ONLINE APPENDIX Persecution Perpetuated: the Medieval Origins of Anti-Semitic Violence in Nazi Germany

Home , Elections in Germany, Judensau, Regierungsbezirk

ONLINE APPENDIX Persecution Perpetuated: The Medieval Origins of Anti-Semitic Violence in Nazi Germany

Nico Voigtländer Hans-Joachim Voth

UCLA and NBER ICREA/UPF and CREi

I. Data Description

I.A. Medieval Data As described in section II in the paper, we use Germania Judaica [GJ] from Avneri (1968) as our principal source. We first establish the presence of a Jewish community based on its inclusion in GJ, volume II, which is for the period 1238-1350. Whenever later work by Alicke (2008) mentions that a Jewish community existed during this period, we use his information instead. For each town, city or village where GJ mentions pogroms, violent attacks on the Jewish population, the burning of Jews, or the wholesale extermination of the Jewish community in 1348-50, we code our dummy variable for Black Death pogroms, POG1349, as unity, and zero otherwise.

I.B. Data on 20C Violence and Anti-Semitic Attitudes We collect data on pogroms in the 1920s, on the number of Stürmer letters, on deportations, and on attacks on synagogues. Our source for all of these variables with the exception of the Stürmer letters is Alicke (2008). For pogroms in the 1920s we use the dummy POG1920 that equals 1 for cities with documented pogroms during this period. Alicke focuses on ‘positive’ information, and mentions when an event actually occurred. We set POG1920 to zero otherwise. We define a pogrom as a violent outrage against the Jewish population, involving physical violence against and/or the killings of people. Therefore, political agitation through Brandreden (incendiary speeches), attacks on Jewish shows, or the desecration of cemeteries are not coded as pogroms. Only when physical violence against at least one Jewish inhabitant is mentioned in Alicke does this variable take the value of unity.

From Alicke (2008), we also take data on the existence of a synagogue in 1933 (coded as 1 if mentioned as such, and 0 otherwise), as well as on the extent of attacks in 1938 during the ‘Night of Broken Glass’ (Reichskristallnacht). We construct two dummy variables – one for destroyed

synagogues, and one for damaged ones. For the former, we assign a value of 0 if no synagogue in use in 1933 was destroyed during Reichskristallnacht. “Destruction” occurred if the relevant building was damaged at least to an extent that it became unusable, in which cases Alicke mostly uses the term “zerstört” (destroyed). The variable then takes the value 1. We code our variable for synagogue damage in a locality as 0 if no synagogue/s in use in 1933 was damaged during Reichskristallnacht. “Damage” we define if the inventory of a synagogue was destroyed or the physical fabric of the building itself was damaged but remained intact. The variable takes the value 1 in these cases. From these two variables, we create a combined variable for synagogues destroyed or damaged.

As an additional variable, the number of published letters to the editor from the anti-Semitic newspaper Der Stürmer is included in our dataset. For the four years 1935 to 1938, all letters to the editor published in one of three different categories were counted provided their place of origin matched a locality in our dataset. The three categories are: (1.) Letters that were published as articles, e.g., a schoolgirl writing about her classmates still interacting with Jewish pupils. (2.) Letters in which individual Jews or other people still talking to/doing business with Jews are denounced. (3.) The category “mailbox” in which Der Stürmer answers questions about Jews (“how many Jews live in xyz town?”). We count letters in these three categories for all three years in our analysis, and sum them by locality.

I.C. Election Data We use election data initially collected as part of Jürgen Falter’s (1991) research project on historical elections in Germany (Hänisch 1988). The source for their database are the official statistics of the Weimar Republic (Statistik des Deutschen Reiches). The vote for each party is calculated as the ratio of the number of valid votes received by a party, divided by the total number of valid votes cast. For May 4, 1924, we analyze the results for the Deutsch-Völkische Freiheitspartei (DVFP) as well as for the Deutschnationale Volkspartei (DNVP). For the elections of May 20, 1928, September 14, 1930, and March 5, 1933, we focus on the Nationsozialistische Deutsche Arbeiterpartei (NSDAP). For these years, election data is available at the city level for municipalities with more than 2,000 inhabitants. For smaller towns, we use the observations on Restkreis (‘remaining county’), which correspond to the voting results for all towns and villages with less than 2,000 inhabitants.

For the socio-economic correlates in our section on elections, we use data from Hänisch (1988) and Falter (1991), derived from the censuses of 1925 and 1933. These allow us to control for the number of inhabitants, the percentage of the population that is Protestant, the percentage of Jews, and of blue-collar workers.

I.D. Deportations The German Federal Archives have compiled a comprehensive list of Jewish deportees during the Nazi period (Bundesarchiv 2007). We use the second, expanded and improved edition. It contains information on 159,972 individuals (Jewish or presumed Jewish by the authorities) who lived in what was considered Germany proper between 1933 and 1945. The database is available online (http://www.bundesarchiv.de/gedenkbuch/directory.html). We consulted the database by entering every single locality in our dataset into the search engine (under the query term “Wohnort”), and then recorded the number of listed deportees for the years 1933-45. A typical entry reads:

Lehmann, Helen Lehmann, Helen

geb. Mayer née Mayer, born 24.7.1876 * 24. Juli 1876 in Maikammer living in Meckenheim (Rhineland) wohnhaft in Meckenheim th Deportation: deported from Camp de Vernet on 19 of ab Camp de Vernet August to Auschwitz (extermination camp) 19. August 1942, Auschwitz, Vernichtungslager

Why do deportation rates reflect local sentiment? The paper already details some of the organizational and historical features that make this a valid interpretation. Here, we provide additional detail:

Local initiative could increase the number of deportees, but hardly ever reduce it – by, for example, criminalizing Jews, who would then become “Schutzhäftlinge” (protected prisoners), which were deported as a matter of course. Löw’s (2006) analysis of deportations to the Litzmannstadt ghetto in Lodz (in 1941) argues that “for the transports to Litzmannstadt, there were no central directives; the respective administrative units (Stellen) proceeded in selecting the deportees according to their own criteria (Ermessen).” After the chaotic first stages, there were some attempts to regularize the deportation rules. The general guideline was that “Mischlinge” (Jews with two or less Jewish grandparents), and Jews married to gentiles were not deported; elderly Jews and war veterans went to Theresienstadt (from where they were often deported further East). The last rule was only applied consistently from 1942 (Hilberg 1961). The treatment of “Geltungsjuden” was also inconsistent. “Geltungsjuden” were sometimes deported in Germany, almost never in Austria, but routinely

deported in Poland.1 Differences were often driven by local power holders’ attitudes and their interactions with the Gestapo. In the case of marriages between gentiles and Jews, there was often massive pressure on the “German” part to divorce the Jewish partner, which would then result in deportation. Again, there were large local differences in the pressure applied.

Meyer (2004) argues that “a close look at the deportation history demonstrates that local [Gestapo] units did not simply act in line with central RSHA directives, but were part of complex local and regional power structures… the extent to which [Gestapo] units forged alliances with the local party and other Reich administrative units or competed with them, were enlisted for the purposes of these other units or not, had a direct effect on the persecution and deportation process – and therefore on the ability of Jewish representatives to influence and modify the process.” (p. 63)

I.E. Crime Data for Late Imperial Germany Crime data are from Eric Johnson’s study of Imperial Germany. They are available from the GESIS data archive (study ZA8069 – German Crime, Death and Socioeconomic Data, 1871–1914). We use information on the number of convictions for causing serious bodily harm and for simple theft (at the county level).

II. Additional Empirical Results

II.A. Additional Sample Statistics Table A.1 shows the summary statistics for our main variables for the extended sample, and Table A.2 shows the corresponding correlations. Table A.3 reports the means conditional on the existence of a medieval Jewish community (Panel A) for various 20C outcome variables in the extended sample, as well as regressions of these variables on POG1349 and JewCom1349. Table A.4 shows that there is no significant difference in election turnout across cities with and without Black Death pogroms, either in the main sample or in the extended sample. Guiso, Sapienza, and Zingales (2008) use election turnout as a key indicator of social capital. The fact that there is no significant difference underlines the comparability of cities with and without pogroms along socio-economic dimensions.

II.B. Additional Control Variables for Main Regression Results In our main results in Table VI in the paper, we use an identical set of covariates for all indicators of anti-Semitic sentiment (except where we need to control for population size or the number of Jews). This has the advantage of consistency, but begs the question if other factors – such as economic

1 Jews fulfilling at least one of the following criteria: they had to be half-Jews descended from two Jewish grandparents and were either members of a Jewish religious community after 1935, married to a Jew, or the illegitimate offspring of a Jew. 4

structure – might possibly drive both the settlement of Jews, and anti-Semitic sentiment. In Table A.5, we control for additional socio-economic variables. In particular, we add data on the percentage of the population in blue-collar jobs, on industrial employment, and the percentage of self-employed in retail and trade. The latter is a proxy for the proportion of the population competing with Jews in retailing: Jewish shops and department stores were often the target of anti-Semitic assaults. For outcome variables measured in the 1930s, we add the level of unemployment in 1933 to capture the depth of the Great Depression. Finally, we also control for voter turnout, which serves as a proxy for social capital (Guiso, Sapienza and Zingales, 2008).2

There are few consistent results. In most specifications, the percentage of blue-collar workers is negatively associated with Jew-hatred, but only the NSDAP 1928 result is statistically significant. Industrial employment is also mostly positively associated with anti-Semitic attitudes, but there are no significant effects. The percentage of self-employed in retail and trade is also positively (albeit insignificantly) associated with some of our indicators, which is broadly in line with Olzak (1992) and Jha (2008), who argue that economic competition between minority and non-minority can raise the potential for conflict. Our proxy for social capital and the unemployment rate enter with varying signs. While the former is significantly negatively associated with DVFP votes in 1924, deportations, and Stürmer letters, the latter is never significant. Crucially, the size and significance of the coefficient on 1349 pogroms is never affected compared to the baseline results in Table VI in the paper.

