT CCT IK

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. P-values are robust p-values, computed by the rdrobust routine, as described in Calonico et al. (2017). CCT refers to Calonico et al. (2017) optimal bandwidth and IK is the Imbens-Kalyanaraman optimal bandwidth.

38 Table 11: The effect of martyr status

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: Baseline

Article share -0.021 -2.298 -1.222 (0.109) (5.847) (6.375) Article share × Share towns -1.491∗∗ -2.996∗∗ -3.016∗∗ -2.992∗∗ (0.734) (1.288) (1.306) (1.312)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Panel B: Martyr towns

Article share -0.006 -0.765 0.602 (0.107) (5.895) (6.522) Article share × Share martyr towns -6.340∗∗ -9.627∗∗∗ -8.820∗∗∗ -8.763∗∗∗ (2.394) (2.951) (2.930) (2.912)

Observations 4,159 4,159 4,159 4,159 R-squared 0.261 0.270 0.355 0.393

Panel C: Horse race

Article share -0.014 -1.831 -0.699 (0.112) (5.877) (6.457) Article share × Share towns -0.031 -2.279 -2.805 -2.829 (2.020) (1.999) (2.223) (2.237) Article share × Share martyr towns -7.037∗∗ -8.469∗∗∗ -7.515∗∗ -7.499∗∗ (2.945) (3.126) (3.179) (3.179) Article share × Mean destruction 0.782 2.270 2.970 3.067 (3.650) (3.775) (4.240) (4.252)

Observations 4,159 4,159 4,159 4,159 R-squared 0.261 0.270 0.355 0.394

Panel D: Martyr status predicted by RDD

Article share -0.023 -2.539 -1.198 (0.112) (5.974) (6.479) Article share × Mean destruction -1.055 1.040 1.367 1.488 (2.275) (2.494) (2.782) (2.767) Article share × Predicted share martyr -2.579 -8.308∗∗∗ -8.411∗∗∗ -8.485∗∗∗ (2.397) (2.751) (3.047) (3.048)

Observations 4,159 4,159 4,159 4,159 R-squared 0.259 0.269 0.355 0.393

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. See Section 4.3 for the construction of Predicted share martyr. Share towns is the share of a prefecture’s towns that experienced reprisals by German occupying forces. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

39 Table 12: Memory and official recognition

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: Memorials

Article share -0.017 -2.720 -1.409 (0.111) (5.844) (6.307) Article share × Share memorials -2.179∗∗ -5.078∗∗∗ -5.097∗∗∗ -5.064∗∗∗ (1.009) (1.734) (1.744) (1.749)

Observations 4,159 4,159 4,159 4,159 R-squared 0.260 0.270 0.355 0.394

Panel B: Memorials and martyr status

Article share -0.012 -2.445 -1.344 (0.112) (5.185) (5.640) Article share × Share towns 2.982 10.107∗∗ 10.446∗∗ 10.400∗∗ (5.048) (4.300) (4.653) (4.684) Article share × Share memorials (non-martyr) -4.005 -16.897∗∗∗ -18.070∗∗∗ -18.039∗∗∗ (6.341) (5.603) (5.986) (6.040) Article share × Share martyr -9.671∗ -20.752∗∗∗ -20.654∗∗∗ -20.615∗∗∗ (5.527) (5.436) (5.718) (5.739) Article share × Mean destruction 0.471 0.954 1.552 1.650 (3.690) (3.046) (3.352) (3.354)

Observations 4,159 4,159 4,159 4,159 R-squared 0.261 0.272 0.357 0.396

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Share memorials is the share of a prefecture’s towns that commemorate the German occupation. Share memorials (non-martyr) is the share of a prefecture’s towns that commemorate the German occupation, but do not have martyr status. Share martyr is the share of a prefecture’s towns that have martyr status. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

40 A Additional Figures and Tables

Figure A.1: Lexis-Nexis articles jointly mentioning “anti-German” and “Greece” 30 20 Number of articles 10 0 Jul 2007 Jan 2009 Jul 2010 Jan 2012 Jul 2013 Jan 2015 Date

Notes: Number of international newspaper articles in English that mention the words “anti-German” and “Greece”. Data from LexisNexis.

Figure A.2: Economic policy uncertainty from Google trends 80 60 40 Index 20 0 Jan 2008 Jan 2010 Jan 2012 Jan 2014 Date

Notes: The index is an average of monthly search volume for the terms (in Greek): “economy”, “economic”, “policy”, “uncertainty”, “uncertain”.

