Media, Markets, and Radical Ideas: Evidence from the Protestant ...

Media, Markets, and Radical Ideas:Evidence from the Protestant Reformation

Jeremiah Dittmar and Skipper Seabold∗

February 22, 2016

Abstract

This research studies the role of economic competition in the diffusion of ideasthat challenged an ideological monopoly and powerful elites during the ProtestantReformation. Martin Luther circulated his initial arguments for reform in 1517.We assemble data on all known books and pamphlets printed in German-speakingEurope between 1454 and 1600 and provide a new measure of religious ideas inthe media. We document a dramatic shift towards Protestant ideas after 1517 incities with competitive media markets, but not in cities with media monopolies.We find that competition in media markets mattered most for the diffusion ofProtestant ideas where formal political freedom was more restricted. We studythe relationship between competition and diffusion directly and using the deathsof printers to isolate plausibly exogenous variation in competition. The diffusionof Protestant ideas in the media preceded and predicts the institutionalization ofthe Reformation in municipal law.

Key words: competition, firms, media, technology, institutions, religion, politics, high-dimensional data

∗Dittmar: Department of Economics and Centre for Economic Performance, London School of Eco-nomics. Email: [email protected]. Seabold: Civis Analytics. Email: [email protected] thank Russ Gasdia, Kevin McGee, David Rinnert, Ava Houshmand, David Schlutz, and LuisMolestina-Vivar for research assistance, discussion, and suggestions. We thank colleagues at LSE, UCL,Stanford, Barcelona GSE, Chicago Booth, UC San Diego, Sussex, Michigan, Munich, Berkeley, and theNYC Media Seminar. Noam Yuchtman, Suresh Naidu, Davide Cantoni, and Ralf Meisenzahl providedvaluable feedback. Dittmar acknowledges support through the Deutsche Bank Membership at the In-stitute for Advanced Study (Princeton, NJ), the Centre for Economic Performance, and the EuropeanResearch Council.

1

1 Introduction

The Protestant Reformation delivered one of the most significant and surprising trans-

formations in European history. The pre-Reformation Catholic Church was an extraor-

dinarily powerful institutional and ideological monopoly. The Protestant Reformation of

the 1500s transformed European society by introducing religious competion. During the

Reformation, “religious and moral competition. . . often took place in compact geograph-

ical areas, even within individual cities. . . [and] competition led to a greater variety in

proferred salvation” (Roeck 1999; p 279-280). This radical shift in the market for ideas

occurred in a world where political freedom, inclusion, and voice were highly restricted.

The Reformation spread against the wishes of the Church, but also against the wishes

local elites and city councils (Sehling et al. 2009; Cameron 1991; Dickens 1979). How did

this happen?

In this research we study the role of economic competition in driving the diffusion of

Protestant ideas that transformed the religious and institutional landscape in Europe.

We focus on the implications of economic competition in local media markets, which

had emerged following Gutenberg’s printing press breakthrough in the mid-1400s. We

document that for the diffusion of Protestant ideas the key variation was between more

and less competitive media markets, not between cities with and without printing.

The Reformation provides a canonical and profoundly consequential example of the

role of market competition in shaping political and institutional change – in a setting

with an entrenched elite administering an ideological monopoly. Prior research on the

economics of the media has focused on settings where political competition is already

established and supported by legal institutions (Gentzkow et al. 2011; 2014; DellaVigna

and Kaplan 2007) and on the effects of propaganda by incumbent elites in non-democratic

settings (Adena et al. 2015; Yanagizawa-Drott 2014). Prior research in political economy

emphasizes the importance of politically inclusive institutions as fundamental supports

for economic openness and inclusion (Acemoglu and Robinson 2012). During the Refor-

mation, economic competition in the media shaped the diffusion of revolutionary ideas

that threatened an incumbent elite and opened new spaces for political and religious

participation.

We study the diffusion of the ideas of the Protestant Reformation in data on all

known books and pamphlets from German-speaking Europe between 1454 and 1600. We

construct a new measure of ideas in the media using estimators for high dimensional data

to analyse text data from historic publications. Our measure provides the first systematic

quantitative evidence on the diffusion of Protestant ideas in the media, to the best of

2

our knowledge. We apply the measure to study the role of technology, competition, and

institutions in the diffusion of the Reformation.1 To study the role of competition, we

construct new data on industrial organization at the city level. We focus on variation in

competition in media markets on the eve of the Reformation.2 We then examine how,

after Martin Luther circulated his theses criticizing Church corruption in 1517, Protestant

ideas hit the market and diffused across cities with different institutions and more or less

competitive media markets.