Table A.6 shows that our baseline results are robust to controlling for regional fixed effects. For each town, we code the province that it belonged to in the Weimar Republic.3 Each province, in turn, is subdivided into prefectures (Regierungsbezirk), which we also code. Our main (extended) dataset comprises localities from 20 (24) provinces and 44 (56) prefectures. Panel A of Table A.6 shows that our main regression results, are remarkably stable when controlling for province fixed effects. The coefficients are almost identical to those reported in Table VI in the paper, and so is their statistical significance, with two exceptions. The result for the DVFP election is now significant at the 5% level; the coefficient on POG1349 for deportations has become insignificant. These results are broadly confirmed under the even more restrictive specification with prefecture fixed effects in Panel B.

2 We use the turnout in 1924 in all regressions (except for the 1928 election, where we use the corresponding voter turnout). This addresses the concern that voter turnout in later years may be endogenous to the growing success of the Nazi party, or may be driven by opportunist voters rather than ‘social capital.’ In fact, the change in voter turnout between 1924 and 1933 is strongly positively correlated with the NSDAP vote share in 1933. 3 For most parts of Germany, provinces are equivalent to states (Länder). The most important exception is Prussia, which was subdivided into 15 provinces (Provinzen). We used information from www.verwaltungsgeschichte.de 5

II.C. Alternative Specifications for 20C Outcome Variables Tables A.7-A.13 report alternative specifications for our 20th-century outcome variables. One table is dedicated to each variable, and all extended-sample regressions therein control for the existence of medieval Jewish communities (JewCom1349). Among these tables, two require a more detailed description:

First, Table A.10 (Panel B) addresses a potential concern with the deportation results reported in Table VI, column 4 in the paper: After 1933, more than half of Germany’s Jews emigrated. More anti-Semitic tendencies may have triggered more emigration before 1939, and thus fewer deportations thereafter. If this was the case, more deportations would reflect less anti-Semitism, when running regressions that control for Jewish population in 1933. To address this, we repeat the analysis, controlling instead for the remaining Jewish population in 1939.4 We find that, again, deportation rates were significantly higher in towns with Black Death pogroms.

Second, Table A.12 reports the results for an alternative specification in the deportation and Stürmer regressions. Instead of using the log-levels of these variables as dependent variables (controlling for log-population), we now use their proportions: Deportations between 1933-45 per 100 Jews in 1933 (columns 1-3), and Stürmer letters per 10,000 inhabitants (columns 4-6) at the city level.5 All three specifications (OLS, propensity score matching based on controls, and geographical matching) yield highly significant coefficients for POG1349. The estimates indicate that deportations were up to 15 percentage points higher in cities with pogroms in 1349, and there were about 0.25 additional Stürmer letters per 10,000 inhabitants. This is the same order of magnitude as obtained in Table A.10 and Table A.11.

II.D. Spatial Correlation, Sample Splits, and Additional Voting Results Table A.14 controls for spatial correlation. This complements our geography-based matching estimation in addressing the concern that unobserved local characteristics are correlated with both medieval and 20th century anti-Semitism. While the geographic matching estimation above compared nearby cities, the analysis here uses all observations and a weighting matrix that is based on each city’s geographic location. We consider cities with less than 4 degrees distance (about 440km) as ‘neighbors,’ assigning them a non-zero spatial weight. As the table shows, our earlier OLS results are confirmed: Black Death pogroms are significantly positively associated with 1920s pogroms,

4 One standard of comparison will yield deportation ratios that are too low, and the other one, ones that are too high. Since we are interested in the difference between towns with and without Black Death pogroms, we do not take a stand on which one is more correct; instead, we simply focus on the differential conditional on Black Death pogroms. 5 The regressions use city population as analytic weights: An observation where 2,000 out of 10,000 Jews were deported has more informational content than 2 out of 10. The same argument applies to the proportion of Stürmer letters. 6

NSDAP votes in 1928, and DVFP votes in 1924.6 The magnitude of the coefficients is also very similar to the baseline results.

Table A.15a – Table A.15c report results for various sample splits. In the baseline regressions, we have controlled directly for the percentage of Protestants in our regressions. We also want to know if medieval anti-Semitism was perpetuated in the same way in Catholic and Protestant areas. We expect that transmission should be similar where non-electoral outcomes are concerned. We subdivide our sample depending on the share of the population that is Protestant.7 Table A.15a reports the results for our preferred specification – matching by geographic location. For predominantly Protestant areas, we find strong and significant results for every outcome variable except deportations. In Catholic areas, we obtain positive and significant results for 1920s pogroms and all of the 1930s outcome variables. The electoral results, however, are insignificant for Catholic areas. In our view, this reflects the fact that Catholics voted differently for one reason – the Zentrum party was explicitly set up to represent Catholic voters. As a minority in Germany, Catholics sought to defend their interests through a single party. As most electoral studies of Weimar Germany show, this reduced the electoral appeal of right-wing anti-Semitic parties.

Overall, results suggest that differences in religious affiliation did not significantly alter the way in which beliefs were transmitted over centuries. Results are very similar, albeit somewhat weaker, for OLS/ML-Poisson estimation and propensity score matching (Table A.15b and Table A.15c, respectively). The same pattern of broadly similar transmission over six centuries emerges if we split our sample into Eastern vs. Western cities, or into larger vs. smaller cities. While not each result in each sub-category is significant, there is no pattern of instability across sub-groups.

Table A.16 provides additional evidence that NSDAP vote shares after 1928 are weakly associated with Black Death pogroms (section V.B in the paper). We find that the effect becomes weaker in 1930 (columns 1-4) and vanishes in 1933 (columns 5-8), where all coefficients on POG1349 are negative and insignificant.

Table A.17 gives more detail on the votes lost by the DNVP in cities with Black Death pogroms in 1924. In section V.B in the paper we show that this is about 2.6 percent. This number is consistent with the various specifications in Table A.17, which shows a range of 1.5 – 3.9 percent votes lost, significant in almost all cases.

6 While being close to zero, the statistical significance of the parameters λ imply that the spatial error model is preferable to OLS in the two election regressions in columns 2 and 3. For the two dependent variables with heavily skewed distributions where ML-Poisson was our preferred specification (deportations and Stürmer letters), we do not obtain significant results. 7 Where it is above 50%, we consider an area pre-dominantly Protestant. Where the share of Catholics is above 50%, we classify it as Catholic.

II.E. Additional Analyses of Medieval Jewish Settlement and Black Death Pogroms Table A.18 addresses the potential concern that historical sources such as GJ might underreport medieval Jewish settlements without pogroms, because reports of pogroms were more likely to be recorded. While this is possible, it is not clear if it would bias our results. We address the underreporting issue as follows: We use all medieval correlates from Table VIII (in the paper) in a Probit regression to derive the predicted probability of medieval Jewish settlement. Cities where this probability is above 50% are added to the list of existing medieval Jewish communities. This gives a sample of 419 cities with actual or likely medieval Jewish settlement – the latter category comprises 94 cities, all of which had no documented Black Death pogroms. The results shown in Table A.18 confirm all of our main findings.

Table A.19 follows up on our analysis in section IV.A in the paper. We first use all medieval correlates that were significantly associated with POG1349 in columns 5 or 6 of Table VIII in the paper. These comprise indicators for free imperial cities, Staufer cities, market rights in 1349, as well as ln(age of the city in 1349). We then predict pogrom probabilities in 1349 and include the prediction in regressions with 20C outcome variables. While POG1349 remains highly significant, predicted pogrom probability is insignificant in all specifications.

Table A.20 addresses the concern that, when using the extended sample, we include cities that probably did not exist in 1349 and thus could not possibly have had a medieval Jewish community. The table reports results for the subsample of cities that were first mentioned in public records before 1349. Almost all results are identical to our main specification.

Table A.21 examines if the link between Black Death pogroms and anti-Semitism in interwar Germany is present even in towns and cities whose Jewish communities vanished completely in 1348-50.8 We demonstrate this by splitting the indicator variable for POG1349 into two parts: First, Jewish communities that vanished in 1349 as a result of attacks, and second, Jewish communities that suffered pogroms but survived. Both generally have a positive and significant effect for most dependent variables, and they are mostly indistinguishable in a statistical sense from each other. The only exception is for deportations. This suggests that 20th century anti-Semitism was broadly similar (or somewhat stronger) in cities where Jewish communities were extinguished in 1349, compared to those places where they were attacked but survived as a community.

