《社会保障概论》课程教学资源（课后阅读资料）APSR_Volume112. Issue 4.Nov 2018_Examining_a_Most_Likely_Case_for_Strong_Campaign_Effects_Hitlers_Speeches_and_the_Rise_of_the_Nazi_P

American Political Science Review (2018)112.4,1050-1066 doi:10.1017/S0003055418000424 American Political Science Association 2018 Examining a Most Likely Case for Strong Campaign Effects:Hitler's Speeches and the Rise of the Nazi Party,1927-1933 PETER SELB University of Konstanz SIMON MUNZERT Hertie School of Governance itler's rise to power amidst an unprecedented propaganda campaign initiated scholarly interest in campaign effects.To the surprise of many,empirical studies often found minimal effects.The predominant focus of early work was on U.S.elections,though.Nazi propaganda as the archety- pal and,in many ways,most likely case for strong effects has rarely been studied.We collect extensive data about Hitler's speeches and gauge their impact on voter support at five national elections preceding the dictatorship.We use a semi-parametric difference-in-differences approach to estimate effects in the face of potential confounding due to the deliberate scheduling of events.Our findings suggest that Hitler's speeches,while rationally targeted,had a negligible impact on the Nazis'electoral fortunes.Only the 1932 presidential runoff,an election preceded by an extraordinarily short,intense,and one-sided campaign, yielded positive effects.This study questions the importance of charismatic leaders for the success of populist movements. "I am conscious that I have no equal in the art of sway- [the national parliament]against the Catholic and Marx- ing the masses."-Adolf Hitler in a reported conversa ist deputies.If out-voting them takes longer than out- tion(1932-34)with early copartisan Hermann Rauschning shooting them,at least the results will be guaranteed by (1939).The authenticity of these records has been chal- their own constitution." 4号元 lenged:see Janssen (1985). Hitler was released on parole at the end of 1924 PROLOGUE and caused the ban on the NSDAP to be lifted by affirming his party's new loyalty to the constitution. n November 11.1923.almost ten years before His previous assertions notwithstanding.Hitler held a the Nazi seizure of power,Adolf Hitler was ar- rabble-rousing public speech at the party's relaunch on rested and subsequently sentenced to five years February 27 1925 in Munich.The regional authorities' in prison for his leading role in the Beer Hall Putsch, reaction came swiftly:five forthcoming public appear- a failed coup d'etat against the national government. ances were cancelled immediately,and,on March 9,the While detained,Hitler ordered the banned and disinte- Bavarian cabinet issued a two-year gag order against grating National Socialist German Workers'Party(Na- Hitler (Rosch 2002,56-68).Many other regional gov- tionalsozialistische Deutsche Arbeiterpartei,NSDAP) ernments,including those of Prussia,Saxony,Hesse to transform from a subversive battle group to a viable Oldenburg,Anhalt,Hamburg,and Lubeck,followed political party(Stachura 1980).As Hitler put it in a per- (Bruppacher 2012,159-72).Although the NSDAP was 685:50190 sonal conversation with Nazi fundraiser Kurt Ludecke not banned once again,party organs and meetings (quoted in Pridham 1973,27): were subject to increased surveillance.According to police reports,turnout at NSDAP meetings and rallies "Instead of working to achieve power by an armed coup, declined markedly in the subsequent period(Rosch we shall have to hold our noses and enter the Reichstag 2002,208-10),and the NSDAP's poor results at the state elections in Saxony,Mecklenburg-Schwerin,and Peter Selb is a Professor of Survey Research,University of Konstanz. Thuringia reinforced the view that "the NSDAP with Department of Politics and Public Administration,P.O.Box 85,D. a Hitler free to speak in public would cause no fu- 78457 Konstanz,Germany(peter.selb@uni.kn). ture concern to the authorities"(Pridham 1973,77). Simon Munzert is a Lecturer of Political Data Science,Hertie On March 6,1927 the Bavarian government revoked School of Governance,Friedrichstr.180,D-10117 Berlin,Germany (munzert@hertie-school.org). its gag order,and Adolf Hitler would take up his un- We are grateful to Johannes Haussler and Sascha Gobel for their precedented campaign activities.In the period of time superb research assistance;Fred Hockney for his proof-reading and between the repeal of the speaking ban on March 6, language editing;Birgit Jacob and Hannah Laumann for their edit. 1927 and the eve of the Reichstag election of March 5, ing:Christian Spinner,who sounded out the terrain in his Bachelor's 1933,Hitler had 455 public appearances,with a gross es- thesis;Juirgen W.Falter,Jonas MeBner,Dieter Ohr,and Paul Thurner. timated attendance of at least 4.5 million.In only four who provided their data;the participants of the research colloquium of the Graduate School of Decision Sciences at the University of years,the NSDAP evolved from a radical fringe group, Konstanz;the panel on media and politics at the EPSA Conference garnering less than 3%of the vote at the 1928 Re- 2016 in Brussels:Alexander De Juan.Thomas Gschwend,Moritz ichstag election,into the most popular German party, Marbach,and the reviewers for valuable comments;and the responsi- ble editor for his enduring support during a long and controversial re- view process.Replication files are available at the American Political Science Review Dataverse:https://doi.org/10.7910/DVN/3KOQWQ We estimated attendance figures from available police reports and imputed missing values from press releases when necessary.For data Received:September 13,2016:revised:September 11,2017;accepted: sources,estimation procedures,and detailed descriptive statistics,see June 24,2018.First published online:August 7,2018. Appendix C. 1050

American Political Science Review (2018) 112, 4, 1050–1066 doi:10.1017/S0003055418000424 © American Political Science Association 2018 Examining a Most Likely Case for Strong Campaign Effects: Hitler’s Speeches and the Rise of the Nazi Party, 1927–1933 PETER SELB University of Konstanz SIMON MUNZERT Hertie School of Governance Hitler’s rise to power amidst an unprecedented propaganda campaign initiated scholarly interest in campaign effects. To the surprise of many, empirical studies often found minimal effects. The predominant focus of early work was on U.S. elections, though. Nazi propaganda as the archety￾pal and, in many ways, most likely case for strong effects has rarely been studied.We collect extensive data about Hitler’s speeches and gauge their impact on voter support at five national elections preceding the dictatorship. We use a semi-parametric difference-in-differences approach to estimate effects in the face of potential confounding due to the deliberate scheduling of events. Our findings suggest that Hitler’s speeches, while rationally targeted, had a negligible impact on the Nazis’ electoral fortunes. Only the 1932 presidential runoff, an election preceded by an extraordinarily short, intense, and one-sided campaign, yielded positive effects. This study questions the importance of charismatic leaders for the success of populist movements. “I am conscious that I have no equal in the art of sway￾ing the masses.” —Adolf Hitler in a reported conversa￾tion (1932-34) with early copartisan Hermann Rauschning (1939). The authenticity of these records has been chal￾lenged; see Janssen (1985). PROLOGUE On November 11, 1923, almost ten years before the Nazi seizure of power, Adolf Hitler was ar￾rested and subsequently sentenced to five years in prison for his leading role in the Beer Hall Putsch, a failed coup d’état against the national government. While detained, Hitler ordered the banned and disinte￾grating National Socialist German Workers’ Party (Na￾tionalsozialistische Deutsche Arbeiterpartei, NSDAP) to transform from a subversive battle group to a viable political party (Stachura 1980).As Hitler put it in a per￾sonal conversation with Nazi fundraiser Kurt Lüdecke (quoted in Pridham 1973, 27): “Instead of working to achieve power by an armed coup, we shall have to hold our noses and enter the Reichstag Peter Selb is a Professor of Survey Research, University of Konstanz, Department of Politics and Public Administration, P.O. Box 85, D- 78457 Konstanz, Germany (peter.selb@uni.kn). Simon Munzert is a Lecturer of Political Data Science, Hertie School of Governance, Friedrichstr. 180, D-10117 Berlin, Germany (munzert@hertie-school.org). We are grateful to Johannes Häussler and Sascha Göbel for their superb research assistance; Fred Hockney for his proof-reading and language editing; Birgit Jacob and Hannah Laumann for their edit￾ing; Christian Spinner, who sounded out the terrain in his Bachelor’s thesis; Jürgen W. Falter, Jonas Meßner,Dieter Ohr, and Paul Thurner, who provided their data; the participants of the research colloquium of the Graduate School of Decision Sciences at the University of Konstanz; the panel on media and politics at the EPSA Conference 2016 in Brussels; Alexander De Juan, Thomas Gschwend, Moritz Marbach, and the reviewers for valuable comments; and the responsi￾ble editor for his enduring support during a long and controversial re￾view process. Replication files are available at the American Political Science Review Dataverse: https://doi.org/10.7910/DVN/3KOQWQ. Received: September 13, 2016; revised: September 11, 2017; accepted: June 24, 2018. First published online: August 7, 2018. [the national parliament] against the Catholic and Marx￾ist deputies. If out-voting them takes longer than out￾shooting them, at least the results will be guaranteed by their own constitution.” Hitler was released on parole at the end of 1924 and caused the ban on the NSDAP to be lifted by affirming his party’s new loyalty to the constitution. His previous assertions notwithstanding, Hitler held a rabble-rousing public speech at the party’s relaunch on February 27, 1925 in Munich. The regional authorities’ reaction came swiftly: five forthcoming public appear￾ances were cancelled immediately, and, on March 9, the Bavarian cabinet issued a two-year gag order against Hitler (Rösch 2002, 56–68). Many other regional gov￾ernments, including those of Prussia, Saxony, Hesse, Oldenburg, Anhalt, Hamburg, and Lübeck, followed (Bruppacher 2012, 159–72). Although the NSDAP was not banned once again, party organs and meetings were subject to increased surveillance. According to police reports, turnout at NSDAP meetings and rallies declined markedly in the subsequent period (Rösch 2002, 208–10), and the NSDAP’s poor results at the state elections in Saxony, Mecklenburg-Schwerin, and Thuringia reinforced the view that “the NSDAP with a Hitler free to speak in public would cause no fu￾ture concern to the authorities” (Pridham 1973, 77). On March 6, 1927, the Bavarian government revoked its gag order, and Adolf Hitler would take up his un￾precedented campaign activities. In the period of time between the repeal of the speaking ban on March 6, 1927 and the eve of the Reichstag election of March 5, 1933,Hitler had 455 public appearances,with a gross es￾timated attendance of at least 4.5 million.1 In only four years, the NSDAP evolved from a radical fringe group, garnering less than 3% of the vote at the 1928 Re￾ichstag election, into the most popular German party, 1 We estimated attendance figures from available police reports and imputed missing values from press releases when necessary. For data sources, estimation procedures, and detailed descriptive statistics, see Appendix C. 1050 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Hitler's Speeches and the Rise of the Nazi Party with more than 37%of the national vote in July 1932 portrayal of early Nazi propaganda by later broadcast- effectively paving the way for the Nazi takeover in ing director and Goebbels'deputy Eugen Hadamovsky March 1933 (1933.44): "All the means of public opinion were denied to Hitler INTRODUCTION His newspapers were banned,he was denied use of the ra- After observing the influence Hitler seemed to wield dio,his brochures and leaflets were confiscated.He had no through the use of propaganda,refugee-scholar Paul choice but to reach the masses directly through constantly Lazarsfeld fielded a panel survey in Erie County,Ohio growing mass rallies." during the run-up to the 1940 U.S.presidential elec- Ohr(1997)collects data on local party events in Hes- tion.This marked the beginning of modern research into campaign effects (Hillygus 2010).The study's sian communities and finds that in the period between findings were surprising-they did not substantiate the 1930 national and the 1931 regional parliamentary Lazarsfeld's motivating concern that campaigns could elections,there is a positive relationship in the changes of municipal Nazi vote shares.Wernette (1974)also arbitrarily manipulate the public.Instead,the pres- codes local Nazi election activities during the run-up idential campaign was found to merely activate the voters'predispositions (Lazarsfeld,Berelson,and to the 1930 Reichstag election from a national news- Gaudet 1948).Subsequent studies came to simi- paper and finds a positive correlation with changes in lar conclusions (Berelson,Lazarsfeld,and McPhee municipal Nazi vote shares in the period 1928-1930. 