The evidence we find points to the fundamental importance of economic competi-

tion in the diffusion of the Protestant Reformation. We find that the most significant

variation in the diffusion of Protestant ideas occurs when we compare cities with two or

three firms to cities with monopolies. Cities with incumbent monopolies produced no

more Protestant content than cities in which no firms were initially active, some of which

experienced entrance during the Reformation. We also find that the relationship between

competition and diffusion was strongest in cities where where political freedom was more

restricted, by comparing cities with greater and lesser legal freedom under the constitu-

tion of the Holy Roman Empire. This evidence suggests that competition in the media

was especially important where the barriers to the dissemination of radical ideas were

relatively high. We show that initial industrial structure predicts the diffusion of both

Protestant and Catholic content during the Reformation era, but that the competition

effects are significantly stronger for the diffusion of Protestant ideas.

Comparisons between cities with more or less competitive media markets provide

evidence on how competitive conduct drove local responses to a common supply shock –

Luther’s intervention and the sudden appearance of new ideas in 1517. These comparisons

allow us to focus on the implications of differences in industrial structure across cities

with media industries that were otherwise similar in the size and composition of pre-

Reformation output. But they also naturally make us wonder whether cities with more

or less competitive media markets differed on unobserved dimensions.

Motivated by questions concerning the identification of cause and effect, we examine

the timing of printer deaths as a source of exogenous variation in competition. We con-

struct comprehensive data on the deaths of printers who were firm owner-managers in the

decade just before the Reformation. We study variation in city-level industrial structure

1Previous research on the Reformation has studied measures of religious belief observed in the 1800s(Becker and Woessmann 2009) or the binary Protestant-or-Catholic denomination of territories or citiesin the mid-1500s (Cantoni 2014; Rubin 2014).

2This focus enables us to study variation in competition that does not reflect Reformation-specificincentives, for example endogenous entrance decisions by printers and/or citizens’ and city council actionsdesigned to favor particular content. This imposes limits on the scope of our study. We currently deferan examination of selective entry that may have been induced by the Reformation for future work.

3

induced by the deaths of active printers and find support for a causal interpretation of

the relationship between competition and diffusion. In contrast, the deaths of retired

printers induce no response in industrial structure or output. The effects of these shocks

reflect competitive environments in media markets which were highly concentrated and

in which cartels, monopolies, and strategic behavior were endemic. In this setting, the

deaths of owner-managers disturbed restraints on trade, induced significant increases in

entrance, and increased competition.

Our focus on the role of competition cuts across the classic explanations for the Ref-

ormation, which emphasize technology and political freedom (Brady 2009). Historical

research has identified printing technology as providing a critical platform the early re-

formers used to disseminate ideas and mobilize citizens (Cameron 1991; Eisenstein 1980).

Research in economics has revisited the hypothesis the Reformation was “the child of the

printing press” and finds that cities with at least one printing press in 1500 were sig-

nificantly more likely to adopt Protestantism than cities that did not have printing in

1500 (Rubin 2014). Historians also observe that constitutionally designated “free cities”

outside the jurisdiction of authoritarian lords provided settings in which the Reforma-

tion developed and set down roots (Moeller 1972).3 Unlike prior research, we study the

diffusion of ideas in the media and show that for the diffusion of Protestant ideas the

key variation was between more and less competitive media markets – not between cities

with and without printing technology.4 We find that the relationship between competi-

tion and diffusion was stronger cities that were not free. We then document how ideas

in the media predict formal institutional change.

A big picture take-away from our research is that fundamental transformations in

European markets for ideas were driven by the use of new information technology by

firms competing in markets (Febvre and Martin 1958). Put another way, the Romerian

meta-idea behind the diffusion of the Reformation and the broader print media revolution

in Europe was the use of new information technology by for-profit firms.5

3This freedom was relative. Free cities were ruled by local oligarchies and subject to the emperor.4This finding contrasts with Rubin’s (2014) evidence suggesting that that cities with printing presses

were more likely to become Protestant than cities without printing. However, our focus is on a continuousmeasure of the diffusion of ideas, whereas Rubin examines a broad binary definition of whether a citybecame Protestant and does not examine systematic evidence on media content. Also as Rubin carefullyexplains, the design in Rubin (2014) cannot definitively disentangle supply and demand side determinantsof religious change to rule out alternate ways the correlation between having printing and adoptingProtestantism could emerge. In related research, Cantoni (2012) focuses on the adoption of Protestantismby territorial princes and does not study systematic or high frequency evidence on ideas in the media.