8 To code communities as vanished, they need to be explicitly mentioned as such in GJ. For many cities, several centuries passed before Jewish communities returned. In others, Jews settled again after only a few decades. However, as we explain in section II.A. in the paper, resettlement never recreated the same density of Jewish communities as in the early 14th century. Jewish presence was transient, and major towns typically expelled the few remaining Jews in the 14th or 15th century. 8

II.F. When Transmission Failed – Additional Analysis Table A.22 investigates the robustness of our results presented in section IV.B in the paper, for the three cases where we find that transmission fails (Hanseatic; Open Cities in Southern Germany; rapidly growing cities). In the paper we used an “open” city index, which can vary between 0 and 4, depending on how many of the four “openness” characteristics a city has: incorporated by 1349, free Imperial city, market charter by 1349, or located on a navigable river. Column 1 shows that our results also hold for a dummy variables that equals 1 whenever the ‘open’ city index is larger than zero. Similarly, instead of the growth rate itself, column 2 uses a dummy for cities with above- median population growth between 1750-1933. Again, our results are confirmed. Column 3 includes an additional interaction term for cities with above-median population growth and above-median industrial employment in 1933. This is meant to differentiate between two interpretations of the negative and significant interaction term in column 2:

Was it population inflow itself that weakened subsistence, or did industrialization (and modernization) drive both weaker persistence and population inflow? The insignificant interaction term for Igrowth,ind suggests that the former is correct, with population inflow alone accounting for most of the weakening of persistence. In addition, in columns 4-6, we split the sample to verify whether persistence fails or survives in the respective smaller subsamples. We find that POG1349 is insignificant in all subsamples that indicate trading, openness, or fast-growing cities. On the other hand, POG1349 is highly significant and positive in non-Hanseatic, non-Open, and below-median growth towns. For Hanseatic cities in column 4, the insignificant result may merely reflect the relatively small sample (29 obs). However, in column 5 we also find no evidence for persistence in the much larger subsample of ‘open’ cities (174 obs), while the coefficient on POG1349 is highly significant in the smaller subsample of non-open towns (40 obs). For above- vs. below-median city growth in column 6, the two subsamples are of almost identical sizen. Altogether, these results support our finding that persistence is weaker in trading and fast-growing cities.

III. Persistence of Anti-Semitism During the Centuries Before and after the Black Death

The main text of the paper focuses on the correlation between 20th century anti-Semitism and Black Death pogroms. Here, we add information for the period before 1349, and for two indicators of anti- Semitism in early modern and 19th century Germany. We show that additional measures are strongly correlated with each other, as well as with pogroms in 1349, and with 20C outcomes. Together, the results presented in this section support our interpretation that the empirical results presented in the paper reflect persistent anti-Semitism at the local level.

III.A. Pre-Plague attacks The historical background section showed that the pogroms at the time of the Black Death were only one particularly violent example in a long history of attacks on Jews during the Middle Ages. If hatred of Jews was a locally persistent phenomenon, then the frequency of pre-plague attacks should be higher in places with pogroms in 1349. Here, we code up the number of attacks on Jews during the centuries before the Black Death. We take particular care to confirm the existence of a Jewish community during the relevant period.9 We find that pogroms before 1349 are (i) good predictors of whether an attack occurred at the time of the Black Death, and (ii) have some predictive power for the intensity of persecutions in the 20th century.

We construct the number of attacks on Jewish communities for three periods of different lengths before 1347, using data from Haverkamp (2002). Some attacks at the city level were part of larger waves of persecution during specific years. These include attacks during the crusades in 1096 and 1146, “Guter Werner” and “Rintfleisch” pogroms in 1287 and 1298, respectively, and “Armleder” attacks in 1336. In addition, numerous local pogroms are reported, often following accusations of ritual murder or host desecration. From these observations, we derive the number of attacks on Jews in a city between year y and 1347, Number POG y-1347. For y, we use 1096, 1175, and 1225. The first cut-off year includes all recorded pogroms prior to the Black Death, and in particular both crusades that led to attacks on Jews on German territory. However, there are only 16 cities with confirmed Jewish settlement before 1096. To provide results for larger samples, we use later cut-off years, when more Jewish communities were recorded: 1175 is the first year with more than 30 observations, and 1225, with more than 60. Correspondingly, we only count attacks after the respective cut-off year when calculating Number POG y-1347. In addition to counting the number of local attacks, we derive the indicator Dummy POG y-1347, which equals 1 if at least one pogrom occurred during the years y-1347, and 0 otherwise.

There is a trade-off between exploiting local variation and sample size: In terms of information at the local level, earlier cut-off years are preferable, because attacks are counted over a longer horizon. The 16 Jewish communities that existed in 1096 were attacked on average 1.38 times until 1347; the 31 communities for y=1175 were attacked .97 times, and the 67 towns for y=1225, .69 times. In Table A.23 we show that the three subsamples confirm the findings of section IV.B in the paper: The conditional means show that cities with Black Death pogroms also saw significantly more attacks before 1347.

9 This is important because strictly speaking, only cities with documented Jewish communities are a feasible control group. For example, cities with attacks during the First Crusade in 1096 can only be compared to cities with documented Jewish settlement prior to 1096. All other cities were – at least potentially – not feasible for ‘treatment:’ Without Jews, attacks could not occur. We thus think of the results here as the most conservative way of confirming the results in the paper (Table IX column 1), where we only conditioned on the existence of a Jewish community in 1349. 10

Next, we report regression results for the three subsamples in Panels A-C of Table A.24. Column 1 shows a strong and positive correlation between Black Death pogroms and the number of pre-plague attacks over the respective period. Each additional pre-plague attack increases the probability of Black Death pogroms by 9 to 19 percentage points. The coefficient estimate is very similar (but not always significant) when including medieval control variables in column 2.10 In contrast to Black Death pogroms, pre-plague attacks were mostly geographically concentrated; for example, Armleder pogroms occurred only in South-Western Germany in the wake of the Armleder peasant revolt. To address this local dimension, column 3 shows that the results are robust to controlling for province fixed effects.11 Columns 4-6 repeat the analysis for the indicator variable Dummy POG y-1347. The results confirm the findings in the first three columns. The same is true for the results reported in columns 7 and 8 – matching estimation based on medieval correlates and geographic distance, respectively.

Table A.25 shows that pre-plague pogroms are also a good predictor of 20th century anti-Semitism. Again, we report results for the three cut-off years 1096, 1175, and 1225, and include the usual 20C controls (percentage of Jews and Protestants). Both Number POG y-1347 and Dummy POG y-1347 are strongly and significantly correlated with our principal component measure. This result is robust to including province fixed effects (for the two cut-off years with larger sample size). In addition, geographic matching yields strong positive coefficients – 20C anti-Semitism is .40-.85 standard deviations higher in towns with pre-plague pogroms, as compared to towns with documented Jewish communities over the same pre-plague period.

10 We control for all variables that are significantly correlated with Black Death pogroms in Table VIII (in the paper), columns 5 and 6. 11 See Appendix II.B for details on the classification of provinces. 11

III.B. Judensau Sculptures Churches, bridges, public buildings and private homes in medieval and early modern Germany were often adorned with images of Jews performing demeaning acts with a female pig. Some of these are still in place today. We use (highly fragmentary) information on these forms of adornment, as compiled by the Zionism and Israel – Encyclopedic Dictionary (http://www.zionism- israel.com/dic/Judensau.htm) and the list compiled by the Institut für Kunst und Forschung.12 Here, we use this data as an indicator for wide-spread, publicly-acceptable form of Jew-hatred. In total, there are 21 towns and cities in our dataset where information on Judensau sculptures has survived to the present, 13 of those belong to our main dataset.13

To analyze the differences associated with these sculptures, we total up the number of attacks before 1349, and condition on the existence of a Jewish community in 1349. As Table A.26 shows, there are large differences in conditional probabilities. The attack frequency before 1349 in towns and cities with a Judensau was more than twice as high as compared to places without (column 1). Column 2 shows that a similar difference can be seen for the period of the plague. We also find that both our indicators for Hep-Hep riots (see below) are much higher for towns with anti-Semitic images in public places, with the frequency of attacks about six times higher (columns 3 and 4). Finally, our composite measure of 20th century anti-Semitism also registers much higher values in locations with a Judensau sculpture (column 5). All differences – with the exception of the last – are statistically highly significant. In Table A.27, we analyze whether the strong association between Black Death pogroms and Judensau sculptures is driven by local characteristics or regional patterns. As columns 2 and 3 show, the result remains unchanged if we control for medieval correlates and province fixed effects. The same is true when we exploit local variation by using geographical matching in column 4.

III.C. Hep-Hep Riots The Hep-Hep riots of 1819 occurred in the context of economic difficulties after the end of the Napoleonic wars. They started in Würzburg. At the same time when local merchants were agitating in favor of curtailing the activities of Jewish traders, there was extensive discussion of granting Jews full emancipation in Bavaria (Sterling 1950). When delegates to the Bavarian parliament (who had supported Jewish emancipation) returned to the city in early August, there were attacks on Jewish property, accompanied by the cry “Hep-Hep.” The police did little to restore order. Eventually, the king of Bavaria dispatched army units to quell the riots.

12 http://www.christliche-sauerei.de/thema/thema.html 13 Even if the sculpture has been removed (as it was, for example, in Kehlheim in 1945 on the orders of a US Army officer), we code this as a location with a Judensau sculpture. 12

As the news of these disturbances spread, riots occurred in many places. They spread to Frankfurt, Baden, and Saxony, to Austria as well as to the North of Germany. In some towns and cities, there were attempts at intimidation (taking the form of scrawls on the walls of Jewish homes and shops), or shouts of “Hep-Hep,” but no riots occurred. In other locations, the historical record notes that strong police presence stopped hostile crowds from sacking Jewish property (Katz 1994). We code

all places with actual riots, using the indicator variable HepHep. In addition, HepHepAll also includes those towns and cities where police or army units stopped threatening mobs from attacking Jews.

In Table A.28, we analyze the Hep-Hep riots as a conditioning variable for other forms of anti- Semitism in Germany after 1096 in our main sample. We first find that pre-plague pogroms were almost twice as frequent in places with attacks on Jews in 1819 (column 1). The frequency of pogroms at the time of the Black goes from 71 to 93 percent if a place was to have Hep-Hep riots (column 2). Column 3 confirms (of course) the strong association between the 1819 riots and Judensau sculptures that we documented in column 3 of Table A.26. Anti-Semitism in the years 1920-45 also differs strongly for places with and without Hep-Hep riots (column 4). With the exception of the last result, all differences are statistically significant.