1 Figure A.3: Partisan air drop supply points and landing locations, original scan

Source: Military maps of Heeresgruppe E (Army Group E), US National Archives. Note: The German title reads ”Resupply of bandit groups by sea and air”. Sets of adjacent round holes along the creases are from hole punchers used in the archives, and do not indicate supply points.

2 Figure A.4: Partisan air drop supply points

Figure A.5: IV estimation – Robustness to dropping outliers −2 −4 −6 Marginal effect and 90% CIs −8 −10 0 10 20 30 40 50 Prefectures ranked by population

Source: The figure plots coefficient estimates and 90% confidence intervals from the specification in Column 4 of Table 8, dropping one prefecture at a time.

3 Figure A.6: Assessing the degree of spatial correlation in the data

Source: The figure plots Moran’s I for the residuals resulting from the specification in Column 4 of Table 4 for every month in the dataset. The spatial weighting scheme assigns equal weight on all prefectures. The vertical line is drawn at 1.96.

Figure A.7: Simulating spatial noise

Source: The figure plots the distribution of p-values resulting from 1,000 simulations of the specification in Column 4 of Table 4, where Share towns has been replaced by generated spatial noise. The vertical line indicates the actual p-value from Column 4 of Table 4. Spatial noise is correlated according to the Matern function, with a variance and shape of 1 and a correlation range of 30 degrees, following Kelly (2019).

4 Table A.1: Key events – Greek sovereign debt crisis

Date Events

Oct 2009 Center-left party PASOK wins parliamentary elections Oct-Dec 2009 Greek budget deficit reaches 12.5% of GDP Consecutive downgrades of Greece’s credit ratings Feb 2010 First austerity package Mar 2010 Second austerity package Apr 2010 First bailout; Greece’s sovereign debt downgraded to junk bond status Greek consumer association calls for boycott of German products May 2010 Third austerity package passed amidst country-wide riots and strikes Oct 2010 Brussels EU summit accepts German-inspired new bailout mechanism Jan 2011 Greek case against Germany for WWII reparations heard in Den Haag Sep 2011 German finance minister Wolfgang Schaeuble suggests Greece may want to leave the Euro Germany refuses extension of repayment terms for Greek loan Oct 2011 Haircut on Greek debt announced Nov 2011 Prime minister Papandreou resigns Feb 2012 Second bailout package May- Jun 2012 Parliamentary elections Nov 2012 7th austerity package passed; massive protests and unrest outside the parliament Den Haag court rules in Germany’s favor in reparations case Jan 2013-Nov 2014 Relative political and economic stability after the EU/IMF bailout and haircut Dec 2014 Failure to elect new Greek president results in new elections Jan-Jul 2015 Left-wing party Syriza elected Jul 2015 Finance minister Varoufakis resigns Varoufakis criticizes German-led dominance of the EU Continuous conflict between Greece and Germany

5 Table A.2: Summary statistics, survey

Variable Non-martyr Martyr Difference

Age 52.709 52.435 0.275 (15.551) (15.189) (1.009) Female 0.647 0.617 0.0300 (0.478) (0.487) (0.0317) Lost job in last 5 years 0.135 0.132 0.00361 (0.342) (0.338) (0.0224) Year car bought 2005.7 2005.8 -0.111 (6.702) (7.145) (0.551) Education: high school or higher 0.848 0.846 0.00264 (0.359) (0.362) (0.0237) Education: university or higher 0.291 0.272 0.0189 (0.455) (0.445) (0.0295) Income bracket (1=under 500, 7=over 5000) 0.857 0.842 0.0147 (0.500) (0.515) (0.0345) % destruction 0 0.658 -0.658∗∗∗ (0.000) (0.371) (0.0173)

Observations 468 460

Table A.3: Dropping expensive cars

Dep. Variable Share German cars (1) (2) Sample of brands All cars VW category Article share -1.222 -4.384 (6.375) (7.189) Article share × Share towns -3.016∗∗ -2.880∗ (1.306) (1.711)

Observations 4,243 4,220 R-squared 0.353 0.326

Pre-controls × Article share !! Prefecture × Calendar month FE !! Time FE × Article share !!

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. VW Category includes the following brands: Volkswagen, Opel, Citroen, Ford, Honda, Hyundai, Nissan, Peugeot, Renault, Seat, Skoda and Toyota. Standard errors are clustered at the prefecture level.