1954;Campbell,Converse,and Miller 1960)and en- Ciolek-Kumper (1976)focuses on Hitler appearances trenched the minimal effects paradigm that would and-roughly-compares changes in ward-level Nazi dominate mass communication research for decades vote shares at the regional election in Lippe on January (Klapper 1960;Bennett and Iyengar 2008).In a way, 15,1933 relative to the preceding Reichstag election in November 1932.She finds no evidence of the effective- however,the bulk of early empirical work focused on a least likely case.As Iyengar and Simon(2000,151) ness of Hitler's intense campaign efforts.Plockinger (1999)looks at differences between local-and regional- note,identifiable net effects should be limited to highly unbalanced campaigns in which one candidate has a level Nazi vote shares at the July and November 1932 clear resource or skills advantage-a condition that Reichstag elections in Bavaria but does not find any de- viations between municipalities visited by Hitler and is rarely met in U.S.presidential campaigns (also see Gelman and King 1993).Hitler's campaign,by con- those that were not.Though inventive,the latter stud- trast,far exceeded any of his rivals'efforts;the ma- ies are limited in their geographic and temporal scope. nipulative techniques employed were novel and so- They also potentially suffer from causality issues,which received much less attention at the time these studies phisticated,and the use of modern technology,such were conducted than they do today. as aircraft and loudspeakers,guaranteed Hitler an un- In this study,we revisit the question of how effective paralleled geographic penetration and public attention (Paul 1990).Tentative evidence for strong effects is early Nazi propaganda was in garnering electoral sup- indeed compelling:the numerous gag orders already port in Weimar Germany.Our substantive focus is on speak volumes about the authorities'beliefs in Hitler's Hitler's public speeches as the Nazis'chief campaign tool at that time.We rely on extensive original data agitational potency,there were signs of electoral stag- nation and organizational decay when the bans were that has superior geographic and temporal scope and resolution.We draw on the campaign resource alloca- 8 in force,the unparalleled campaign activities that fol- lowed their repeal coincided with a steep electoral rise tion literature and use a semi-parametric difference-in- for the Nazi party,and plenty of reports from contem- differences estimation strategy to account for often ig- nored endogeneity problems in the assessment of local porary witnesses further corroborate the importance of campaign effects.In doing so,we also provide rare in- Hitler's campaign for the Nazis'success (Abel 1965). sight into the campaign strategy of the Nazi party.We Only a handful of studies have thus far attempted discuss the limitations of our study and provide several to systematically assess the effectiveness of early Nazi robustness checks.Finally,we consider the implications propaganda.2 Most recently,Adena and her colleagues (2015)measure local exposure to radio broadcasting of our findings for current research into campaign and using a method for predicting the spatial attenuation leader effects of radio signals.They find that exposure was negatively ESTIMATING CAMPAIGN EFFECTS: 四 related to NSDAP support before the Nazi seizure of PROBLEMS AND STRATEGIES power and positively related soon afterward as the party assumed control over the mass media.In line with Campaign effects on voting behavior and election re- this finding,the authors also provide evidence that ra- sults are notoriously difficult to detect in a campaign dio content before 1933 was largely pro-government realm characterized by the selective exposure of vot- and against the Nazis.These results corroborate the ers to a diffuse stream of countervailing campaign mes- sages.Only since the late 1980s has the paradigmatic view of minimal campaign effects been challenged by 2Voigtlander and Voth(2014,2015)examine the effectiveness of scholars using novel data and sophisticated method- Nazi propaganda under the dictatorship (the years 1933-1945)using ologies such as laboratory experiments(Iyengar and novel data and empirical strategies. Kinder 1987),rolling surveys (Johnston et al.1992), 1051

Hitler’s Speeches and the Rise of the Nazi Party with more than 37% of the national vote in July 1932, effectively paving the way for the Nazi takeover in March 1933. INTRODUCTION After observing the influence Hitler seemed to wield through the use of propaganda, refugee-scholar Paul Lazarsfeld fielded a panel survey in Erie County, Ohio during the run-up to the 1940 U.S. presidential elec￾tion. This marked the beginning of modern research into campaign effects (Hillygus 2010). The study’s findings were surprising—they did not substantiate Lazarsfeld’s motivating concern that campaigns could arbitrarily manipulate the public. Instead, the pres￾idential campaign was found to merely activate the voters’ predispositions (Lazarsfeld, Berelson, and Gaudet 1948). Subsequent studies came to simi￾lar conclusions (Berelson, Lazarsfeld, and McPhee 1954; Campbell, Converse, and Miller 1960) and en￾trenched the minimal effects paradigm that would dominate mass communication research for decades (Klapper 1960; Bennett and Iyengar 2008). In a way, however, the bulk of early empirical work focused on a least likely case. As Iyengar and Simon (2000, 151) note, identifiable net effects should be limited to highly unbalanced campaigns in which one candidate has a clear resource or skills advantage—a condition that is rarely met in U.S. presidential campaigns (also see Gelman and King 1993). Hitler’s campaign, by con￾trast, far exceeded any of his rivals’ efforts; the ma￾nipulative techniques employed were novel and so￾phisticated, and the use of modern technology, such as aircraft and loudspeakers, guaranteed Hitler an un￾paralleled geographic penetration and public attention (Paul 1990). Tentative evidence for strong effects is indeed compelling: the numerous gag orders already speak volumes about the authorities’ beliefs in Hitler’s agitational potency, there were signs of electoral stag￾nation and organizational decay when the bans were in force, the unparalleled campaign activities that fol￾lowed their repeal coincided with a steep electoral rise for the Nazi party, and plenty of reports from contem￾porary witnesses further corroborate the importance of Hitler’s campaign for the Nazis’ success (Abel 1965). Only a handful of studies have thus far attempted to systematically assess the effectiveness of early Nazi propaganda.2 Most recently, Adena and her colleagues (2015) measure local exposure to radio broadcasting using a method for predicting the spatial attenuation of radio signals. They find that exposure was negatively related to NSDAP support before the Nazi seizure of power and positively related soon afterward as the party assumed control over the mass media. In line with this finding, the authors also provide evidence that ra￾dio content before 1933 was largely pro-government and against the Nazis. These results corroborate the 2 Voigtländer and Voth (2014, 2015) examine the effectiveness of Nazi propaganda under the dictatorship (the years 1933–1945) using novel data and empirical strategies. portrayal of early Nazi propaganda by later broadcast￾ing director and Goebbels’ deputy Eugen Hadamovsky (1933, 44): “All the means of public opinion were denied to Hitler. His newspapers were banned, he was denied use of the ra￾dio, his brochures and leaflets were confiscated. He had no choice but to reach the masses directly through constantly growing mass rallies.” Ohr (1997) collects data on local party events in Hes￾sian communities and finds that in the period between the 1930 national and the 1931 regional parliamentary elections, there is a positive relationship in the changes of municipal Nazi vote shares. Wernette (1974) also codes local Nazi election activities during the run-up to the 1930 Reichstag election from a national news￾paper and finds a positive correlation with changes in municipal Nazi vote shares in the period 1928–1930. Ciolek-Kümper (1976) focuses on Hitler appearances and—roughly—compares changes in ward-level Nazi vote shares at the regional election in Lippe on January 15, 1933 relative to the preceding Reichstag election in November 1932. She finds no evidence of the effective￾ness of Hitler’s intense campaign efforts. Plöckinger (1999) looks at differences between local- and regional￾level Nazi vote shares at the July and November 1932 Reichstag elections in Bavaria but does not find any de￾viations between municipalities visited by Hitler and those that were not. Though inventive, the latter stud￾ies are limited in their geographic and temporal scope. They also potentially suffer from causality issues, which received much less attention at the time these studies were conducted than they do today. In this study, we revisit the question of how effective early Nazi propaganda was in garnering electoral sup￾port in Weimar Germany. Our substantive focus is on Hitler’s public speeches as the Nazis’ chief campaign tool at that time. We rely on extensive original data that has superior geographic and temporal scope and resolution. We draw on the campaign resource alloca￾tion literature and use a semi-parametric difference-in￾differences estimation strategy to account for often ig￾nored endogeneity problems in the assessment of local campaign effects. In doing so, we also provide rare in￾sight into the campaign strategy of the Nazi party. We discuss the limitations of our study and provide several robustness checks. Finally, we consider the implications of our findings for current research into campaign and leader effects. ESTIMATING CAMPAIGN EFFECTS: PROBLEMS AND STRATEGIES Campaign effects on voting behavior and election re￾sults are notoriously difficult to detect in a campaign realm characterized by the selective exposure of vot￾ers to a diffuse stream of countervailing campaign mes￾sages. Only since the late 1980s has the paradigmatic view of minimal campaign effects been challenged by scholars using novel data and sophisticated method￾ologies such as laboratory experiments (Iyengar and Kinder 1987), rolling surveys (Johnston et al. 1992), 1051 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Peter Selb and Simon Munzert field experiments (Gerber and Green 2000),and nat- that group-level observations can be highly misleading ural experiments(Huber and Arceneaux 2007). when aggregate data