Hep-Hep riots occurred predominantly in Southern-Central Germany, where pogroms in 1349 were also relatively frequent (see Figure 1 in the paper). Next, we check whether this regional concentration may be responsible for our results. Table A.29 shows that Black Death pogroms predict Hep-Hep attacks even after controlling for local characteristics (column 2), or province fixed effects (column 3). Finally, the highly significant coefficient from geographic matching confirms that our results are driven by differences between nearby cities, rather than by regional patterns. In terms of magnitude, a history of Black Death attacks increases the probability of attacks in 1819 by about 4 percentage points. Our results are confirmed when using the alternative measure that also counts attempted attacks on Jewish communities during the Hep-Hep riots (columns 5-8).

FIGURES

Figure A.1: Electoral results for the DVFP and NSDAP, 1924-1933 Note: The figure plots the kernel density of the vote share for the Nazi party (NSDAP) at the city/county level in 1928, 30, and 33. The data used corresponds to our extended sample – including all cities with data on interwar Germany. The vote distributions for the main sample (only towns with documented Medieval Jewish settlement) look almost identical.

TABLES

Table A.1: Descriptive statistics – Extended Sample mean Std.dev min max Obs. $ Population in 1933 23,825 139,413 138 4,449,125 1,427 % Jewish in 1933 2.31 3.18 .000047 .377 1,174 % Protestant in 1925 53.82 35.60 .00133 .992 1,427 Synagogue in 1933 .83 .37 0 1 1,386 Indicators for 20thC anti-Semitism POG1920s .027 .163 0 1 1,396 NSDAP1928 .032 .044 0 .513 1,427 DVFP1924 .067 .083 0 .593 1,427 DEPORT 112 1,582 0 55,807 1,356 STÜRMER 1.75 11.0 0 354 1,427 SYNATTACK .817 .387 0 1 1,152 Medieval Jewish settlement .276 .447 0 1 1,427 Black Death pogrom (POG1349) .181 .385 0 1 1,294

Table based on extended sample (including all cities with Jewish population in 1920/30). POG1920s is an indicator variable for pogroms in each location during the 1920s; NSDAP1928 is the vote share of the NSDAP in the May 1928 election, DVFP1924 is the vote share for the Deutsch-Völkische Freiheitspartei in May 1924; DEPORT is the number of deportees from each locality; STÜRMER is the number of anti-Semitic letters to Der Stürmer; SYNATTACK takes the value 1 if a synagogue was destroyed or damaged in the ‘Night of Broken Glass’ in 1938, and 0 otherwise. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. $ The largest city in the sample is Berlin, which is composed of several boroughs. It is followed by Hamburg, Cologne, and Munich.

Table A.2: Correlations between main variables – Extended Sample (1) (2) (3) (4) (5) (6) (7) (1) POG1349 1 (2) POG1920s .117*** 1 (3) DVFP1924 .262*** .403*** 1 (4) NSDAP1928 .242*** .312** .726*** 1 (5) %DEPORT .042*** .174*** -.071** -.017 1 (6) STÜRMER prop. .069*** -.001 .093*** .128*** -.010 1 (7) SYNATTACK .127*** .073** -.040 -.026 .034 -.008 1

Table based on extended sample (including all cities with Jewish population in 1920/30). P-values for pairwise correlations (weighted by city population in 1933): * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. POG1920s is an indicator variable for pogroms in each location during the 1920s; NSDAP1928 is the vote share of the NSDAP in the May 1928 election, DVFP1924 is the vote share for the Deutsch-Völkische Freiheitspartei in May 1924; %DEPORT is the percentage of deportees from each locality (relative to Jewish population in 1933); STÜRMER/pop is the number of anti-Semitic letters to Der Stürmer per 10,000 inhabitants; SYNATTACK takes the value 1 if a synagogue was destroyed or damaged in the ‘Night of Broken Glass’ in 1938, and 0 otherwise.

Table A.3: City-level outcome variables and medieval pogroms – Extended Sample (1) (2) (3) (4) (5) (6) (7) (8) City pop. growth %Pro- %Jewish %Blue- %Unem- %Manu- %Retail 1300- 1750- testant Collar ployed facturing & Trade 1933 1933 ‘25 ’33 ’33 ’33 ’33 ‘33 Panel A: Means by existence of Jewish community in 1349 JewCom 1349=1 2.36 1.99 48 .4 1.4 4 40 .8 16 .5 34 .3 21 .2 (std dev) (1.25) (1.02) (34.0) (1.45) (11.1) (7.97) (13.1) (10.6) JewCom 1349=0 4.61 2.41 55 .6 2.7 2 38 .8 12 .9 29 .1 12 .8 (std dev) (1.59) (.816) (36.0) (3.51) (12.6) (7.50) (14.3) (8.47) 1349 Panel B: Regressions on POG POG 1349 .0857 .168 -6.470 .503** -1.478 -.133 .347 .291 (.507) (.249) (4.445) (.195) (1.150) (.757) (1.377) (.920) JewCom 1349 -1.569** -.516* -4.625 -.481*** -3.733*** -1.628** -3.436** 1.586* (.676) (.272) (4.253) (.181) (1.163) (.725) (1.396) (.846) Obs ervations 54 147 1,061 1,061 1,061 1,061 1,061 1,061 Adjusted R2 .386 .008 .031 .218 .355 .476 .325 .633 All regressions run by OLS for the extended sample. Standard errors in parentheses (clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. All regressions include our standard set of control variables: City population, %Protestants, and %Jewish (except for columns 3 and 4 which exclude %Protestant and %Jewish, respectively). City population corresponds to the year of the dependent variable: ln(City population) in 1300 in column 1, ln(City population) in 1750 in column 2, ln(City population) in 1925 in column 3, and ln(City population) in 1933 in columns 4-8. City population data for 1300 and 1750 are from Bairoch et al. (1988).

Table A.4: Election turnout (%) in various years – Main Sample (1) (2) (3) (4) (5) (6) (7) (8) Main Sample Extended Sample Election year 1924 1928 1930 1933 1924 1928 1930 1933 Panel A: Means by Pogrom in 1349 POG 1349=1 76 .5 73 .5 82 .5 90 .0 (std dev) (8.1) (7.4) (5.6) (3.2) POG 1349=0 77 .5 74 .7 82 .8 89 .0 (std dev) (8.9) (9.0) (5.5) (3.5)

JewCom1349=1 76.8 73.9 82.6 90.0 (std dev) (8.4) (7.9) (5.5) (3.2) JewCom 1349=0 76 .9 72 .9 80 .2 89 .7 (std dev) (8.9) (9.4) (7.2) (3.2) 1349 Panel B: Regressions on POG POG 1349 -.165 -.609 -.00590 .504 -.539 -.923 -.0344 .418 (.995) (.981) (.658) (.390) (1.051) (1.046) (.693) (.396) JewCom 1349 .724 2.094** 2.738*** .257 (.942) (1.021) (.655) (.361) Observations 325 325 325 325 1,295 1,295 1,295 1,295 Adjusted R2 .139 .134 .067 .128 .022 .019 .036 .095 All regressions run by OLS for the extended sample. Standard errors in parentheses (clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. All regressions include our standard set of control variables: City population, %Protestants, and %Jewish. City population corresponds to the year of the election.

Table A.5: Robustness of Main Results (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Baseline Regressions OLS OLS OLS ML$ ML$ OLS

POG1349 .0643*** .0147** .0128 .116* .282** .121** (.0244) (.00581) (.0108) (.0687) (.128) (.0560)

ln(Pop) .0420** -.00246 -.00450 .280*** .950*** .0367*** (.0166) (.00231) (.00467) (.0879) (.0460) (.0134) % Jewish .0120 -.000209 .00209 .0718** .161*** .0345** (.0117) (.00189) (.00398) (.0316) (.0384) (.0147) % Protestant .000368 .000312*** .00120*** -.00189 -.00132 .00000496 (.000426) (.0000896) (.000198) (.00144) (.00211) (.000635) ln(# Jews '33) .786*** (.0801)

% Blue collar .000761 -.00131*** -.00135 -.00929 -.0218 -.000993 (.00163) (.000404) (.000904) (.0137) (.0267) (.00493) % Industry employ. -.000987 .000367 .000818 .00355 .00465 .00183 (.00144) (.000327) (.000754) (.00847) (.0157) (.00318) % Self-employed in .00172 .000259 .000380 -.00477 .0209* -.00145 Retail & Trade (.00176) (.000312) (.000875) (.00585) (.0117) (.00293) ‘Social Capital’ -.000280 .000387 -.0039*** -.00896* -.0245** .00484* (% Voter turnout) (.00179) (.000312) (.000695) (.00469) (.0121) (.00266) % Unemployed ‘33 -.00205 -.0165 .00266 (.00966) (.0181) (.00356)

Observations 320 325 325 278 325 278 Adjusted R2 .045 .082 .187 .109 Panel B: Matching by all Characteristics§

POG1349 .0773*** .0119** .0248** 111.0*** 2.197*** .161*** (.0174) (.00526) (.00989) (37.37) (.539) (.0618)

Observations 320 325 325 278 325 278 All regressions run at the city level. Standard errors in parentheses, clustered at the county (Kreis) level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. %Blue-collar workers and %Industry employment is from the 1925 census in columns 1-3, and from the 1933 census in columns 4-6. The % of self-employed individuals in Retail and Trade is available only from the 1925 census (if no observation is available for 1925, the % of self-employed in overall employment from the 1933 census is used as a proxy). ‘Social capital’ is proxied by voter turnout in the 1924 election – with the exception of column 2, where the 1928 turnout is used. $ Poisson maximum likelihood estimation. § Matching estimation based on geographic longitude and latitude, and on the same set of control variables as used in Panel A. Treatment variable is Pogrom 1349. The average treatment effect for the treated is reported, using robust nearest neighbor estimation with the four closest matches.