6 Table A.4: Pairwise correlations, instrument

Predsupply Dsupply Share towns Predsupply 1.0000 Dsupply -0.1950∗∗∗ 1.0000 (0.0000) Share towns 0.2636∗∗∗ -0.3098∗∗∗ 1.0000 (0.0000) (0.0000)

Notes: P-values in parentheses. Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Predsupply is the interaction of average ruggedness and the standard deviation of ruggedness in a prefecture. Dsupply is a prefecture’s average distance from partisan supply points. Share towns is the share of a prefecture’s 1940 towns that experienced reprisals in WWII.

Table A.5: Dropping famous reprisals

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: Drop Distomo, Kalavryta, Kommeno

Article share -0.019 -2.248 -1.178 (0.108) (5.761) (6.280) Article share × Share towns -1.547∗∗ -3.125∗∗ -3.157∗∗ -3.134∗∗ (0.739) (1.298) (1.317) (1.323)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Panel B: Drop entire prefectures

Article share -0.021 -2.298 -1.222 (0.109) (5.847) (6.375) Article share × Share towns -1.491∗∗ -2.996∗∗ -3.016∗∗ -2.992∗∗ (0.734) (1.288) (1.306) (1.312)

Observations 4,243 4,243 4,243 4,243 R-squared 0.258 0.267 0.353 0.391

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Panel A drops Distomo, Kalavryta and Kommeno before aggregating information at the prefecture level. Panel B drops the prefectures containing these three towns. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

7 Table A.6: Google index

Dep. Variable Share German cars (1) (2) (3) (4)

Google index -0.001 -0.001 -0.001 (0.000) (0.000) (0.000) Google index × Share towns -0.006∗∗ -0.007∗∗ -0.008∗∗ -0.008∗∗ (0.002) (0.003) (0.003) (0.003)

Observations 4,243 4,243 4,243 4,243 R-squared 0.260 0.269 0.355 0.391

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. The dependent variable is the monthly German share of new cars registered in a prefecture. Google Index is the average monthly search index for the terms “Germans”, “German reparations” and “Distomo”, compiled from Google Trends. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

8 Table A.7: Martyr status predicted by discontinuity, narrow range

Dep. Variable Share German cars (1) (2) (3) (4)

Panel A: At least 50% of towns in 0.25-0.75 destruction range

Article share 0.185 -50.527∗∗∗ -46.179∗∗∗ (0.250) (0.287) (0.310) Article share × Mean destruction -0.184 6.663∗∗∗ 7.590∗∗∗ 7.623∗∗∗ (0.150) (0.008) (0.007) (0.022) Article share × Predicted martyr status 0.151 -3.172∗∗∗ -4.552∗∗∗ -4.559∗∗∗ (0.157) (0.002) (0.001) (0.004)

Observations 831 831 831 831 R-squared 0.181 0.203 0.270 0.367

Panel B: At least 65% of towns in 0.25-0.75 destruction range

Article share 0.485 0.226 0.486 (0.444) (0.425) (0.460) Article share × Mean destruction -0.400 0.843∗∗∗ 0.949∗∗∗ 0.973∗∗∗ (0.506) (0.005) (0.005) (0.015) Article share × Predicted martyr status -0.446 -1.273∗∗∗ -1.366∗∗∗ -1.368∗∗∗ (0.546) (0.000) (0.000) (0.001)

Observations 495 495 495 495 R-squared 0.152 0.166 0.219 0.355

Panel C: All towns in 0.25-0.75 destruction range

Article share -0.299 0.794∗∗ 0.906∗∗ (0.343) (0.264) (0.298) Article share × Mean destruction 0.632∗∗ 0.275∗∗ 0.252∗∗∗ 0.252∗∗∗ (0.153) (0.000) (0.000) (0.000) Article share × Predicted martyr status -0.633∗ -0.764∗∗∗ -0.799∗∗∗ -0.799∗∗∗ (0.231) (0.000) (0.000) (0.000)

Observations 420 420 420 420 R-squared 0.306 0.316 0.373 0.548

Pre-controls × Article share !!! Prefecture FE × Calendar month FE !! Time FE !