are used to make inferences about While the historical perspective of our research pre- individuals.Even under unconfoundedness (e.g.,by cludes attractive design options that employ random- random assignment of campaign appearances),we can- ization or survey data,3 a number of recent stud- not unambiguously attribute higher turnout or voter ies try to gauge the impact of campaign stops on support in visited localities to increased propensities voting behavior by using widely available informa- among those who attended the campaign events to turn tion about the candidates'campaign itineraries and out and vote for the candidate.Such aggregate effects local-level election results (Campbell 2008;Herr 2002; may also come about.for instance,in the fashion of Hill,Rodriquez,and Wooden 2010;Holbrook 2002; an indirect two-step flow of communication (Katz and Jones 1998;King and Morehouse 2004;Sellers and Lazarsfeld 1955),in which opinion leaders who would Denton 2006;Vavreck,Spiliotes,and Fowler 2002). turn out and support their candidate anyway (i.e.,for Such studies typically struggle with identification is- whom the individual effect of attendance is essentially sues that challenge causal claims.Like any observa- zero)attend the event and are then motivated to mo- tional study,they are subject to potential confound- bilize and persuade others within their personal net- ing (see Goldstein and Holleque 2010).Confound- works (Rosenstone and Hansen 1993).In our partic- ing would occur,for instance,if candidates and their ular case,positive effects of Hitler's campaign visits staff deliberately selected locations for their appear- on local Nazi vote shares might also have occurred in- ances where they expected a large pool of easy-to- directly through intimidation.As Childers and Weiss mobilize supporters or anticipated a close race.If a re (1990)document,violence was an integral part of Nazi searcher failed to properly take into account such con- mobilization strategy at the end of the Weimar Repub- founders (e.g.,latent support,marginality),she would lic.If Hitler's appearances were regularly accompanied probably overestimate the effect of appearances on by assaults on political opponents,increases in Nazi the candidate vote share and voter turnout.We use a vote shares at the following election could have been 4号元 semi-parametric difference-in-differences estimation the result of selective abstention by supporters of op- strategy to account for potential confounding due to ponent parties.Either way,campaign effects on local- observed and unobserved variables (see Abadie 2005: level election outcomes are,like other neighborhood Heckman,Ichimura,and Todd 1997).In doing so,we effects,emergent properties of the social interaction of specify a parametric model to predict the probability of the residents (Oakes 2004).Therefore,one has to be Hitler appearances in a geographic unit and match vis- cautious not to interpret even internally valid aggre- ited units with control units that feature a similar pre- gate estimates of the effect of campaign appearances dicted probability before the difference-in-differences on election results in terms of the impact of individual analysis. attendance on voting behavior. Beyond potential confounding,such studies of can- Finally,spatial ecological studies potentially suffer didate appearances may be considered what epidemi- from ambiguities in separating exposure from nonex- ologists call spatial ecological studies (Wakefield 2004). posure units.Effects of campaign events need not be re- Spatial ecological studies use geographic proximity to stricted to the areal units for which we observe the out- a presumed cause(in our case:campaign appearances) comes of interest.For instance,voters and opinion lead- 675:.101 as a surrogate for individual exposure to the cause ers from neighboring units may also attend the events (attendance to the campaign event)and measure the and thereby carry individual and network effects back response (voting behavior)at the level of geographic home.Likewise,the geographic range of news media units(communities or counties).5 A number of addi- that cover the events may well exceed the borders of tional biases may arise in such a design.It is well known the units of analysis.Such spatial spillovers would vio- late the non-interference assumption underlying most methods for causal inference.Non-interference is an See Collier (1944)for an early(non-randomized)experiment on the attitudinal effects of Nazi propaganda materials on a sample of essential aspect of the stable unit-treatment value as- U.S.college students in 1941-1942.Also see Reuband (2006),who sumption (SUTVA),which implies that a treatment uses a retrospective survey conducted in 1949 to assess mass support applied to one unit does not affect the outcome of during the Nazi regime. other units.This allows researchers to employ multiple 4 In their original study.,Shaw and Gimpel(2012)randomize a can. units for estimating causal effects(Rubin 1980).To il- didate's travel schedule during the 2006 Texas gubernatorial race to make campaign appearances statistically independent of other fac lustrate the implications,imagine that Hitler's appear- 四 tors related to the outcome of interest.While such a randomized field ances actually had their intended effect on Nazi sup- experiment is a powerful design for valid causal inference,even the port in the visited county,but that this effect carried authors seem surprised that the candidate's staff actually agreed to over to neighboring counties through travel activity, let scholars interfere in their strategic planning (Shaw and Gimpel 2012,140).Moreover,this is an apparently infeasible approach for a personal networks,or media coverage.If these neigh- retrospective study like ours. boring counties served as controls when assessing the Shaw and Gimpel(2012)field a large-scale survey of registered effect of Hitler's appearance on the NSDAP vote in voters that includes items on both exposure to the campaign events the exposure county,the effect estimate would obvi- and candidate support.Such data would have allowed them to esti- ously be biased downward because the average over- mate causal effects of individual exposure by using an instrumental variable approach(see Angrist,Imbens,and Rubin 1996).However time difference in outcomes among control units would Shaw and Gimpel (2012)limit their empirical analysis to before- not properly reflect the expected developments in after comparisons within and between geographic units. the absence of the appearance.We,therefore,exclude 1052

Peter Selb and Simon Munzert field experiments (Gerber and Green 2000), and nat￾ural experiments (Huber and Arceneaux 2007). While the historical perspective of our research pre￾cludes attractive design options that employ random￾ization or survey data,3 a number of recent stud￾ies try to gauge the impact of campaign stops on voting behavior by using widely available informa￾tion about the candidates’ campaign itineraries and local-level election results (Campbell 2008; Herr 2002; Hill, Rodriquez, and Wooden 2010; Holbrook 2002; Jones 1998; King and Morehouse 2004; Sellers and Denton 2006; Vavreck, Spiliotes, and Fowler 2002). Such studies typically struggle with identification is￾sues that challenge causal claims. Like any observa￾tional study, they are subject to potential confound￾ing (see Goldstein and Holleque 2010). Confound￾ing would occur, for instance, if candidates and their staff deliberately selected locations for their appear￾ances where they expected a large pool of easy-to￾mobilize supporters or anticipated a close race. If a re￾searcher failed to properly take into account such con￾founders (e.g., latent support, marginality), she would probably overestimate the effect of appearances on the candidate vote share and voter turnout.4 We use a semi-parametric difference-in-differences estimation strategy to account for potential confounding due to observed and unobserved variables (see Abadie 2005; Heckman, Ichimura, and Todd 1997). In doing so, we specify a parametric model to predict the probability of Hitler appearances in a geographic unit and match vis￾ited units with control units that feature a similar pre￾dicted probability before the difference-in-differences analysis. Beyond potential confounding, such studies of can￾didate appearances may be considered what epidemi￾ologists call spatial ecological studies (Wakefield 2004). Spatial ecological studies use geographic proximity to a presumed cause (in our case: campaign appearances) as a surrogate for individual exposure to the cause (attendance to the campaign event) and measure the response (voting behavior) at the level of geographic units (communities or counties).5 A number of addi￾tional biases may arise in such a design. It is well known 3 See Collier (1944) for an early (non-randomized) experiment on the attitudinal effects of Nazi propaganda materials on a sample of U.S. college students in 1941-1942. Also see Reuband (2006), who uses a retrospective survey conducted in 1949 to assess mass support during the Nazi regime. 4 In their original study, Shaw and Gimpel (2012) randomize a can￾didate’s travel schedule during the 2006 Texas gubernatorial race to make campaign appearances statistically independent of other fac￾tors related to the outcome of interest.While such a randomized field experiment is a powerful design for valid causal inference, even the authors seem surprised that the candidate’s staff actually agreed to let scholars interfere in their strategic planning (Shaw and Gimpel 2012, 140). Moreover, this is an apparently infeasible approach for a retrospective study like ours. 5 Shaw and Gimpel (2012) field a large-scale survey of registered voters that includes items on both exposure to the campaign events and candidate support. Such data would have allowed them to esti￾mate causal effects of individual exposure by using an instrumental variable approach (see Angrist, Imbens, and Rubin 1996). However, Shaw and Gimpel (2012) limit their empirical analysis to before– after comparisons within and between geographic units. that group-level observations can be highly misleading when aggregate data are used to make inferences about individuals. Even under unconfoundedness (e.g., by random assignment of campaign appearances), we can￾not unambiguously attribute higher turnout or voter support in visited localities to increased propensities among those who attended the campaign events to turn out and vote for the candidate. Such aggregate effects may also come about, for instance, in the fashion of an indirect two-step flow of communication (Katz and Lazarsfeld 1955), in which opinion leaders who would turn out and support their candidate anyway (i.e., for whom the individual effect of attendance is essentially zero) attend the event and are then motivated to mo￾bilize and persuade others within their personal net￾works (Rosenstone and Hansen 1993). In our partic￾ular case, positive effects of Hitler’s campaign visits on local Nazi vote shares might also have occurred in￾directly through intimidation. As Childers and Weiss (1990) document, violence was an integral part of Nazi mobilization strategy at the end of the Weimar Repub￾lic. If Hitler’s appearances were regularly accompanied by assaults on political opponents, increases in Nazi vote shares at the following election could have been the result of selective abstention by supporters of op￾ponent parties. Either way, campaign effects on local￾level election outcomes are, like other neighborhood effects, emergent properties of the social interaction of the residents (Oakes 2004). Therefore, one has to be cautious not to interpret even internally valid aggre￾gate estimates of the effect of campaign appearances on election results in terms of the impact of individual attendance on voting behavior. Finally, spatial ecological studies