Table A.6: Main Results with province and prefecture fixed effects (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Province Fixed Effects OLS OLS OLS ML$ ML$ OLS

POG1349 .0655*** .0147*** .0233** .129 .386** .144** (.0199) (.00537) (.00913) (.0968) (.175) (.0623) ln(Pop) .0350** -.00118 .00104 .00819 .867*** .0449*** (.0143) (.00170) (.00280) (.0968) (.0355) (.0118)

% Jewish .0109 -.000725 .00540 -.0621 .366*** .0371* (.0111) (.00180) (.00329) (.0655) (.0607) (.0218)

% Protestant .000410 .000514*** .000881*** -.00562*** -.000968 .000350 (.000555) (.000121) (.000209) (.00125) (.00237) (.000821) ln(# Jews '33) 1.047*** (.0937)

Province FE yes yes yes yes yes yes

Observations 320 325 325 278 325 278 Adjusted R2 .090 .385 .526 .091 Panel B: Prefecture Fixed Effects OLS OLS OLS ML$ ML$ OLS

POG1349 .0443* .0119** .0238*** .140 .233 .143** (.0235) (.00512) (.00628) (.122) (.195) (.0701) ln(Pop) .0342** -.00161 -.00232 .0323 .855*** .0451*** (.0142) (.00175) (.00203) (.132) (.0332) (.0132)

% Jewish .0176 .00279 .0114*** -.0664 .234*** .0288 (.0116) (.00186) (.00289) (.0662) (.0526) (.0214)

% Protestant .000226 .000384*** .000773*** -.00384** .00188 .000896 (.000483) (.000125) (.000121) (.00180) (.00240) (.00104) ln(# Jews '33) 1.036*** (.134)

Prefecture FE yes yes yes yes yes yes

Observations 320 325 325 278 325 278 Adjusted R2 .124 .513 .782 .113 All regressions run at the city level. There are 20 provinces in the main dataset and 44 prefectures (Regierungsbezirk). Standard errors in parentheses, clustered at the county (Kreis) level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. $ Poisson maximum likelihood estimation.

Table A.7: Dependent variable: Pogrom 1920s (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS OLS OLS Probit ME# GeoMatch§ Sample --- extended ------main --- ext. main

POG1349 .0705*** .0638*** .0705*** .0607*** .896** .0744*** .0819*** .0819*** (.0213) (.0210) (.0214) (.0226) (.411) (.0182) (.0166) (.0162)

JewCom1349 -.00336 -.00391 (mv) (mv) (.0120) (.0114) ln(Pop '25) .0224*** .0390** .281*** (mv) Median Distance (.00631) (.0152) (.0898) 20.2 20.4

% Jewish '25 .0130** .0135 .116 (mv) Mean Distance (.00574) (.0114) (.0926) 23.5 23.5

% Protestant .00026** .00034 .0043 (mv) '25 (.00012) (.00042) (.0041)

Observations 1,271 1,271 320 320 320 320 1,271 320 Adjusted R2 .025 .053 .014 .054 .025 .053 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. # Matching estimation based on the full set of control variables in column (5). Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.8: Dependent variable: % vote for NSDAP in May 1928 election (1) (2) (3) (4) (5) (6) (7) (8) OLS – all cities OLS OLS ML$ ME# GeoMatch§ Sample --- extended ------main --- ext. main 1349 POG .0124** .0139*** .0124** .0142** .424** .0132*** .0112*** .0116** (.00507) (.00536) (.00508) (.00567) (.174) (.00486) (.00431) (.00456)

JewCom1349 -.00379 -.00312 (mv) (.00411) (.00433) ln(Pop '25) -.000240 -.00254 -.0661 (mv) Median Distance (.00138) (.00219) (.0576) 20.0 20.0

% Jewish .000611 .00174 .0465 (mv) Mean Distance '25 (.00128) (.00190) (.0478) 23.1 23.1

% Protestant .00028*** .00029*** .0081*** (mv) '25 (.00005) (.00009) (.0023)

Observations 1,295 1,295 325 325 325 325 1,295 325 Adjusted R2 .005 .051 .010 .043 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. $ Poisson maximum likelihood estimation. # Matching estimation based on the full set of control variables in column (5). Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.9: Dependent variable: % vote for DVFP/NSFP in May 1924 election (1) (2) (3) (4) (5) (6) (7) (8) OLS – all cities OLS OLS ML$ ME# GeoMatch§ Sample --- extended ------main --- ext. main 1349 POG .0115 .0145 .0115 .0147 .188 .0203** .0226*** .0238*** (.0105) (.0105) (.0105) (.0110) (.144) (.0102) (.00765) (.00746)

JewCom1349 .00930 .00941 (mv) (.00892) (.00901) ln(Pop '24) -.00267 -.00123 -.00628 (mv) Median Distance (.00219) (.00418) (.0483) 20.0 20.0

% Jewish .00648** .00701 .0918* (mv) Mean Distance '25 (.00299) (.00442) (.0491) 23.1 23.1

% Protestant .000716*** .00083*** .0108*** (mv) '25 (.000085) (.00018) (.00241)

Observations 1,295 1,295 325 325 325 325 1,295 325 Adjusted R2 .008 .106 .000 .080 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. $ Poisson maximum likelihood estimation. # Matching estimation based on the full set of control variables in column (5). Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.10: Deportations 1933-1944

Dep. Variable: ln(1 + # of deported Jews) # of deported Jews ln(1+#dep. Jews) (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS ME# ME# ML$ ML$ GeoMatch§ Sample ext. --- main ------main --- ext. main

PANEL A: Controlling for Jewish population in 1933

POG1349 .266** .165 .452*** .528*** .232*** .142** .596** .718*** (.132) (.138) (.172) (.150) (.0672) (.0706) (.276) (.242)

JewCom1349 -.150 (mv) (.120)

ln(# Jews '33) .932*** 1.047*** (mv) (mv) 1.028*** .815*** Median Distance (.0236) (.0341) (.0205) (.0822) 42.8 39.3

Additional no no (mv) no yes Mean Distance Controls 48.6 44.2 Observations 981 278 278 278 278 278 981 278 Adjusted R2 .659 .774

PANEL B: Controlling for Jewish population in 1939

POG1349 .416** .346* .650** .726*** .112 .0638 .871*** .955*** (.167) (.180) (.269) (.168) (.105) (.0959) (.336) (.357) 1349 JewCom -.0496 (mv) (.151)

ln(# Jews '39) .728*** .783*** (mv) (mv) .933*** .471*** Median Distance (.0269) (.0444) (.0225) (.105) 51.6 48.6

Additional Controls no no (mv) no yes Mean Distance 55.3 52.0 Observations 642 204 204 204 204 204 642 204 Adjusted R2 .741 .776 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. Additional controls include: ln(pop ‘33), %Protestant, and %Jewish. $ Poisson maximum likelihood estimation. # Matching estimation based on the full set of control variables in columns (1) and (2), respectively. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude, latitude, and ln(# Jews '33) in panel A / ln(# Jews '39) in panel B for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.11: Letters to the editor, Der Stürmer Dep. variable: ln(1 + number of letters) number of letters ln(1+#letters) (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS ME# ME# ML$ ML$ GeoMatch§ Sample extended --- main ------main --- ext. main

POG1349 .207** .103 .189** .227** .458*** .369** .274** .305*** (.0844) (.0841) (.0877) (.0896) (.149) (.144) (.109) (.109) 1349 JewCom -.0247 (mv) (.0724)

ln(Pop '33) .278*** .451*** (mv) (mv) .849*** .848*** Median Distance (.0195) (.0308) (.0551) (.0419) 38.8 36.3

Additional no no (no) (mv) no yes Mean Distance Controls 44.2 40.2

Observations 1,295 325 325 325 325 325 1,294 325 Adjusted R2 .394 .517 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. Additional controls include %Protestant and %Jewish. $ Poisson maximum likelihood estimation. # Matching estimation based on the full set of control variables in columns (1) and (2), respectively. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude, latitude, and ln(Pop ‘33) for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.12: Proportions of deportations and Stürmer letters (1) (2) (3) (4) (5) (6) OLS ME# GeoMatch§ OLS ME# GeoMatch§ Dep. Var. --- %DEPORT ------STÜRMER prop.---

POG1349 1.09** 13.24*** 15.01*** .254*** .185** .334*** (4.067) (4.541) (5.762) (.0971) (.0862) (.0984) ln(Pop) 1.191* (mv) -.167*** (mv) (.701) (.0402)

%Jewish’33 -.0848* (mv) -.00433** (mv) (.0452) (.00194)

%Protestant -.640 (mv) .283*** (mv) '25 (1.134) (.0694)

Observations 278 278 278 325 325 325 Adjusted R2 .066 .166 Standard errors in parentheses (for OLS, clustered at the county level). * p < .10, ** p < .05, *** p < .01. %DEPORT is the percentage of deportees from each locality (relative to Jewish population in 1933); STÜRMER/pop is the number of anti-Semitic letters to Der Stürmer per 10,000 inhabitants. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. Regressions are weighted by city population in 1933. # Matching estimation based on the full set of control variables in columns (1) and (4), respectively. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude, latitude, and ln(Pop ‘33) for each city, using the two closest matches.