Notes: Significance levels: ∗∗∗ p< 0.01, ∗∗ p< 0.05, ∗ p< 0.1. Years 2008–2014. . The dependent variable is the monthly German share of new cars registered in a prefecture. Article share is the share of conflict related articles in Kathimerini. Predicted martyr status is the share of towns predicted to have martyr status by a regression controlling for the discontinuity at 50% destruction and a quadratic polynomial in destruction in the town-level dataset, as described in Section 4.3. The sample in each panel consists of prefectures with a minimum share of towns in the 0.25-0.75 destruction range, as indicated in the title. All regressions control for year, prefecture and calendar month fixed effects. Prefecture controls include log population in 2001, share employed in agriculture in 2001, share employed in industry in 2001, share with higher education in 2001, share with secondary education in 2001, unemployment rate in 2001, share civil servants in 2001, log GDP per capita in 2008-2009, GDP per capita growth in 2008-2009, ruggedness, average distance from the nearest road in 1940, average distance from the nearest railway line in 1940 and share of parliamentary seats to communists in 1936. Standard errors are clustered at the prefecture level.

9 B Possible Issues and their Solutions

Concern Solutions

German cars are luxury This effect would have to be more pronounced in reprisal prefec- products and likely to be tures. We show below that reprisal prefectures do not differen- affected more in a downturn tially experience the economic impact of the crisis. Additionally:

1. Dropping luxury cars from the sample does not materially affect estimates (Table A.3).

2. Survey respondents in martyr towns are less likely to name German cars as their ideal cars (Figure 3 and Ta- ble 6).

Economic effect of the crisis The effect would have to differ by reprisal status. This is unlikely could directly reduce German for the following reasons: market share 1. Throughout, we control for pre-crisis economic controls and the conflict article share.

2. Results are robust to controlling for prefecture-level per- formance through an interaction of national-wide unem- ployment and prefecture reprisal status, prefecture-level nightlight density, and a national-wide policy uncertainty index interacted with the share of reprisal towns (Table 5).

3. Survey respondents in martyr towns are no more likely to have become unemployed during the crisis (Tables A.2 and 7) and do not differ in terms of income (Table A.2).

10 Concern Solutions

Results reflect the persistence of local attitudes (such as 1. Prefectures with and without reprisals are not different in nationalism), and not the effect terms of historical political attitudes (Table 3). of German wartime massacres 2. Results are robust to an IV strategy that predicts the lo- cation of reprisals from geographical features associated with partisan activity (Table 8).

3. Survey respondents in martyr towns are not more nation- alistic or prone to collective action today (Table 7).

4. Falsification test shows that prefectures that experienced reprisals by Italians and Bulgarians do not show backlash against German products (Table 9).

5. Martyr status assigned exogenously based on discontinu- ity at 50% destruction (Figures 5 and 6 and Table 10) pre- dicts reactions to German-Greek conflict (Tables 11 and A.7).

Kathimerini reporting may The extent to which there is local interest in German-Greek con- reflect local sentiment, but not flict is also a proxy of collective memory of Germany’s war the extent of German-Greek crimes and thus part of our mechanism. Additionally, the pat- conflict tern of conflict similar in LexisNexis results which come from international news articles (Figure A.1).

The measure of German-Greek conflict is just a measure of 1. Results are robust to using search volume of words un- economic policy uncertainty or related to crisis (‘Germans’, ‘German reparations’, ‘Dis- magnitude of the crisis. Any tomo’) from Google trends (Table A.6). other measure of political conflict during the debt crisis 2. Results are robust to directly controlling for the differen- will generate the same effects, tial effect of economic policy uncertainty (Table 5). The as it just proxies for economic measure of uncertainty itself does not follow the pattern uncertainty and lower output of the conflict article share (Figure A.2). in the future. 3. A share of news about Italy (replacing ‘German’ with ‘Ital- ian’ and ‘Merkel’ with ‘Renzi’) does not generate signifi- cant results, despite overlap in terms referring to crisis (Table 9).

11 Concern Solutions

Results are just driven by a handful of famous massacres. 1. Results are robust to dropping famous reprisal towns or the entire prefectures that contain them (Table A.5).

2. IV results are not sensitive to dropping outliers (Fig- ure A.5).

Reprisal and non-reprisal Table 3 shows this not to be the case. The samples are balanced prefectures are too different to on all characteristics, with the exception of pre-crisis German be compared. market share.

Significance inflated due to Figure A.6 shows that spatial correlation is not an issue in our spatial correlation. data. Figure A.7 shows that spatial noise outperforms in ex- planatory power our main geographic variable (Share towns) less than 5% of the time.