potentially suffer from ambiguities in separating exposure from nonex￾posure units.Effects of campaign events need not be re￾stricted to the areal units for which we observe the out￾comes of interest.For instance, voters and opinion lead￾ers from neighboring units may also attend the events and thereby carry individual and network effects back home. Likewise, the geographic range of news media that cover the events may well exceed the borders of the units of analysis. Such spatial spillovers would vio￾late the non-interference assumption underlying most methods for causal inference. Non-interference is an essential aspect of the stable unit-treatment value as￾sumption (SUTVA), which implies that a treatment applied to one unit does not affect the outcome of other units. This allows researchers to employ multiple units for estimating causal effects (Rubin 1980). To il￾lustrate the implications, imagine that Hitler’s appear￾ances actually had their intended effect on Nazi sup￾port in the visited county, but that this effect carried over to neighboring counties through travel activity, personal networks, or media coverage. If these neigh￾boring counties served as controls when assessing the effect of Hitler’s appearance on the NSDAP vote in the exposure county, the effect estimate would obvi￾ously be biased downward because the average over￾time difference in outcomes among control units would not properly reflect the expected developments in the absence of the appearance. We, therefore, exclude 1052 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Hitler's Speeches and the Rise of the Nazi Party neighboring areas from the pool of controls to account minds us that a leading candidate's campaign is not a for possible spillovers unitary entity but rather a set of efforts undertaken by various agents.For instance,Kelley (1961)found in his study of U.S.presidential campaigns that contenders THE TARGETING OF CANDIDATE used their running mates'schedules in either comple- APPEARANCES mentary or duplicative ways.Geographically or tem- porally complementary campaign schedules carry the In light of the potential confounding factors discussed risk of offsetting campaign effects(Finkel 1993),while above,the first step in assessing the impact of can- duplicative itineraries bring about potential misattri- didate appearances on election returns is to theorize bution of campaign effects.One way or the other,we how such visits are being targeted and in what way need to account for the activities of other Nazi speak- these factors relate to the outcome of interest (Althaus ers when assessing the effects of Hitler's public appear- Nardulli,and Shaw 2002).In doing so,we adopt an in- ances.The following section introduces measurements strumental view and assume that campaign activities for all relevant confounders serve to maximize votes while accounting for mobiliza- tion costs(Brams and Davis 1974:Colantoni.Levesque. and Ordeshook 1975;Cox 1999).Candidates should fo- DATA COLLECTION AND MEASUREMENT cus their scarce resources accordingly-that is,on lo- Period of observation.Our empirical analysis covers cations they expect personal appearances to favorably the period between the repeal of the speaking ban translate into additional votes,and specifically where on Hitler in Bavaria on March 6.1927 and the last additional votes would be decisive for winning man- (halfway)competitive Reichstag election on March 5, dates.The former suggests that rational campaigners 1933.5 Five national parliamentary elections (on May primarily target locales with large numbers of eligibles 20.1928:September 14,1930:July 31.1932:November 6 and a high expected share of supportive voters.The lat- 1932;and March 5,1933)and a two-round presidential 4号 ter expectation implies that candidates are more likely election (on March 13,1932 and April 10,1932)were to visit competitive districts in which small vote shifts held within this period.Our difference-in-differences could change the allocation of mandates in their favor approach focuses on changes between the four consec- or to their detriment.All these factors-the number utive parliamentary elections and both rounds of the of eligibles,expected electoral support,and a party's presidential election,respectively. expected competitiveness-are potential confounders Areal units.The availability of election statistics insomuch as they may also influence the outcome of in- dictates our choice of areal units.Thanks to a data terest.Applied to the present case,the size of the local collection effort of epic proportions by Jurgen Falter electorate may be negatively linked to Nazi vote shares and collaborators (Falter and Hanisch 1990),digital- since rural areas were less populous and,at the same ized community-level election statistics are available 是 time,on average more supportive of the NSDAP than for the 1928,1930,and 1933 elections;however,they urban areas for programmatic reasons(Heberle 1978; are not available for the elections in 1932,whose re- Thurner,Klima,and Kuchenhoff 2015).Likewise,the sults the Reich Statistical Office (Statistisches Reich- classical decision-theoretic model of voting suggests samt)reported only for the higher administrative levels 5795.801g that a party's expected competitiveness may directly of counties (Kreise)and county boroughs (kreisfreie affect the relative strength of parties through selective Stadte)(see Hanisch 1989,45).This leaves us with participation and strategic voting(Cox 1999). a single election pair (1928-1930)for a community- Local infrastructure is also relevant to mobilization level difference-in-differences analysis.We will use the costs.In the context of U.S.presidential campaigns, community-level data(N=3,864),among other things, Holbrook's(2002)study of Truman's 1948 whistle-stop to check the sensitivity of our empirical results for campaign and Althaus,Nardulli,and Shaw's (2002) potential violations of the non-interference assump- narratives of the boat trips down the Mississippi river tion discussed above.Other than this,our analytical by Al Gore in 2000 and by George H.W.Bush in 1988 focus will be on the counties and county boroughs(N provide good examples of campaigns in which ground =1.000).While election statistics are available at the transportation connectivity mattered.A remarkable county and,partly,the municipal level,neither level feature of Hitler's 1932 campaign was that,for the was relevant for the apportionment of parliamentary first time in history,he chartered a plane to transport seats.Mandates were allocated at the level of the 35 四 him to certain campaign events.Under the ambigu- primary districts (Wahlkreise)and 16 secondary dis- ous label "Hitler over Germany,"he made nearly 150 tricts (Wahlkreisverbande),which will serve as addi- appearances from April until November 1932.There. tional geographical layers to compute measures of the fore,distance to the nearest airfield should have mat NSDAP's competitiveness. tered for targeting event locations starting with the Areal units are key in generating and combining sub- 1932 elections.The strength of local party organiza- stantive variables.Unfortunately,the Falter data do not tions is often quoted as an important source of logistic support and secondary mobilization(Cox 1999;Rosen- stone and Hansen 1993).Some authors even consider 6 Gag orders in other regions were remitted successively:Saxony local organizational strength as another campaign tool on January 26,1927;Hamburg on March 23,1927;Baden on Apri 22,1927;Luibeck on May 18,1927;and both Anhalt and Prussia on subject to strategic allocation(Bartels 1985),which re- November 16,1928(Bruppacher 2012,181-198). 1053

Hitler’s Speeches and the Rise of the Nazi Party neighboring areas from the pool of controls to account for possible spillovers. THE TARGETING OF CANDIDATE APPEARANCES In light of the potential confounding factors discussed above, the first step in assessing the impact of can￾didate appearances on election returns is to theorize how such visits are being targeted and in what way these factors relate to the outcome of interest (Althaus, Nardulli, and Shaw 2002). In doing so, we adopt an in￾strumental view and assume that campaign activities serve to maximize votes while accounting for mobiliza￾tion costs (Brams and Davis 1974; Colantoni,Levesque, and Ordeshook 1975; Cox 1999). Candidates should fo￾cus their scarce resources accordingly—that is, on lo￾cations they expect personal appearances to favorably translate into additional votes, and specifically where additional votes would be decisive for winning man￾dates. The former suggests that rational campaigners primarily target locales with large numbers of eligibles and a high expected share of supportive voters. The lat￾ter expectation implies that candidates are more likely to visit competitive districts in which small vote shifts could change the allocation of mandates in their favor or to their detriment. All these factors—the number of eligibles, expected electoral support, and a party’s expected competitiveness—are potential confounders insomuch as they may also influence the outcome of in￾terest. Applied to the present case, the size of the local electorate may be negatively linked to Nazi vote shares since rural areas were less populous and, at the same time, on average more supportive of the NSDAP than urban areas for programmatic reasons (Heberle 1978; Thurner, Klima, and Küchenhoff 2015). Likewise, the classical decision-theoretic model of voting suggests that a party’s expected competitiveness may directly affect the relative strength of parties through selective participation and strategic voting (Cox 1999). Local infrastructure is also relevant to mobilization costs. In the context of U.S. presidential campaigns, Holbrook’s (2002) study of Truman’s 1948 whistle-stop campaign and Althaus, Nardulli, and Shaw’s (2002) narratives of the boat trips down the Mississippi river by Al Gore in 2000 and by George H.W. Bush in 1988 provide good examples of campaigns in which ground transportation connectivity mattered. A remarkable feature of Hitler’s 1932 campaign was that, for the first time in history, he chartered a plane to transport him to certain campaign events. Under the ambigu￾ous label “Hitler over Germany,” he made nearly 150 appearances from April until November 1932. There￾fore, distance to the nearest airfield should have mat￾tered for targeting event locations starting with the 1932 elections. The strength of local party organiza￾tions is often quoted as an important source of logistic support and secondary mobilization (Cox 1999; Rosen￾stone and Hansen 1993). Some authors even consider local organizational strength as another campaign tool subject to strategic allocation (Bartels 1985), which re￾minds us that a leading candidate’s campaign is not a unitary entity but rather a set of efforts undertaken by various agents. For instance, Kelley (1961) found in his study of U.S. presidential campaigns that contenders used their running mates’ schedules in either comple￾mentary or duplicative ways. Geographically or tem￾porally complementary campaign schedules carry the risk of offsetting campaign effects (Finkel 1993), while duplicative itineraries bring about potential misattri￾bution of campaign effects. One way or the other, we need to account for the activities of other Nazi speak￾ers when assessing the effects of Hitler’s public appear￾ances. The following section introduces measurements for all relevant confounders. DATA COLLECTION AND MEASUREMENT Period of observation. Our empirical analysis covers the period between the repeal of the speaking ban on Hitler in Bavaria on March 6, 1927 and the last (halfway) competitive Reichstag election on March 5, 1933.6 Five national parliamentary elections (on May 20, 1928; September 14, 1930; July 31, 1932;November 6, 1932; and March 5, 1933) and a two-round presidential election (on March 13, 