Table A.13: Dependent variable: Synagogue damaged or destroyed in 1938 (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS OLS OLS Probit ME# GeoMatch§ Sample (cities w/ synagogue) extended --- main --- ext. main

POG1349 .147*** .114** .147*** .124** .650*** .103* .152** .152** (.0524) (.0512) (.0525) (.0506) (.248) (.0553) (.0688) (.0677)

Jew. Comm. .00758 -.0508 (mv) 1349, NoPog (.0519) (.0516) ln(Pop '33) .0600*** .0498*** .532*** (mv) Median Distance (.00831) (.0113) (.124) 23.7 23.7

Additional no yes no yes yes (mv) Mean Distance Controls 27.6 27.6

Observations 1,040 891 278 278 278 278 1,040 278 Adjusted R2 .024 .075 .042 .098 Standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348 -50, and 0 otherwise. Additional controls include: % Protestant ‘25, % Jewish ‘33. # Matching estimation based on the full set of control variables in column (5). Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.14: Spatial Error Regression Model$ (1) (2) (3) (4) (5) (6) Dep. 1920s NSDAP DVFP Depor- Stürmer Synagogue Variable: pogroms 1928 1924 tations letters attacks 1349 ** ** *** *** POG .0684 .0140 .0241 .112 -.00586 .140 (.0305) (.00563) (.00894) (.122) (.0872) (.0423)

ln(Pop) .0365*** .000915 .00201* .135 .518*** .0473*** (.0105) (.000676) (.00110) (.113) (.0274) (.0127)

% Jewish .0155 .00240 .00980*** -.00767 .110*** .0278** (.0111) (.00197) (.00310) (.0588) (.0268) (.0126)

% Protestant .00025 .00023** .00048*** -.00331** -.00125 .00059 (.00041) (.000095) (.00015) (.0015) (.00110) (.00053)

ln(# Jews .953*** '39) (.103)

Observations 320 320 320 274 320 260 λ# .00054 .01316*** .01291*** -.00028 -.00033** -.0005 (.00053) (.00140) (.00082) (.00040) (.00014) (.0011) Standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. Population is taken from the election data for the respective year. $ Estimated by maximum likelihood, using each city’s geographic location to derive the weighting matrix. All cities with distance less than 4 degrees (~440km) are considered spatially contiguous and are assigned a nonzero spatial weight. # If λ=0, the spatial error model reduces to OLS. For λ≠0, OLS is unbiased and consistent, but inefficient.

Table A.15a: Sample splits. All regressions run by geographical matching. (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Protestant Majority POG 1349 .114*** .0311*** .0516*** 244.0*** 3.043*** .174** (.0280) (.00655) (.0115) (65.02) (.705) (.0742) Observations 165 16 7 167 146 167 146 Panel B: Catholic Majority POG 1349 .0526*** -.00211 .00589 192.0*** 2.888*** .189** (.0186) (.00514) (.00851) (33.36) (.990) (.0833) Observations 151 154 154 128 154 128 Panel C: Western Cities POG 1349 .0379 ** .00928** .0116** 244.9*** 3.132*** .128 (.0152) (.00434) (.00508) (47.94) (.768) (.0784) Observations 160 162 162 138 162 148 Panel D: Eastern Cities POG 1349 .140*** .0149* .0370** 173.8*** 3.054*** .169* (.0330) (.00810) (.0174) (44.45) (.883) (.0948) Observations 160 163 163 140 163 130 Panel E: Larger Cities POG 1349 .115*** .0136** .0314*** 338*** 5.354*** .183* (.0272) (.00547) (.00902) (64.33) (1.051) (.0934) Observations 159 162 162 143 162 142 Panel F: Smaller Cities POG 1349 .0455*** .0155** .0215* 1.767 -.339 .121 (.0155) (.00697) (.0116) (6.374) (.247) (.0904) Observations 161 163 163 135 163 136 All regressions run at the city level by matching estimation based on geography: Matching characteristics are geographic longitude and latitude. Column 4 uses the city’s Jewish population in 1933 as additional matching variable, and column 5 uses city population in 1933. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the two closest matches. Distance (in miles) between each city and its two closest matches is reported. Standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. Protestant and Catholic majority are cities with more than 50% of their population belonging to the respective confession in 1925. The cutoff for Western vs. Eastern cities is the median longitude in the main sample (9°11′ E). The cutoff for larger vs. smaller cities is the median city size in the main sample: 9,022 inhabitants in 1933.

Table A.15b: Sample splits. Baseline specification (OLS / ML). (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks OLS OLS OLS ML$ ML$ OLS Panel A: Protestant Majority POG 1349 .111*** .0350*** .0471*** .180* .360** .0965 (.0312) (.00772) (.0142) (.0993) (.142) (.0696) Observations 165 167 167 146 167 146 Adjusted R2 .034 .084 .095 .057 Panel B: Catholic Majority POG 1349 .00853 -.0126 -.0233 .349** .457 .145* (.0339) (.00788) (.0169) (.145) (.314) (.0791) Observations 151 154 154 128 154 128 .109 .034 .023 .126 Panel C: Western Cities POG 1349 .0415** .0114* .0110** -.00161 .506** .191* (.0200) (.00643) (.00546) (.192) (.248) (.103) Observations 160 162 162 138 162 148 .028 .090 .305 .122 Panel D: Eastern Cities POG 1349 .0934** .0275*** .0524*** .209** .327* .0690 (.0365) (.00835) (.0157) (.0878) (.182) (.0602) Observations 160 163 163 140 163 130 .091 .042 .063 .067 Panel E: Larger Cities POG 1349 .0653 .0131** .0223* .165* .544*** .142** (.0501) (.00617) (.0130) (.0880) (.145) (.0650) Observations 159 162 162 143 162 142 .056 .003 .033 .109 Panel F: Smaller Cities POG 1349 .0449** .0132 .00478 -.0325 -.523 .0654 (.0202) (.00918) (.0173) (.217) (.331) (.0851) Observations 161 163 163 135 163 136 .008 .067 .115 .074 All regressions run at the city level. Standard errors in parentheses, clustered at the county level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. All regressions control for: ln(city population), % Protestant, % Jewish. Column 4 uses the city’s Jewish population in 1933 as additional control variable. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. Protestant and Catholic majority are cities with more than 50% of their population belonging to the respective confession in 1925. The cutoff for Western vs. Eastern cities is the median longitude in the main sample (9°11′ E). The cutoff for larger vs. $ smaller cities is the median city size in the main sample: 9,022 inhabitants in 1933. Poisson maximum likelihood estimation.

Table A.15c: Sample splits. All regressions run by propensity score matching. (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Protestant Majority POG 1349 .114*** .0333*** .0532*** 188.2*** 2.624*** .0963 (.0308) (.00684) (.0135) (7.93) (.762) (.0631) Observations 165 167 167 146 167 146 Panel B: Catholic Majority POG 1349 .0452*** -.0100 -.0159 188.1*** 2.787*** .153 (.0165) (.00672) (.0155) (49.23) (.695) (.109) Observations 151 154 154 128 154 128 Panel C: Western Cities POG 1349 .0379** .00865* .00889 216.6*** 2.534*** .178 (.0160) (.00462) (.00642) (5.71) (.747) (.119) Observations 160 162 162 138 162 148 Panel D: Eastern Cities POG 1349 .130*** .0266*** .0562*** 148.3*** 2.956*** .0552 (.0335) (.00726) (.0152) (49.66) (.806) (.0638) Observations 160 163 163 140 163 130 Panel E: Larger Cities POG 1349 .0963*** .0131* .0254 29.9*** 4.756*** .109 (.0302) (.00757) (.0166) (74.80) (1.002) (.0676) Observations 159 162 162 143 162 142 Panel F: Smaller Cities POG 1349 .0455** .0142* .0103 .825 -.174 .0354 (.0186) (.00837) (.0156) (4.730) (.221) (.0860) Observations 161 163 163 135 163 136 Standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. All regressions run at the city level by propensity score matching. Treatment variable is POG1349, which takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. Matching variables are ln(city population), % Protestant, and % Jewish. Column 4 uses the city’s Jewish population in 1933 as additional matching variable. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. Protestant and Catholic majority are cities with more than 50% of their population belonging to the respective confession in 1925. The cutoff for Western vs. Eastern cities is the median longitude in the main sample (9°11′ E). The cutoff for larger vs. smaller cities is the median city size in the main sample: 9,022 inhabitants in 1933.

Table A.16: Dependent variable: % vote for NSDAP after 1928 (1) (2) (3) (4) (5) (6) (7) (8) Year: ---1930 ------1933 --- OLS OLS ME# GeoM.§ OLS OLS ME# GeoM.§ Sample extended --- main --- extended --- main ---

POG1349 .0164 .0137 .0123 .00901 -.0123 -.0113 -.0129 -.0236 (.0104) (.0101) (.0121) (.0130) (.0126) (.0125) (.0124) (.0184)

JewCom1349 -.00275 (mv) -.0207* (mv) (.0100) (.0117) ln(Pop) -.0070*** -.00816** (mv) .0093*** -.0111*** (mv) (.0026) (.00320) (.0035) (.0036)

%Jewish’33 -.00034 .00240 (mv) .0115*** .0010*** (mv) (.0024) (.0032) (.0034) (.0038)

%Protestant .00167*** .00128*** (mv) .0028*** .0023*** (mv) '25 (.00011) (.00015) (.00014) (.00017)

Observations 1,295 325 325 325 1,295 325 325 325 Adjusted R2 .330 .219 .495 .426 Standard errors in parentheses (clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. Additional controls include: %Blue-Collar ‘33, %Unemployed ‘33. Population is taken from the election data for the respective year. # Matching estimation based on the full set of control variables in columns (2) and (4), respectively. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance (in miles) between each city and its two closest matches for columns (4) and (8) are respectively: median distance 20.0; mean distance 23.1.