12 C Data Appendix

C.1 News Index Construction

For the construction of the index, we use the 2008–2014 Kathimerini archive of the sections “Greece”, “Politics” and “Economy”, containing a total of 101,889 articles. First, we sample 10% of articles containing the stem “German-” and manually classify them into relevant and non-relevant to political tension between Germany and Greece. An article classified as relevant must contain a reference, however short, to German–Greek conflicting political interests or political interactions in the context of foreign relations, the Eurozone or the issue of German war reparations. Articles that refer to German-Greek relations in another context – e.g. tourism flows, economic transactions between German and Greek firms etc. – are classified as non-relevant. We split the audited sample into a training and test set and use the test set to evaluate the classification performance of our algorithm. We start by assigning a frequency score to each term appearing more than twenty times in the articles of the training set. This score captures how frequently the term appears in conflict-related articles relative to non- conflict-related articles. The frequency score for term i is an empirical index constructed as:

Pr(i|c) Frequency score = ∗ 100 i Pr(i|c) + Pr(i|nc)

where c is a conflict-related article and nc an unrelated article. The score takes on the value 100 when a term appears only in conflict-related articles and the value 0 when it appears only in unrelated articles. Terms like bankruptcy, memorandum, austerity, but also subtler terms like discipline, painful, tolerance, score above 95 in this index. We use the 40 highest-scoring terms and form all combinations of 5 to 6 terms from this list. For each combination, we classify an article in our test set as conflict-related if it contains at least one of the terms in the combination. We then compare this classification to the human audit and evaluate each combination of terms based on a compound measure of precision and recall known as the F1-score. Since we are interested in minimizing both false positive and false negative classifications of articles, we put equal weight on the two types of errors and pick the combination that maximizes:

2 ∗ true positive F = 2 ∗ 1 2 ∗ true positive + false negative + false positive

Based on the above procedure, we end up classifying an article as related to German- Greek conflict if it contains the stem “german-” and at least one of the words in the set {haircut, summit, Merkel, troika, Eurozone}. This gives us a monthly count of conflict- related articles, which we normalize by the total number of articles Kathimerini published

13 in the month.

14 C.2 Variable Descriptions

Variable Description and source

Share German cars Monthly share of German-manufactured cars in a prefecture?s total new car registrations. Source: Ministry of Transportation and Communications and ELSTAT. Article share Monthly Kathimerini articles relevant to German–Greek conflict, normalized by the total number of Kathimerini articles in the month. For details on the selection of the relevant articles see Appendix B.1.1. Source: Kathimerini electronic archive 2008– 2014, sections on “Greece”, “Politics” and “Economy”. EPA Index Average monthly search volume for the terms “economic”, “economy”, “policy”, “uncertain” and “uncertainty” in the pe- riod 2008?2014, for the geographic area of Greece. The index is a normalization of the share of total searches represented by each term in a given time and region and ranges between 0 and 100. Source: Google Trends. Google Index Average monthly search volume for the terms “Germans”, “Ger- man reparations” and “Distomo” in the period 2008?2014, for the geographic area of Greece. The index is a normalization of the share of total searches represented by each term in a given time and region and ranges between 0 and 100. Source: Google Trends. Share towns Share of a prefecture’s towns in 1940 that experienced reprisals by the German occupying forces during WWII. Original list of 396 towns that experienced reprisals compiled from the follow- ing sources: Destroyed towns and villages owing to the War of 1940-1945, Ministry of Social Welfare (1946), Tables of Destruc- tion of Buildings of Greece, Ministry of Reconstruction (1947), Re- port on Atrocities in Crete (Iraklion, 1983), Greek Holocausts 1940- 1945 (Athens, 2010), Meyer (2002), Hagen (1979), Fotiou (2008). Additional information collected from primary archival work in the Libraries of the Hellenic Parliament and the Archives of the Committee of article 18 of the Law 2508/97 at the Ministry of the Interior in Greece. Mean destruction Average destruction in a prefecture during WWII. Estimates of destruction (in %) assigned to each town in our list of reprisal towns from the sources stated above. Towns not in the list are assigned destruction of zero. Share 1940 pop. Share of a prefecture’s 1940 population in reprisal towns. Source: 1940 Greek census.