1932 and April 10, 1932) were held within this period. Our difference-in-differences approach focuses on changes between the four consec￾utive parliamentary elections and both rounds of the presidential election, respectively. Areal units. The availability of election statistics dictates our choice of areal units. Thanks to a data collection effort of epic proportions by Jürgen Falter and collaborators (Falter and Hänisch 1990), digital￾ized community-level election statistics are available for the 1928, 1930, and 1933 elections; however, they are not available for the elections in 1932, whose re￾sults the Reich Statistical Office (Statistisches Reich￾samt) reported only for the higher administrative levels of counties (Kreise) and county boroughs (kreisfreie Städte) (see Hänisch 1989, 45). This leaves us with a single election pair (1928–1930) for a community￾level difference-in-differences analysis. We will use the community-level data (N = 3, 864), among other things, to check the sensitivity of our empirical results for potential violations of the non-interference assump￾tion discussed above. Other than this, our analytical focus will be on the counties and county boroughs (N = 1,000). While election statistics are available at the county and, partly, the municipal level, neither level was relevant for the apportionment of parliamentary seats. Mandates were allocated at the level of the 35 primary districts (Wahlkreise) and 16 secondary dis￾tricts (Wahlkreisverbände), which will serve as addi￾tional geographical layers to compute measures of the NSDAP’s competitiveness. Areal units are key in generating and combining sub￾stantive variables.Unfortunately, the Falter data do not 6 Gag orders in other regions were remitted successively: Saxony on January 26, 1927; Hamburg on March 23, 1927; Baden on April 22, 1927; Lübeck on May 18, 1927; and both Anhalt and Prussia on November 16, 1928 (Bruppacher 2012, 181–198). 1053 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Peter Selb and Simon Munzert readily contain the geographic information necessary maintained by a dutch amateur historian that provides to do so.?In an independent data collection effort,we a daily account of Adolf Hitler's life compiled from var- retrieve the center coordinates of the municipalities ious sources.In sum.we are able to identify 566 Hitler and county boroughs included in Falter and Hanisch appearances in the specified period,455 of which were (1990)from the Google Maps API service(Google Inc public.s Figure 1 provides a timeline of events for the 2015).Whole counties are georeferenced in terms of entire observation period.Apparently,Hitler's cam- the center coordinates of their main towns or cities to paign activities had already been intense in the period account for the uneven population distribution within between the 1928 and 1930 elections(84 appearances) counties.Details and links to computer code can be and culminated in the run-up to the 1932 presiden- found in Appendix A. tial and Reichstag elections(243 appearances).In early Outcome variables.The primary outcome of interest 1933,as Hitler had already been appointed head of gov- is the number of NSDAP votes in a Reichstag election ernment,the NSDAP's strategy quickly switched from in an areal unit relative to the voting-eligible electorate. winning votes to repression (Evans 2005).Of the mea- The reason for standardizing against the number of ger 36 events recorded between November 7,1932 and eligible-not actual-voters is that the specific appor- March 5,1933,as many as 16 related to the January 15, tionment method used in Reichstag elections guaran- 1933 regional elections in Lippe.10 teed each party one seat per 60,000 votes (see Ap- To assess the effect of Hitler's appearances on lo- pendix E).Therefore,an increase in the number o cal election returns,we would ideally know whether or votes from one election to the next would increase a not (or rather:to what extent)the local population for party's chances of winning additional seats irrespec- which we observe the outcome had been exposed to tive of the vote gains and losses for other parties.We the event.This information is simply not available.Like also use the vote share of the communist party (Kom- other spatial ecological studies,we use spatial prox- munistische Partei Deutschlands.KPD)as the Nazi's imity to the events as a surrogate for exposure.First major adversary and the turnout rate as alternative Hitler appearances are point-referenced according to 4 outcomes.For the 1932 presidential elections,we use their location(municipality or borough center,possi- vote shares for Adolf Hitler and Ernst Thalmann, bly even the specific venue)using the Google Maps the KPD's candidate.as outcomes of interest.The rel- API service.We then classify those counties and county ative size of the areal units'voting-eligible popula boroughs (municipalities in the municipal-level anal- tion serves as a weighting variable in the difference- ysis)whose center coordinates were situated within a in-differences analysis.Given its vast variability across radius of ten kilometers from an event location at any units-county-level electoral populations range be- time during an interelection period as exposure units tween 184 and 871,764,with a mean of 41,358 Alternative specifications of the exposure variable with eligibles-appropriate weights are imperative for our varying geographical and temporal scope will be con- analysis(Ridgeway et al.2015).All election data are sidered in the section on robustness taken from Falter and Hanisch(1990) Competitiveness.The rational-actor perspective Exposure.Our main source of information about suggests that Hitler and the NSDAP focused their Hitler's public appearances is the multivolume edi- scarce campaign resources on close races-that is, tion"Hitler.Reden,Schriften,Anordnungen"(Hitler. on electoral districts where the stakes of winning 5795.801g Speeches,Writings,Directives)by the Institute for additional seats and of losing seats won in previous Contemporary History Munich (Hitler 1992,1994a, elections were high.To measure the NSDAP's loca 1994b,1994c,1995,1996,1997,1998).For each appear-- electoral stakes,we follow lines similar to Grofman ance within the specified period,we hand code its date and Selb (2009)and develop party-specific com- location,type of event(public or private)and,if avail- petition indices that are sensitive to the nature of able,attendance figures by source(police reports,press the automatic apportionment method used at that coverage,Nazi press coverage).As the edition only time (Schanbacher 1982).The indices are specified covers the period until Hitler's appointment as head at the level of the primary and secondary electoral of the national government on January 30,1933,we districts(Competitiveness 1 and 2,respectively).They consult Bruppacher(2012)to fill the gap until the reflect the minimum vote distance of the NSDAP Reichstag election on March 5,1933.For validation to winning an additional seat or to losing their final we use Bruppacher(2012)and Domarus (1973)and seat at the respective level at the previous election retrieve information from hitlerpages.com,a website relative to the highest possible vote distance to a seat gain or loss.An index value of one indicates maximum competitiveness,a value of zero indicates 7 O'Loughlin and colleagues(O'Loughlin,Flint,and Anselin 1994: the theoretical minimum.Details of the calculations O'Loughlin and Shin 1995)digitalize county boundaries for the 1930 Reichstag election based on U.S.military maps from World War II and descriptive statistics are given in Appendix E. However,this map does not contain geographic information about most county boroughs and about none of the communities.The for- L mer would lead to the exclusion of a substantial share of the voting eligible population from the analysis.The latter would prevent us s Most of the events we classified as nonpublic were speeches on the from disaggregating our study to the municipality level.Paul Thurner occasion of meetings with party officials. and his colleagues (2015)made a fresh geocoding effort,including A detailed description of the coding procedure and additional fig all the counties and county boroughs.For the last-mentioned reason. ures can be found in Appendix B. however,we do not use their materials either. 10 See the clustering of events in Figure B1 in the Appendix. 1054

Peter Selb and Simon Munzert readily contain the geographic information necessary to do so.7 In an independent data collection effort, we retrieve the center coordinates of the municipalities and county boroughs included in Falter and Hänisch (1990) from the Google Maps API service (Google Inc. 2015). Whole counties are georeferenced in terms of the center coordinates of their main towns or cities to account for the uneven population distribution within counties. Details and links to computer code can be found in Appendix A. Outcome variables. The primary outcome of interest is the number of NSDAP votes in a Reichstag election in an areal unit relative to the voting-eligible electorate. The reason for standardizing against the number of eligible—not actual—voters is that the specific appor￾tionment method used in Reichstag elections guaran￾teed each party one seat per 60,000 votes (see Ap￾pendix E). Therefore, an increase in the number of votes from one election to the next would increase a party’s chances of winning additional seats irrespec￾tive of the vote gains and losses for other parties. We also use the vote share of the communist party (Kom￾munistische Partei Deutschlands, KPD) as the Nazi’s major adversary and the turnout rate as alternative outcomes. For the 1932 presidential elections, we use vote shares for Adolf Hitler and Ernst Thälmann, the KPD’s candidate, as outcomes of interest. The rel￾ative size of the areal units’ voting-eligible popula￾tion serves as a weighting variable in the difference￾in-differences analysis. Given its vast variability across units—county-level electoral populations range be￾tween 184 and 871,764, with a mean of 41,358 eligibles—appropriate weights are imperative for our analysis (Ridgeway et al. 2015). All election data are taken from Falter and Hänisch (1990). Exposure. Our main source of information about Hitler’s public appearances is the multivolume edi￾tion “Hitler. Reden, Schriften, Anordnungen” (Hitler. Speeches, Writings, Directives) by the Institute for Contemporary History Munich (Hitler 1992, 1994a, 1994b, 1994c, 1995, 1996, 1997, 1998). For each appear￾ance within the specified period, we hand code its date, location, type of event (public or private) and, if avail￾able, attendance figures by source (police reports, press coverage, Nazi press coverage). As the edition only covers the period until Hitler’s appointment as head of the national government on January 30, 1933, we consult Bruppacher (2012) to fill the gap until the Reichstag election on March 5, 1933. For validation we use Bruppacher (2012) and Domarus (1973) and retrieve information from hitlerpages.com, a website 7 O’Loughlin and colleagues (O’Loughlin, Flint, and Anselin 1994; O’Loughlin and Shin 1995) digitalize county boundaries for the 1930 Reichstag election based on U.S. military maps from World War II. However, this map does not contain geographic information about most county boroughs and about none of the communities. The for￾mer would lead to the exclusion of a substantial share of the voting eligible population from the analysis. The latter would prevent us from disaggregating our study to the municipality level. Paul Thurner and his colleagues (2015) made a fresh geocoding effort, including all the counties and county boroughs. For the last-mentioned reason, however, we do not use their materials either. maintained by a Dutch amateur historian that provides a daily account of Adolf Hitler’s life compiled from var￾ious sources. In sum, we