Table A.17: Dependent variable: % vote for DNVP in May 1924 election (1) (2) (3) (4) (5) (6) (7) (8) OLS – all cities OLS OLS ML$ ME# GeoMatch§ Sample --- extended ------main --- extended main 1349 POG -.0403** -.0272** -.0403** -.0267** -.170** -.0258* -.0197 -.0173 (.0172) (.0131) (.0172) (.0131) (.0859) (.0138) (.0191) (.0170) 1349 JewCom .00401 .0120 (mv) (.0167) (.0128) ln(Pop '24) .00234 -.00505 -.0214 (mv) Median Distance (.00385) (.00419) (.0310) 20.0 20.0

% Jewish -.00440 -.00337 -.00727 (mv) Mean Distance '25 (.00314) (.00403) (.0354) 23.1 23.1 *** *** *** % Protestant .00215 .00201 .0160 (mv) '25 (.00016) (.00016) (.0010)

Observations 1,295 1,295 325 325 325 325 1,294 325 Adjusted R2 .009 .310 .023 .372 Standard errors in parentheses (clustered at the county level). * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. $ Poisson maximum likelihood estimation. # Matching estimation based on the full set of control variables in column (5). Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. ‘mv’ indicates match variable. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two closest matches. Distance between each city and its two closest matches is reported.

Table A.18: Main Results – Cities with documented or likely Jewish settlement (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Baseline Regressions OLS OLS OLS ML$ ML$ OLS

POG1349 .0508** .0146*** .0201** .0418 .267** .113*** (.0214) (.00482) (.00872) (.0535) (.108) (.0375) ln(Pop) .0340*** -.00280 -.00139 .200*** .843*** .0436*** (.0126) (.00186) (.00356) (.0689) (.0378) (.0117)

% Jewish .0125 .00126 .00666* .0537* .188*** .00450 (.00986) (.00175) (.00400) (.0277) (.0283) (.0155)

% Protestant .000238 .000287*** .000801*** -.00404*** -.00484** -.000135 (.000331) (.0000710) (.000144) (.00103) (.00207) (.000596) ln(# Jews '33) .849*** (.0669)

Observations 411 419 419 348 402 346 Adjusted R2 .052 .053 .095 .074 Panel B: Matching Estimation#

POG1349 .0606*** .0138*** .0230*** 107.5** 1.854*** .0604* (.0206) (.00443) (.00807) (48.32) (.612) (.0353)

Observations 411 419 419 348 402 346 Panel C: Geographic Matching§

1349 POG .0797*** .00646 .0148** 154.7*** 1.981*** .104** (.0203) (.00514) (.00730) (31.52) (.580) (.0482)

Observations 411 419 419 348 419 356 All regressions run at the city level, including all cities with documented Jewish community in 1349; in addition, we include cities with a predicted probability of medieval settlement larger 50%. The latter is based on a probit regression in the style of columns 1 and 2 in Table VIII in the paper, where all medieval correlates mentioned in Table VIII are used. Standard errors in parentheses, clustered at the county level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. $ Poisson maximum likelihood estimation. # Matching estimation based on the same set of control variables as used in Panel A. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude as well as JewCom1349. Column 4 uses the city’s Jewish population in 1933 as additional matching variable, and column 5 uses city population in 1933. Treatment variable is Pogrom 1349. ATT are reported, using robust nearest neighbor estimation with the two closest matches. Distance (in miles) between each city and its two closest matches is reported.

Table A.19: Controlling for predicted pogrom probability (1) (2) (3) (4) (5) (6) (7) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue Principal pogroms 1928 1924 tations letters attacks Component$ OLS OLS OLS ML# ML# OLS OLS

POG1349 .0614*** .0155** .0169 .123* .385*** .122** .313** (.0236) (.00642) (.0116) (.0716) (.145) (.0546) (.142)

Prob(POG1349) -.0116 -.0172 -.0306 .136 -.173 -.142 -.419 (.116) (.0219) (.0375) (.217) (.509) (.121) (.490) ln(Pop) .0418*** -.00216 -.000053 .184 .854*** .0548*** -.0704 (.0157) (.00228) (.00442) (.131) (.0441) (.0130) (.0700)

% Jewish .0167 .00179 .00773* .0111 .225*** .0266* .0160 (.0119) (.00197) (.00460) (.0857) (.0533) (.0137) (.101)

% Protestant .00032 .00028*** .00081*** -.0041*** -.0053** .00038 .276*** (.00043) (.000086) (.00018) (.0012) (.0022) (.00062) (.0755) ln(# Jews '39) .860*** (.126)

Observations 318 323 323 276 323 276 309 Adjusted R2 .055 .042 .080 .096 .051

All regressions run at the city level. Standard errors in parentheses, clustered at the county level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. Prob(POG1349) is predicted pogrom probability from a Probit regression, using all medieval variables that are significantly correlated with POG1349 in columns 5 or 6 of Table VIII in the paper. These comprise indicators for free imperial cities, Staufer cities, market rights in 1349, as well as ln(age of the city in 1349). City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-7, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-7. %Protestants is from the 1925 census. # Poisson maximum likelihood estimation. $ First principal component (standardized) obtained from six 20th century proxies for anti-Semitism: Pogrom in the 1920s, DVFP votes 1924, NSDAP votes 1928, deportations, Stürmer letters, and an indicator variable for synagogue attacks. See section III.F in the paper for details.

Table A.20: Main Results – Extended Sample, only for cities that were first mentioned before 1349 (1) (2) (3) (4) (5) (6) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue pogroms 1928 1924 tations letters attacks Panel A: Baseline Regressions OLS OLS OLS ML$ ML$ OLS

POG1349 .0593*** .0144*** .0165 .193** .367** .118** (.0213) (.00550) (.0108) (.0820) (.152) (.0529)

JewCom1349 -.0218 .000633 .00764 -.225** -.197 -.0580 (.0138) (.00463) (.00948) (.111) (.143) (.0561) ln(Pop) .0369*** -.00300* -.00586** .214*** .869*** .0520*** (.0101) (.00179) (.00288) (.0694) (.0236) (.0100)

% Jewish .0205** .00191 .00590 .0200 .217*** .0110 (.00936) (.00164) (.00365) (.0557) (.0264) (.0138)

% Protestant .000333 .00032*** .000921*** -.00282* -.00313* -.00084* (.00021) (.000058) (.00010) (.00150) (.00180) (.00049)

ln(# Jews '33) .839*** (.0674)

Observations 659 673 673 552 618 523 Adjusted R2 .068 .060 .120 .071 Panel B: Matching Estimation#

POG1349 .0733*** .0132*** .0198* 42.69 2.336*** .111** (.0191) (.00489) (.0103) (68.02) (.591) (.0550)

Observations 659 673 673 552 618 523 Panel C: Geographic Matching§ 1349 POG .0819*** .0113*** .0238*** 218.0*** 2.955*** .152** (.0162) (.00439) (.00766) (36.54) (.539) (.0677)

Median distance 20.2 20.0 20.0 21.5 21.5 23.6 Mean distance 23.4 23.0 23.0 26.1 28.6 27.5

Observations 659 673 673 552 673 558 All regressions run at the city level, including all cities that were first mentioned in public records before 1349. Standard errors in parentheses, clustered at the county level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. JewCom1349 equals 1 if a Jewish community is documented before 1349. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. $ Poisson maximum likelihood estimation. # Matching estimation based on the same set of control variables as used in Panel A. Treatment variable is Pogrom 1349. Average Treatment Effects for the Treated (ATT) are reported, using robust nearest neighbor estimation with the four closest matches. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude as well as JewCom1349. Column 4 uses the city’s Jewish population in 1933 as additional matching variable, and column 5 uses city population in 1933. Treatment variable is Pogrom 1349. ATT are reported, using robust nearest neighbor estimation with the two closest matches. Distance (in miles) between each city and its two closest matches is reported.

Table A.21: Extinction of Jewish communities in 1349 (1) (2) (3) (4) (5) (6) (7) Dep. Variable: 1920s NSDAP DVFP Depor- Stürmer Synagogue Principal pogroms 1928 1924 tations letters attacks Component$ OLS OLS OLS ML# ML# OLS OLS

PANEL A: Main Sample

Pogrom 1349, .0404 .0156** .0183 .221*** .387*** .122** .331** Comm. vanished (.0248) (.00628) (.0123) (.0662) (.149) (.0516) (.150)

Pogrom 1349, .0971*** .0117 .00837 .0110 .329 .126** .226 Comm. survived (.0364) (.00718) (.0130) (.0806) (.213) (.0607) (.160)

Observations 320 325 325 278 325 278 311 Adjusted R2 .061 .041 .079 .094 .050

PANEL B: Extended Sample

Pogrom 1349, .0454** .0152** .0185 .253*** .396** .113** .382*** Comm. vanished (.0231) (.00600) (.0119) (.0731) (.159) (.0524) (.144)

Pogrom 1349, .0972*** .0117* .00748 .0339 .331 .116* .250 Comm. survived (.0360) (.00689) (.0126) (.0784) (.222) (.0616) (.156)

JewCom1349 -.00393 -.00312 .00940 -.195** -.172 -.0509 .0170 (.0114) (.00433) (.00901) (.0968) (.153) (.0551) (.105)

Observations 1,271 1,295 1,295 981 1,061 891 1,042 Adjusted R2 .056 .051 .106 .074 .124 All regressions run at the city level. Standard errors in parentheses, clustered at the county level. * p < .10, ** p < .05, *** p < .01. POG1349 takes the value 1 if a pogrom occurred in the years 1348-50, and 0 otherwise. All regressions control for: ln(city population), % Protestant, % Jewish. Column 4 uses the city’s Jewish population in 1933 as additional control variable. City population is taken from the 1925 census in column 1, from the election data for the respective year in columns 2 and 3. In columns 4-6, city population is from the 1933 census. %Jews is from the 1925 census for columns 1-3, and from 1933 census in columns 4-6. %Protestants is from the 1925 census. $ First principal component (standardized) obtained from six 20th century proxies for anti-Semitism: Pogrom in the 1920s, DVFP votes 1924, NSDAP votes 1928, deportations, Stürmer letters, and an indicator variable for synagogue attacks. See section III.F in the paper for details.