15 Variable Description and source

Share martyr towns Share of prefecture?s towns awarded martyr status. Sources: Archive of the Committee on martyr towns, Presidential decrees 399(1998), 99(2000), 40(2004), 140(2005), 111(2009), 203(2012)139(2014), 79(2017), Law 2130(1993). Share memorials Share of a prefecture’s towns that commemorate the German occupation, with ceremonies or physical memorials. Original phone survey of mayors’ offices for the list of reprisal towns. Share Italian/Bulgarian Share of a prefecture’s towns in 1940 that experienced reprisals by the Italian or Bulgarian occupying forces during WWII. Sources used in the construction of Share towns. Population Source: 2001 Greek census. Share agriculture Source: 2001 Greek census. Share industry Source: 2001 Greek census. Share civil servants Source: 2001 Greek census. Share secondary education Source: 2001 Greek census. Share higher education Source: 2001 Greek census. Unemployment rate Source: 2001 Greek census. Income p.c. Source: Eurostat. Population in 1940 Source: 1940 Greek census. Share seats to communists The share of a prefecture’s seats allocated to the coalition of the Greek Communist Party and the Greek Agrarian Party (Pallaiko Metopo) in the 1936 parliamentary elections. Source: Hellenic Parliament, Registry of Parliament Members. Ruggedness Terrain ruggedness index computed as in Riley et al. (1999) and averaged over each prefecture’s surface. The shapefile of prefecture boundaries is from ELSTAT and elevation data from GMTED2010. Average distance 1940 road To compute this measure we first compute the distance to the nearest road from the centroid of each 50×50 km grid cell in an equidistant projection and then average over each prefec- ture’s surface. We digitize a physical map of Greece?s pre-WWII road network from Doxiadis (1947). The shapefile of prefecture boundaries is from ELSTAT. Average distance 1940 rail Similarly to the above measure, we first compute the distance to the nearest railway line from the centroid of each 50×50 km grid cell in an equidistant projection and then average over each prefecture’s surface. We digitize a physical map of Greece?s pre- WWII railway network from Doxiadis (1947). The shapefile of prefecture boundaries is from ELSTAT. Nighttime light density Source: NOAA (National Centers for Environmental Observa- tion).

16 Variable Description and source

Farflat Interaction of distance to partisan supply points and differ- ence between prefecture’s average ruggedness and maximum ruggedness in sample of prefectures. Predsupply Interaction of average ruggedness and standard deviation of ruggedness, prefecture-level.

C.3 List of Martyr Towns

1. Aetos, Messinia 25. Kalami, Iraklio 49. Mournies, Lasithi

2. Agia Efthimia, Fokida 26. Kalavryta, Achaia 50. Mousiotitsa, Ioannina

3. Agii Anargiri, Lakonia 27. Kaloskopi, Fokida 51. Nea Kerdillia, Serres

4. Agios Vassilios, Iraklio 28. Kaloutas, Ioannina 52. Pefko, Iraklio

5. Amira, Iraklio 29. Kandanos, Chania 53. Pentapoli, Fokida

6. Ano Meros, Rethymno 30. Karoutes, Fokida 54. Pirgi, Kozani

7. Ano Viannos, Iraklio 31. Kato Simi, Iraklio 55. Prosilio, Fokida

8. Anogia, Rethymno 32. Kato Viannos, Iraklio 56. Pteri, Achaia

9. Arginia, Kefallonia 33. Kefalovryso, Iraklio 57. Riza, Lasithi

10. Asprageloi, Ioannina 34. Klisoura, Kastoria 58. Rizomilo, Magnisia

11. Chondros, Iraklio 35. Kommeno, Arta 59. Rodakinou, Rethymno

12. Chortiatis, Thessaloniki 36. Koxare, Rethymno 60. Rogi, Achaia

13. Damasta, Iraklio 37. Kria Vrissi, Rethymno 61. Saktouria, Rethymno

14. Distomo, Viotia 38. Lechovo, Florina 62. Sarchos, Iraklio

15. Drakia, Magnisia 39. Lidoriki, Fokida 63. Sidironero, Drama

16. Drosopigi, Florina 40. Ligiades, Ioannina 64. Sikologos, Iraklio

17. Elati, Ioannina 41. Lilea, Fokida 65. Skines, Chania

18. Emparos, Iraklio 42. Lochria, Rethymno 66. Sougia, Chania

19. Eptalofos, Fokida 43. Magarikari, Iraklio 67. Tibaki, Iraklio

20. Erimanthia, Achaia 44. Malathiros, Chania 68. Vachos, Iraklio

21. Gdochia, Lasithi 45. Manassis, Ioannina 69. Vlacherna, Arkadia

22. Gerakari, Rethymno 46. Mesovouni, Ioannina 70. Vorizia, Iraklio

23. Ipati, Fthiotida 47. Mesovouno, Kozani 71. Vounichora, Fokida

24. Kakopetro, Chania 48. Mirtos, Lasithi 72. Vrisses, Rethymno

17