are able to identify 566 Hitler appearances in the specified period, 455 of which were public.8 Figure 1 provides a timeline of events for the entire observation period.9 Apparently, Hitler’s cam￾paign activities had already been intense in the period between the 1928 and 1930 elections (84 appearances) and culminated in the run-up to the 1932 presiden￾tial and Reichstag elections (243 appearances). In early 1933, as Hitler had already been appointed head of gov￾ernment, the NSDAP’s strategy quickly switched from winning votes to repression (Evans 2005). Of the mea￾ger 36 events recorded between November 7, 1932 and March 5, 1933, as many as 16 related to the January 15, 1933 regional elections in Lippe.10 To assess the effect of Hitler’s appearances on lo￾cal election returns, we would ideally know whether or not (or rather: to what extent) the local population for which we observe the outcome had been exposed to the event.This information is simply not available.Like other spatial ecological studies, we use spatial prox￾imity to the events as a surrogate for exposure. First, Hitler appearances are point-referenced according to their location (municipality or borough center, possi￾bly even the specific venue) using the Google Maps API service.We then classify those counties and county boroughs (municipalities in the municipal-level anal￾ysis) whose center coordinates were situated within a radius of ten kilometers from an event location at any time during an interelection period as exposure units. Alternative specifications of the exposure variable with varying geographical and temporal scope will be con￾sidered in the section on robustness. Competitiveness. The rational-actor perspective suggests that Hitler and the NSDAP focused their scarce campaign resources on close races—that is, on electoral districts where the stakes of winning additional seats and of losing seats won in previous elections were high. To measure the NSDAP’s local electoral stakes, we follow lines similar to Grofman and Selb (2009) and develop party-specific com￾petition indices that are sensitive to the nature of the automatic apportionment method used at that time (Schanbacher 1982). The indices are specified at the level of the primary and secondary electoral districts (Competitiveness 1 and 2, respectively). They reflect the minimum vote distance of the NSDAP to winning an additional seat or to losing their final seat at the respective level at the previous election relative to the highest possible vote distance to a seat gain or loss. An index value of one indicates maximum competitiveness, a value of zero indicates the theoretical minimum. Details of the calculations and descriptive statistics are given in Appendix E. 8 Most of the events we classified as nonpublic were speeches on the occasion of meetings with party officials. 9 A detailed description of the coding procedure and additional fig￾ures can be found in Appendix B. 10 See the clustering of events in Figure B1 in the Appendix. 1054 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Hitler's Speeches and the Rise of the Nazi Party FIGURE 1.Timeline of Hitler's public appearances,March 1927 to March 1933. Elections(R=Reichstag,P=Presidential) 5 3 2 1927 1928 1929 1930 1931 1932 1933 Organizational strength.It has been argued that lo- ers.Therefore,we include the number of eligibles (in cal infrastructure is relevant to mobilization costs and 100,000)and the vote share of the NSDAP (for the that the strength of local party organization is one 1932 presidential election:Hitler)at the previous elec- important infrastructural factor.Apart from some re- tion (i.e.,the first election in each election pair con- 4号 gional studies (Anheier 2003;Anheier,Neidhardt,and sidered).Information on both variables is taken from Vortkamp 1998),there is no systematic nationwide in- Falter and Hanisch(1990). formation about the local organizational strength of Goebbels'appearances.To account for the eventual- 'asn the NSDAP and its development over time.As a proxy ity that Hitler's campaign schedule was complemented for organizational strength,we estimate county-level or duplicated by those of other high-profile Nazi speak- party membership totals (in 1,000)based on samples ers,we collect information about the public appear- drawn from the two original NSDAP member files ances of Joseph Goebbels,the second most important archived at the Berlin Document Center by teams of Nazi speaker after Hitler.We first conduct an auto- researchers in Berlin(Falter and Kater 1993)and Min. matic search of keywords12 using a digital version of his neapolis(Brustein 1998).The sample data include,in- diaries(Goebbels 1992)and then manually code infor- ter alia,information about the place and date that the mation on places,dates,and types of speeches(public members joined.The thorough description of the sam or private).Finally,we geocode the appearances using pling procedures in Schneider-Haase (1991)allow us the Google Maps API service.In total,we are able to to calculate appropriate design weights.Unfortunately, collect data on 200 public speeches,an overwhelming 575.1018 the researchers used fixed yearly quotas for entries in majority of which(110)were held in Berlin.Figure B2 the period 1930-1933 so that it is impossible to calcu- in the Appendix maps the locations of Goebbels' late time-varying weights.Auxiliary information about appearances. the annual development of national membership fig- Previous appearances.In addition to the matching ures from Kater (1980)is used to generate election- variables listed,we include a binary variable that in- specific estimates.Details of the estimation procedure dicates whether Hitler previously (i.e.,before the last and descriptive statistics are given in Appendix F. election)visited a county to help control for unob- Distance to nearest airfield.Of particular impor- served confounders,assuming that those factors had tance for mobilization costs as of 1932 is the distance already affected past targeting decisions.Election- to the nearest airfield.We consult several Wikipedia specific summary statistics of all the variables and sup- entries and a privately run website!to identify a total plementary maps are given in Appendix H. of 70 civilian airfields in operation at that time.We use the Google Maps API service to geocode the airfields, which provides the basis for calculating minimum dis- PREDICTING HITLER'S APPEARANCES tances to an airfield (in 100 km)for each municipal- In the first stage of our empirical analysis,we model ity,county borough,and county.Figure Gl in the Ap- the election-specific probability of a Hitler visit to a pendix maps the location of the airfields as well as the community or county as a function of the above covari- minimum distances. ates.Predictions from these models will then be used Number of eligibles and previous vote share.Our in the causal inference step to trim the sample to in- theoretical considerations suggest that rational cam- clude as controls only those counties and communities paigners primarily target locales with large numbers of eligibles and a high expected share of supportive vot- 12 The list of keywords includes(root)words related to speeches and rallies:sprech,gesprochen,rede,kundgebung,ansprache,veranstal- 11 http://www.forgottenairfields.com/. tung,vortrag. 1055

Hitler’s Speeches and the Rise of the Nazi Party FIGURE 1. Timeline of Hitler’s public appearances, March 1927 to March 1933. Organizational strength. It has been argued that lo￾cal infrastructure is relevant to mobilization costs and that the strength of local party organization is one important infrastructural factor. Apart from some re￾gional studies (Anheier 2003; Anheier, Neidhardt, and Vortkamp 1998), there is no systematic nationwide in￾formation about the local organizational strength of the NSDAP and its development over time. As a proxy for organizational strength, we estimate county-level party membership totals (in 1,000) based on samples drawn from the two original NSDAP member files archived at the Berlin Document Center by teams of researchers in Berlin (Falter and Kater 1993) and Min￾neapolis (Brustein 1998). The sample data include, in￾ter alia, information about the place and date that the members joined. The thorough description of the sam￾pling procedures in Schneider-Haase (1991) allow us to calculate appropriate design weights. Unfortunately, the researchers used fixed yearly quotas for entries in the period 1930–1933 so that it is impossible to calcu￾late time-varying weights. Auxiliary information about the annual development of national membership fig￾ures from Kater (1980) is used to generate election￾specific estimates. Details of the estimation procedure and descriptive statistics are given in Appendix F. Distance to nearest airfield. Of particular impor￾tance for mobilization costs as of 1932 is the distance to the nearest airfield. We consult several Wikipedia entries and a privately run website11 to identify a total of 70 civilian airfields in operation at that time. We use the Google Maps API service to geocode the airfields, which provides the basis for calculating minimum dis￾tances to an airfield (in 100 km) for each municipal￾ity, county borough, and county. Figure G1 in the Ap￾pendix maps the location of the airfields as well as the minimum distances. Number of eligibles and previous vote share. Our theoretical considerations suggest that rational cam￾paigners primarily target locales with large numbers of eligibles and a high expected share of supportive vot- 11 http://www.forgottenairfields.com/. ers. Therefore, we include the number of eligibles (in 100,000) and the vote share of the NSDAP (for the 1932 presidential election: Hitler) at the previous elec￾tion (i.e., the first election in each election pair con￾sidered). Information on both variables is taken from Falter and Hänisch (1990). Goebbels’ appearances.To account for the eventual￾ity that Hitler’s campaign schedule was complemented or duplicated by those of other high-profile Nazi speak￾ers, we collect information about the public appear￾ances of Joseph Goebbels, the second most important Nazi speaker after Hitler. We first conduct an auto￾matic search of keywords12 using a digital version of his diaries (Goebbels 1992) and then manually code infor￾mation on places, dates, and types of speeches (public or private). Finally, we geocode the appearances using the Google Maps API service. In total, we are able to collect data on 200 public speeches, an overwhelming majority of which (110) were held in Berlin. Figure B2 in the Appendix maps the locations of Goebbels’ appearances. Previous appearances. In addition to the matching variables listed, we include a binary variable that in￾dicates whether Hitler previously (i.e., before the last election) visited a county to help control for unob￾served confounders, assuming that those factors had already affected past targeting decisions. Election￾specific summary statistics of all the variables and sup￾plementary maps are given in Appendix H. PREDICTING HITLER’S APPEARANCES In the first stage of our empirical analysis, we model the election-specific probability of a Hitler visit to a community or county as a function of the above covari￾ates. Predictions from these models will then be used in the causal inference step to trim the sample to in￾clude as controls only those counties and communities 12 The list of keywords includes (root) words related to speeches and rallies: sprech, gesprochen, rede, kundgebung, ansprache, veranstal￾tung, vortrag. 