Table A.22: When persistence fails – detailed analysis. Dep. var.: Principal component$ (1) (2) (3) (4) (5) (6) Dummy Open City Growth Dummy Hanseatic Open City City City Growth Panel A: Dummy interactions Panel B: Sample splits by indicator POG1349 1.053*** .877*** .882*** Dummy =1 (.380) (.305) (.308) (Hanseatic / Open / Growth>Median) Dummy .647** .217 .280 POG1349 -.207 -.0198 -.367 (.276) (.389) (.306) (.204) (.223) (.335) Dummy × -1.100** -1.170*** -1.096*** Obs. 29 174 56 POG1349 (.455) (.434) (.382) Adj. R2 .178 .036 -.036 Igrowth,ind -.0936 Dummy =0 (.591) (non-Hanseatic/non-Open/Growth

Table A.23: Pogroms in 1349 and pre-plague attacks by sub-period (1) (2) (3) #POG1096-1347 #POG1175-1347 #POG1225-1347 POG 1349=1 1.69 1.20 .84 POG1349=0 0 0 .30

Difference 1.69** 1.20** .54* p-value .022 .040 .095 Observations 16 30 60 POG1349 indicates whether a city saw a Black Death Pogrom. #POG y -1347 is the number of attacks on Jewish communities between year y (1096, 1175, 1225) and 1347. Each subsample only includes communities with documented first Jewish settlement before year y. * p < .10, ** p < .05.

Table A.24: Pre-plague pogroms and pogroms in 1349. Dep. Var: POG1349 (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS OLS OLS OLS OLS ME# GeoM§ PANEL A: Pre-plague sample period 1096 – 1347 Number POG y -1347 .190** .212 .302 (.0838) (.122) (.183) Dummy POG y -1347 .750*** .906*** .667* .750*** .542* (.231) (.201) (.363) (.274) (.325)

Controls no yes no no yes no (mv) no Province FE no no yes no no yes no no

Observations 16 16 16 16 16 16 16 16 Adjusted R2 .272 .006 .303 .670 .648 .544 PANEL B: Pre-plague sample period 1175 – 1347

Number POG y -1347 .119** .0903 .116** (.0491) (.0561) (.0515)

Dummy POG y -1347 .357** .383 .424*** .133 .594* (.133) (.250) (.149) (.177) (.328)

Controls no yes no no yes no (mv) no Province FE no no yes no no yes no no

Observations 30 29 30 30 29 30 29 30 Adjusted R2 .112 .004 .088 .201 .076 .216 PANEL C: Pre-plague sample period 1225 – 1347

Number POG y -1347 .0878** .0805* .0853** (.0367) (.0414) (.0380)

Dummy POG y -1347 .122 .103 .154* .125 .0690 (.0956) (.100) (.0808) (.115) (.127)

Controls no yes no no yes no (mv) no Province FE no no yes no no yes no no

Observations 60 59 60 60 59 60 59 60 Adjusted R2 .031 .067 .155 .010 .047 .148 Robust standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. Number POG y -1347 is the number of attack s on Jews in a city between year y (1096, 1175, 1225) and 1347. Each regression only includes communities with documented first Jewish settlement before year y. Dummy POG y -1347 equals 1 if at least one pogrom occurred during the years y-1347, and 0 otherwise. ‘Controls’ include all medieval variables that are significantly correlated with POG1349 in columns 5 or 6 of Table VIII in the paper. These comprise indicators for free imperial cities, Staufer cities, market rights in 1349, as well as ln(age of the city in 1349). § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two nearest matches. Average Treatment Effects for the Treated are reported.

Table A.25: Pre-plague pogroms and 20C anti-Semitism. Dep. Var.: Principal Component$ (1) (2) (3) (4) (5) (6) (7) (8) OLS GeoM§ OLS OLS GeoM§ OLS OLS GeoM§ Sample period (y-1347) 1096–1347 --- 1175–1347 ------1225–1347 --- y -1347 Number POG .113 .355** .190** .306*** .204*** (.165) (.154) (.0908) (.0957) (.0673) Dummy POG y -1347 .399** .842*** .458** (.185) (.190) (.180)

Controls yes yes yes yes yes Province FE no no yes no yes

Observations 15 15 29 29 29 59 59 59 Adjusted R2 -.306 .173 .542 .188 .624 Standard errors in parentheses (clustered at the county level). * p < .10, ** p < .05, *** p < .01. Number POG y -1347 is the number of attacks on Jews in a city between year y (1096, 1175, 1225) and 1347. Each regression only includes communities with documented first Jewish settlement before year y. Dummy POG y -1347 equals 1 if at least one pogrom occurred during the years y-1347, and 0 otherwise. Control variables are ln(city population in 1933), %Jews in 1933, and %Protestants in 1925. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two nearest matches. Average Treatment Effects for the Treated are reported. $ First principal component (standardized) obtained from six 20th century proxies for anti-Semitism: Pogrom in the 1920s, DVFP votes 1924, NSDAP votes 1928, deportations, Stürmer letters, and an indicator variable for synagogue attacks. See section III.F in the paper for details.

Table A.26: Judensau sculptures and persecutions, 1096-1945. Conditional means (1) (2) (3) (4) (5) pre-1347 1349 #POG POG HepHep HepHepall 20C PC Yes 1.0 1.0 .23 .31 .335 No .41 .71 .04 .05 .006

Difference .59*** .29** .19*** .26*** .329 p-value .006 .023 .001 .0001 .302 Observations 325 325 325 325 311 Conditional means are reported for cities with and without Judensau sculptures, for our main sample. ** p < .05, *** p < .01. pre-1347 #POG is the number of attacks on Jewish communities in a city before 1347. POG1349 indicates whether a city saw a Black Death Pogrom. HepHep is an indicator variable for whether Hep-Hep riots against Jewish communities broke out in 1819 in the city; HepHepAll also includes attempted attacks on Jews that were stopped by army units or by the police. 20C PC is the first principal component (standardized) obtained from six 20th century proxies for anti-Semitism: Pogrom in the 1920s, DVFP votes 1924, NSDAP votes 1928, deportations, Stürmer letters, and an indicator variable for synagogue attacks. See section III.F in the paper for details.

Table A.27: Black Death pogroms and Judensau sculptures: Regressions (1) (2) (3) (4) OLS OLS OLS GeoM§ 1349 POG .0553*** .0424*** .0517*** .0553*** (.0150) (.0143) (.0157) (.0141) Controls no yes no Province FE no no yes

Observations 325 323 325 325 Adjusted R2 .008 .014 .038 Dependent variable is a dummy that equals 1 if a city had a Judensau sculpture . Robust standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. POG1349 indicates whether a city saw a Black Death Pogrom. ‘Controls’ include all medieval variables that are significantly correlated with POG1349 in columns 5 or 6 of Table VIII in the paper. These comprise indicators for free imperial cities, Staufer cities, market rights in 1349, as well as ln(age of the city in 1349). There are 20 provinces in the main dataset, for which fixed effects are included in column 3. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two nearest matches. Average Treatment Effects for the Treated are reported.

Table A.28: Hep-Hep riots in 1819 and anti-Semitism, 1096-1945 (1) (2) (3) (4) #POGpre-1347 POG1349 Judensau 20C PC Yes .80 .93 .20 .370 No .42 .71 .03 .004

Difference .38* .22* .17*** .366 p-value .055 .063 .001 .233 Observations 325 325 325 325 Conditional means are reported for cities with and without Hep-Hep attacks on Jews in 1819, for our main pre-1347 sample. * p < .10, ** p < .05, *** p < .01. #POG is the number of attacks on Jewish communities in a city before 1347. POG1349 indicates whether a city saw a Black Death Pogrom. Judensau is an indicator for whether a city had a Judensau sculpture. 20C PC is the first principal component (standardized) obtained from six 20th century proxies for anti-Semitism: Pogrom in the 1920s, DVFP votes 1924, NSDAP votes 1928, deportations, Stürmer letters, and an indicator variable for synagogue attacks. See section III.F in the paper for details.

Table A.29: Black Death pogroms and Hep-Hep attacks: Regressions (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS OLS GeoM§ OLS OLS OLS GeoM§ Dep. Var.: HepHep HepHepall 1349 POG .0485** .0355* .0431** .0447*** .0544** .0383 .0608** .0596*** (.0190) (.0207) (.0208) (.0171) (.0234) (.0244) (.0253) (.0188) Controls no yes no no yes no Province FE no no yes no no yes

Observations 325 323 325 325 325 323 325 325 Adjusted R2 .008 .014 .038 .007 .010 .037 Robust standard errors in parentheses. * p < .10, ** p < .05, *** p < .01. POG1349 indicates whether a city saw a Black Death Pogrom. HepHep is an indicator variable for whether Hep-Hep riots against Jewish communities broke out in 1819 in the city; HepHepAll also includes attempted attacks on Jews that were stopped by army units or by the police. ‘Controls’ include all medieval variables that are significantly correlated with POG1349 in columns 5 or 6 of Table VIII in the paper. These comprise indicators for free imperial cities, Staufer cities, market rights in 1349, as well as ln(age of the city in 1349). There are 20 provinces in the main dataset, for which fixed effects are included in columns 3 and 7. § Matching estimation based on geography: Matching characteristics are geographic longitude and latitude for each city, using the two nearest matches. Average Treatment Effects for the Treated are reported.