1055 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Peter Selb and Simon Munzert TABLE 1. Probit estimates of Hitler appearances by election.Standard errors in parentheses. Sep 1930 Sep 1930(mun.) Apr1932(P) Jul 1932 Nov 1932 Mar 1933 Competitiveness 1 0.311 0.737* -0.653* -0.294 1.251* (0.291) (0.153) (0.386) (0.424) (0.748) Competitiveness 2 0.668* 0.770* 0.333* -0.256 0.094 (0.226) (0.125) (0.179) (0.209) (0.310) Organizational strength -0.704 -0.616* -0.166 -0.106 -0.257* -0.240 (0.513) (0.239) (0.260) (0.190) (0.155) (0.173) Distance to nearest airfield -0.323 0.050 -1.943* -0.334* -0.367 -3.574** (0.282) (0.164) (0.612) (0.190) (0.235) (0.780) Number of eligibles 0.695* 0.736* 0.196 1.200** 0.593* 0.5774 (0.164) (0.103) (0.216) (0.270) (0.214) (0.240) Previous NSDAP vote share 3.394 2.135* 0.579 0.989* 0.158 (2.146) (0.945) (0.847) (0.589) (1.088) Previous Hitler vote share -5.544* (1.823) Previous appearance 0.844** 0.886* 5.719 0.985* 0.967* 0.418 (0.185) (0.086) 240.510) (0.129) (0.129) (0.199) Goebbels appearance 1.077* 1.189* 7.513 0.708* 1.165* 1.618* (0.200) (0.112) (6.609.109) (0.207) (0.381) (0.293) (Intercept) -2.269* -2.715* -4.750 -1.161 -1.714** 2.337* (0.249) (0.140) (240.510) (0.328) (0.425) (0.750) Mc-Fadden's Pseudo R2 0.31 0.27 0.51 0.24 0.23 0.44 Observations 1.000 3.864 685 1.000 953 953 Log Likelihood 229.763 .712.633 63.317 398.459 252.176 118.828 Akaike Inf.Crit. 477.526 1.443.267 140.633 814.918 522.353 255.655 Note:"p<0.1;“p<0.05;**p<0.01 that are similar to exposure units in terms of their es- strength of local party organizations and the probabil- timated propensity score (while also being geographi- ity of a visit.An ad hoc interpretation of this finding cally distant enough from exposure units).The results would be that Hitler appearances were targeted at ar- 是 from the propensity score estimation are interesting in eas lagging behind in terms of organizational develop- their own right,too,since they provide rare system- ment to increase party membership(see Bytwerk 1981 atic insight into early Nazi campaign strategy.While 16).Unfortunately,the available samples from the NS- there are numerous studies on the organization of Nazi DAP member files are too small to detect local changes propaganda(e.g.Anheier 2003;Rosch 2002)and the in membership in the immediate aftermath of Hitler manipulative techniques employed(e.g.Anheier,Nei- appearances.What additional analyses show,however, dhardt,and Vortkamp 1998;Paul 1990),little is known is that local organizational development in the whole about the targeting of candidate appearances. legislative period following an election did not sys- Table 1 reports probit estimates and their standard tematically differ between exposure and nonexposure errors(clustered by primary electoral district)for each units.13.The significantly positive coefficient associated election separately.As one would expect,the size of the with Goebbels'appearances suggests that Goebbels' eligible voting population in a county turns out to be a campaign schedule tended to duplicate Hitler's.Thus,if consistent predictor of Hitler's campaign appearances we ignored Goebbels'activities,there would be a risk in across elections.Also in line with our expectations,the the subsequent analyses of erroneously ascribing cam- distance to the nearest airfield is a significant predictor paign effects to Hitler,whereas they actually trace back of Hitler visits as of the 1932 elections-Hitler's first to Goebbels.Finally,previous campaign appearances campaign trip by plane did not occur before April 3. prove to be useful predictors of current events,indicat- 1932(Bruppacher 2012,265).The exceptionally large ing that there are factors relevant to(past and current) coefficient on airfield distance referring to the 1933 targeting choices that are not appropriately taken into election is due to the extraordinarily intense regional account in our model specification.4 Nevertheless,the election in Lippe in January 1933 and its proximity to the airports of Bielefeld and Hannover-Vahrenwald (see Figure H2 in the Appendix).The campaign trail 13 See Table 139 in the Appendix. also tended to stop where the NSDAP did well in 14 The inflated SEs for"Previous appearances"and the intercept in the previous election-although the coefficients are,at the presidential election model are due to the fact that all 21 appear- most,weakly significant-and where the last election ances(32 units affected using a 10 km radius)took place in counties that had been exposed to earlier appearances.For this election,an was close from the party's viewpoint.The parameter earlier visit was-empirically speaking-a necessary but not suffi- estimates indicate a negative relationship between the cient condition for an appearance. 1056

Peter Selb and Simon Munzert TABLE 1. Probit estimates of Hitler appearances by election. Standard errors in parentheses. Sep 1930 Sep 1930 (mun.) Apr 1932 (P) Jul 1932 Nov 1932 Mar 1933 Competitiveness 1 0.311 0.737∗∗∗ − 0.653∗ − 0.294 1.251∗ (0.291) (0.153) (0.386) (0.424) (0.748) Competitiveness 2 0.668∗∗∗ 0.770∗∗∗ 0.333∗ − 0.256 0.094 (0.226) (0.125) (0.179) (0.209) (0.310) Organizational strength − 0.704 − 0.616∗∗∗ − 0.166 − 0.106 − 0.257∗ − 0.240 (0.513) (0.239) (0.260) (0.190) (0.155) (0.173) Distance to nearest airfield − 0.323 0.050 − 1.943∗∗∗ − 0.334∗ − 0.367 − 3.574∗∗∗ (0.282) (0.164) (0.612) (0.190) (0.235) (0.780) Number of eligibles 0.695∗∗∗ 0.736∗∗∗ 0.196 1.200∗∗∗ 0.593∗∗∗ 0.577∗∗ (0.164) (0.103) (0.216) (0.270) (0.214) (0.240) Previous NSDAP vote share 3.394 2.135∗∗ 0.579 0.989∗ 0.158 (2.146) (0.945) (0.847) (0.589) (1.088) Previous Hitler vote share − 5.544∗∗∗ (1.823) Previous appearance 0.844∗∗∗ 0.886∗∗∗ 5.719 0.985∗∗∗ 0.967∗∗∗ 0.418∗∗ (0.185) (0.086) (240.510) (0.129) (0.129) (0.199) Goebbels appearance 1.077∗∗∗ 1.189∗∗∗ 7.513 0.708∗∗∗ 1.165∗∗∗ 1.618∗∗∗ (0.200) (0.112) (6,609.109) (0.207) (0.381) (0.293) (Intercept) − 2.269∗∗∗ − 2.715∗∗∗ − 4.750 − 1.161∗∗∗ − 1.714∗∗∗ − 2.337∗∗∗ (0.249) (0.140) (240.510) (0.328) (0.425) (0.750) Mc-Fadden’s Pseudo R2 0.31 0.27 0.51 0.24 0.23 0.44 Observations 1,000 3,864 685 1,000 953 953 Log Likelihood − 229.763 − 712.633 − 63.317 − 398.459 − 252.176 − 118.828 Akaike Inf. Crit. 477.526 1,443.267 140.633 814.918 522.353 255.655 Note: *p<0.1; **p<0.05; ∗∗∗p<0.01 that are similar to exposure units in terms of their es￾timated propensity score (while also being geographi￾cally distant enough from exposure units). The results from the propensity score estimation are interesting in their own right, too, since they provide rare system￾atic insight into early Nazi campaign strategy. While there are numerous studies on the organization of Nazi propaganda (e.g. Anheier 2003; Rösch 2002) and the manipulative techniques employed (e.g. Anheier, Nei￾dhardt, and Vortkamp 1998; Paul 1990), little is known about the targeting of candidate appearances. Table 1 reports probit estimates and their standard errors (clustered by primary electoral district) for each election separately.As one would expect, the size of the eligible voting population in a county turns out to be a consistent predictor of Hitler’s campaign appearances across elections. Also in line with our expectations, the distance to the nearest airfield is a significant predictor of Hitler visits as of the 1932 elections—Hitler’s first campaign trip by plane did not occur before April 3, 1932 (Bruppacher 2012, 265). The exceptionally large coefficient on airfield distance referring to the 1933 election is due to the extraordinarily intense regional election in Lippe in January 1933 and its proximity to the airports of Bielefeld and Hannover-Vahrenwald (see Figure H2 in the Appendix). The campaign trail also tended to stop where the NSDAP did well in the previous election—although the coefficients are, at most, weakly significant—and where the last election was close from the party’s viewpoint. The parameter estimates indicate a negative relationship between the strength of local party organizations and the probabil￾ity of a visit. An ad hoc interpretation of this finding would be that Hitler appearances were targeted at ar￾eas lagging behind in terms of organizational develop￾ment to increase party membership (see Bytwerk 1981, 16). Unfortunately, the available samples from the NS￾DAP member files are too small to detect local changes in membership in the immediate aftermath of Hitler appearances. What additional analyses show, however, is that local organizational development in the whole legislative period following an election did not sys￾tematically differ between exposure and nonexposure units.13. The significantly positive coefficient associated with Goebbels’ appearances suggests that Goebbels’ campaign schedule tended to duplicate Hitler’s.Thus,if we ignored Goebbels’ activities, there would be a risk in the subsequent analyses of erroneously ascribing cam￾paign effects to Hitler, whereas they actually trace back to Goebbels. Finally, previous campaign appearances prove to be useful predictors of current events, indicat￾ing that there are factors relevant to (past and current) targeting choices that are not appropriately taken into account in our model specification.14 Nevertheless, the 13 See Table I39 in the Appendix. 14 The inflated SEs for “Previous appearances” and the intercept in the presidential election model are due to the fact that all 21 appear￾ances (32 units affected using a 10 km radius) took place in counties that had been exposed to earlier appearances. For this election, an earlier visit was—empirically speaking—a necessary but not suffi￾cient condition for an appearance. 1056 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Hitler's Speeches and the Rise of the Nazi Party FIGURE 2.Illustration of radius definitions of exposure areas(dark grey;event locations in white), no-matching areas(light grey),and potential control areas(white;potential control units in black). Triangles represent centers of county boroughs.Dots represent municipal centers. N.T Freiberg Sachsen N.609 0 ▲Glauchau Chemnitz ⊙· © Werdau 4号 。 Annaberg-Buchholz ④ 4905 N+F'OS 12-E 12.5-E 13-E 13.5-E models fit the data remarkably well.Pseudo R2 values guard our analysis against potential spillover effects. range between 0.23 in November 1932 and 0.51 in April The definitions of exposure,no-matching,and poten- 1932.That is,Hitler's campaign schedule appears to be tial control areas are illustrated in Figure 2.The match- pretty much in accord with an instrumental account of ing results are illustrated in Figure 3.On one hand,the campaign resource allocation.Further,there does not method throws away a nontrivial amount of informa seem to be much variation in campaign strategy over tion from both the pool of potential controls and the time. exposure group,which may result in loss of statistical efficiency.On the other hand,expanding the caliper and/or oversampling controls for efficiency gains may MATCHING AND COVARIATE BALANCE increase bias in matching estimators due to the inclu- sion of poorer matches.15 There are a variety of matching methods available.The Tables 2 and 3 report mean differences in vari- decision of which to choose from among them involves ables between exposure and control units before balancing the tradeoffs inherent in each between vari- and after matching plus a statistic that indicates the ance and bias(Caliendo and Kopeinig 2008).Our pre- ferred method is 1:1 nearest-neighbor matching with- out replacement subject to a caliper constraint of 0.25 15 Figures 12.13.and 14 in the Appendix report results for 1:5,1:10. standard deviations on the estimated propensity score. and 1:20 nearest-neighbor matching.As expected,these matching As an additional restriction,we draw no-matching ar- estimators are(slightly)more efficient than those based on our pre- ferred method,yet most of the estimates do not differ substantively eas of ten kilometers around the exposure units and If they do,they tend to be smaller in magnitude,which we interpret only match control units outside these areas to safe- as increased bias due to the inclusion of poorer matches. 1057

Hitler’s Speeches and the Rise of the Nazi Party FIGURE 2. Illustration of radius definitions of exposure areas (dark grey; event locations in white), no-matching areas (light grey), and potential control areas (white; potential control units in black). Triangles represent centers of county boroughs. Dots represent municipal centers. models fit the data remarkably well. Pseudo R2 values range between 0.23 in November 1932 and 0.51 in April 1932. That is, Hitler’s campaign schedule appears to be pretty much in accord with an instrumental account of campaign resource allocation. Further, there does not seem to be much variation in campaign strategy over time. MATCHING AND COVARIATE BALANCE There are a variety of matching methods available. The decision of which to choose from among them involves balancing the tradeoffs inherent in each between vari￾ance and bias (Caliendo and Kopeinig 2008). Our pre￾ferred method is 1:1 nearest-neighbor matching with￾out replacement subject to a caliper constraint of 0.25 standard deviations on the estimated propensity score. As an additional restriction, we draw no-matching ar￾eas of ten kilometers around the exposure units and only match control units outside these areas to safe￾guard our analysis against potential spillover effects. The definitions of exposure, no-matching, and poten￾tial control areas are illustrated in Figure 2. The match￾ing results are illustrated in Figure 3. On one hand, the method throws away a nontrivial amount of informa￾tion from both the pool of potential controls and the exposure group, which may result in loss of statistical efficiency. On the other hand, expanding the caliper and/or oversampling controls for efficiency gains may increase bias in matching estimators due to the inclu￾sion of poorer matches.15 Tables 2 and 3 report mean differences in vari￾ables between exposure and control units before and after matching plus a statistic that indicates the 15 Figures I2, I3, and I4 in the Appendix report results for 1:5, 1:10, and 1:20 nearest-neighbor matching. As expected, these matching estimators are (slightly) more efficient than those based on our pre￾ferred method, yet most of the estimates do not differ substantively. If they do, they tend to be smaller in magnitude, which we interpret as increased bias due to the inclusion of poorer matches. 1057 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Peter Selb and Simon Munzert FIGURE 3.Predicted propensity scores by exposure and matching status.Lines indicate matched pairs. Sep 1930 Sep 1930(mun.) Apr1932(P) Unmatched.Matched Matched.Unmatched,Unmatched,Matched. Matched. Unmatched.Unmatched.Matched. Matched. Unmatched trealed trealed control control treated trealed contol control treated treated contro control (n=29) =75 (n=75) =7350 n=54) (n=216份 (0=216m=2914) n=6 (0=26 n=26粉 (n=591) Jul 1932 Nov 1932 Mar 1933 上一 Unmatched.Matched. Matched, Unmatched,Unmatched, Matched. Matched, Unmatched,Unmatched, Matched. Matched, Unmatched treated treated control control treated control control treated treated control control h=55) n=164 (=16) (n=468) (a=170 m=87 (n=87) (=686) (n=20) (n=36 n=36) (=793) TABLE 2. Propensity score and covariate balance before and after matching.Mean differences on variables reported. Sep 1930 Sep 1930(mun.) Apr1932(P) Variable names Before After Impr. Before After Impr. Before After Impr Propensity score 0.32 0.01 97 0.24 0.01 97 0.35 0.00 100 Competitiveness 1 0.05 0.05 -8 0.10 0.05 48 Competitiveness 2 0.16 0.09 44 0.16 0.03 Organizational strength 0.18 -0.02 90 0.09 0.01 92 0.40 0.14 6 Distance to nearest airfield -0.22 -0.09 9 -0.14 0.03 -0.28 -0.03 0 Number of eligibles 0.78 -0.06 9 0.33 -0.01 98 0.85 0.26 69 Previous NSDAP vote share 0.01 0.00 78 0.01 0.00 -0.06 0.02 Previous appearance 0.27 -0.05 0.36 0.04 0.79 0.04 95 Goebbels appearance 0.41 0.03 93 0.34 -0.01 97 0.03 0.00 100 relative improvement of covariate balance through balanced before matching.In these cases,slight distri- matching.Matching markedly improved covariate bal- butional changes had massive consequences for the- ance to the extent that there are barely any mean differ- relative-improvement statistic.Overall,the balancing ences left.The seemingly curious instances of covari- statistics suggest that we can approach the causal anal- ate balance deteriorating after matching occur in situ- ysis step with a certain measure of confidence,at least ations in which the distribution of variables was well with regards to the observed potential confounders. 1058

Peter Selb and Simon Munzert FIGURE 3. Predicted propensity scores by exposure and matching status. Lines indicate matched pairs. TABLE 2. Propensity score and covariate balance before and after matching. Mean differences on variables reported. Sep 1930 Sep 1930 (mun.) Apr 1932 (P) Variable names Before After % Impr. Before After % Impr. Before After % Impr. Propensity score 0.32 0.01 97 0.24 0.01 97 0.35 0.00 100 Competitiveness 1 0.05 0.05 –8 0.10 0.05 48 Competitiveness 2 0.16 0.09 44 0.16 0.03 84 Organizational strength 0.18 –0.02 90 0.09 0.01 92 0.40 0.14 66 Distance to nearest airfield –0.22 –0.09 59 –0.14 –0.03 77 –0.28 –0.03 90 Number of eligibles 0.78 –0.06 93 0.33 –0.01 98 0.85 0.26 69 Previous NSDAP vote share 0.01 0.00 78 0.01 0.00 75 –0.06 0.02 75 Previous appearance 0.27 –0.05 80 0.36 0.04 89 0.79 0.04 95 Goebbels appearance 0.41 0.03 93 0.34 –0.01 97 0.03 0.00 100 relative improvement of covariate balance through matching. Matching markedly improved covariate bal￾ance to the extent that there are barely any mean differ￾ences left. The seemingly curious instances of covari￾ate balance deteriorating after matching occur in situ￾ations in which the distribution of variables was well balanced before matching. In these cases, slight distri￾butional changes had massive consequences for the— relative—improvement statistic. Overall, the balancing statistics suggest that we can approach the causal anal￾ysis step with a certain measure of confidence, at least with regards to the observed potential confounders. 1058 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424

Hitler's Speeches and the Rise of the Nazi Party TABLE 3.Propensity score and covariate balance before and after matching,continued.Mean differences on variables reported. July 1932 November 1932 Mar 1933 Variable names Before After Impr. Before After Impr. Before After Impr. Propensity score 0.28 0.01 96 0.23 0.00 99 0.37 0.01 98 Competitiveness 1 -0.02 0.00 94 0.00 0.03 -5479 -0.01 0.00 42 Competitiveness 2 0.03 -0.02 21 0.00 0.01 -248 0.02 -0.06 -234 Organizational strength 0.54 0.04 93 0.60 0.04 93 0.93 -0.11 8 Distance to nearest airfield -0.16 -0.05 70 -0.19 0.01 6 0.32 -0.02 93 Number of eligibles 0.55 0.06 90 0.66 0.01 99 1.09 0.14 7 Previous NSDAP vote share 0.00 0.00 0.00 0.01 -135 -0.03 0.01 Previous appearance 0.36 -0.05 85 0.51 -0.01 98 0.49 0.14 72 Goebbels appearance 0.20 0.04 82 0.13 0.01 91 0.36 0.03 92 FIGURE 4. Difference-in-differences estimates of exposure effects on NSDAP(Hitler)vote shares, KPD(Thalmann)vote shares,and turnout in national parliamentary and presidential elections 1930-33.Lines represent 80%and 95%confidence bands.Estimates are reported for unmatched and matched samples.For full model statistics,see Tables l1 to 13 in the Appendix. 4号元 NSDAP/Hitler vote share KPD/Thalmann vote share Turnout Sep 1930 Sep 1930 (mun. Apr 1932 (P) Jul 1932 Nov 1932 fult somnte Mar 1933 matched sample 0.06-0.03 0 0.030.06 0.06 -0.03 0.03 0.06 0.06 -0.03 0.03 0.06 ESTIMATING CAMPAIGN EFFECTS: 1933 Reichstag elections.16 Figure 4 reports difference- EMPIRICAL RESULTS in-differences estimates for the five election pairs and their 80%and 95%confidence intervals.both for the The semi-parametric difference-in-differences esti- matched and unmatched samples.The calculation of mate of the effect of Hitler appearances is the mean confidence intervals is based on robust standard er- difference in the overtime changes of outcomes be- rors to account for the clustering of temporal observa- tween the matched exposure and control units,with tions(pre-and post-exposure)within areal units.Re- each unit weighted according to the size of its electoral population.Difference-in-differences estimation criti- sults for three outcomes are reported:NSDAP(Hitler) cally rests on the assumption that observed overtime vote shares,KPD(Thalmann)vote shares,and turnout. Most point estimates are in the range of +1%of the changes in the control group reflect,on average,unob- voting-eligible population,and hardly any coefficient is served changes in the exposure group in the absence significantly different from zero at conventional prob- of treatment.A common plausibility check of such parallel trends is to compare pretreatment changes ability levels.The strongest effect of-2.4%pertains to over time between exposure and control units.Fig- ure I1 in the Appendix does not indicate any differ- 16 Analogous tests are not available for the 1930 Reichstag and the 1932 presidential election since the NSDAP (Hitler)did not run in ences in pre-treatment developments for the 1932 and the reference elections of 1924 and 1925. 1059

Hitler’s Speeches and the Rise of the Nazi Party TABLE 3. Propensity score and covariate balance before and after matching, continued. Mean differences on variables reported. July 1932 November 1932 Mar 1933 Variable names Before After % Impr. Before After % Impr. Before After % Impr. Propensity score 0.28 0.01 96 0.23 0.00 99 0.37 0.01 98 Competitiveness 1 –0.02 0.00 94 0.00 0.03 –5479 –0.01 0.00 42 Competitiveness 2 0.03 –0.02 21 0.00 0.01 –248 0.02 –0.06 –234 Organizational strength 0.54 0.04 93 0.60 0.04 93 0.93 –0.11 88 Distance to nearest airfield –0.16 –0.05 70 –0.19 –0.01 96 –0.32 –0.02 93 Number of eligibles 0.55 0.06 90 0.66 0.01 99 1.09 –0.14 87 Previous NSDAP vote share 0.00 0.00 92 0.00 0.01 –135 –0.03 0.01 76 Previous appearance 0.36 –0.05 85 0.51 –0.01 98 0.49 0.14 72 Goebbels appearance 0.20 0.04 82 0.13 0.01 91 0.36 0.03 92 FIGURE 4. Difference-in-differences estimates of exposure effects on NSDAP (Hitler) vote shares, KPD (Thälmann) vote shares, and turnout in national parliamentary and presidential elections 1930–33. Lines represent 80% and 95% confidence bands. Estimates are reported for unmatched and matched samples. For full model statistics, see Tables I1 to I3 in the Appendix. ESTIMATING CAMPAIGN EFFECTS: EMPIRICAL RESULTS The semi-parametric difference-in-differences esti￾mate of the effect of Hitler appearances is the mean difference in the overtime changes of outcomes be￾tween the matched exposure and control units, with each unit weighted according to the size of its electoral population. Difference-in-differences estimation criti￾cally rests on the assumption that observed overtime changes in the control group reflect, on average, unob￾served changes in the exposure group in the absence of treatment. A common plausibility check of such parallel trends is to compare pretreatment changes over time between exposure and control units. Fig￾ure I1 in the Appendix does not indicate any differ￾ences in pre-treatment developments for the 1932 and 1933 Reichstag elections.16 Figure 4 reports difference￾in-differences estimates for the five election pairs and their 80% and 95% confidence intervals, both for the matched and unmatched samples. The calculation of confidence intervals is based on robust standard er￾rors to account for the clustering of temporal observa￾tions (pre- and post-exposure) within areal units. Re￾sults for three outcomes are reported: NSDAP (Hitler) vote shares, KPD (Thälmann) vote shares, and turnout. Most point estimates are in the range of ±1% of the voting-eligible population, and hardly any coefficient is significantly different from zero at conventional prob￾ability levels. The strongest effect of −2.4% pertains to 16 Analogous tests are not available for the 1930 Reichstag and the 1932 presidential election since the NSDAP (Hitler) did not run in the reference elections of 1924 and 1925. 1059 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000424