5Romer defines a meta-idea as an idea about how to support the production and diffusion of otherideas. To consider how the use of printing technology by firms shaped European economies, a high-level contrast is Korea. A metal type printing technology with “an almost hallucinatory similarity toGutenberg’s” (Briggs and Burke 2010; p. 13) was developed in Korea well before similar innovations inEurope. Printing using this technology in Korea was principally an activity of the state and not profit-

4

We focus on the period from 1518 to 1554 to respond to the core debate in the

historical literature on the role of printing in the diffusion of the Reformation. Starting

in 1518, print media was used to mobilize local, popular movements for reform. In

1555, a new institutional and geographic equilibrium for religion in central Europe was

settled in law by the Peace of Augsburg, changing the incentives in media markets. Most

historians argue that over this period printing was of overwhelming importance (Hamm

1994; Ozment 1980; Brady 2009) and that the publication decisions of individual printers

were pivotal shifters in the diffusion process (Chrisman 1982). But some historians

suggest that printers publishing Protestant material may have been simply responding

to demand (Whaley 2011) and thus raise fundamental questions about cause and effect.6

The debate over the role of the media in the Reformation poses three challenges

that we address in this research. First, existing data on print media are large, high-

dimensional, and do not categorize books or authors by religion. As a result, no research

systematically documents the diffusion of Protestant ideas in the media quantitatively.7

Second, evidence on firms is needed to study how the economics of media markets explain

diffusion.8 Third, we need to isolate sources of plausibly exogenous variation to untangle

cause and effect. We address these challenges as follows.

First, we construct a new measure of religious ideas in the media. We assemble

data on all known books and pamphlets printed in German-speaking Europe and use

estimators for high-dimensional data to classify content. The data consist of over 100,000

publications printed in more than 100 cities between 1454 and 1600. We use historical

sources to identify over 450 leading Protestant and Catholic authors – authors of 18% of

the publications printed in Germany. We then use statistical models for high-dimensional

data to identify the language characteristic of Protestant and Catholic authors in long,

seeking entrepreneurs. In Europe, printing was one of the first industries in which the firm emerged asthe key organizational form (Brady 2009) and was from the outset for-profit (Dittmar 2011).

6A general consensus among historians holds that print media were important despite the fact thatliteracy levels were low. Print media were read aloud and the ideas transmitted in print were furthercirculated in sermons and conversations. In addition, pamphlets often featured illustrations that wereintelligible to non-readers. The narrative evidence also points very strongly to the conclusion that printmedia radically shifted views and was not simply a response to prior variations in demand. See forexample, Brady (2009), Scribner (1994), and Edwards (1994).

7For example, the Universal Short Title Catalogue (of St. Andrews 2012) which we discuss belowsimply classifies book subject matter as “religious” or not. The same is true of Das Verzeichnis derim deutschen Sprachbereich erschienenen Drucke des 16. Jahrhunderts. The history literature has todate examined small subsets of the data. Edwards (1994) is arguably the largest and studies a sampleof pamphlets from the 1500s that comprises 3,183 authored by Martin Luther and 1,763 authored byCatholic activists. But a large share of output is by effectively unknown authors, as discussed below.

8Previous research has not constructed systematic evidence on firms. The exception is Dittmar(2015), which develops related ideas on historic media markets, but relies on fundamentally differentand non-comprehensive evidence on firm-level shocks, studies locations outside Germany, and does notfocus on the Reformation or religious content, as discussed below.

5

historical book titles.9 We use these models to generate a measure of the religious content

at the author and city levels that captures how similar the language used in the media

is to the language of known Protestant and Catholic writers.10

Second, we construct data on firms in order to study the relationship between indus-

trial organization and the diffusion of ideas in the media. We construct data on firms

from comprehensive biographical dictionaries of historic printers.11

Third, we both examine how initial competition shaped local responses to the common

supply shock – the sudden appearance of Protestant ideas – and construct data on the

timing of the deaths of printers as a source of exogenous variation in local competition.

We identify the timing of the death of every individual printer active in German-speaking

Europe during the early 1500s. This individual-level evidence enables us to document the

presence and absence of manager deaths in all firms. No previous research has constructed

similarly comprehensive data on firm-level shocks, to the best of our knowledge.12

Our research also addresses larger questions about the origins of fundamental be-

liefs and institutions. An influential body of scholarship traces differences in contempo-

rary economic performance and behavior to historically determined institutions or beliefs

(Acemoglu et al. 2001; 2011; Voigtlander and Voth 2012; Guiso et al. 2003). This lit-

erature calls attention to the persistence of key institutions and beliefs, but raises the

question: What explains fundamental changes on these dimensions? Here we study the

role of competition in the media in transmitting ideas that led to large scale social change.

We specifically study environments in which the media were used to solve coordination

problems and mobilize citizens’ movements that challenged religious and secular elites –

somewhat in the spirit of Acemoglu and Robinson (2006) but without democracy on the

table.

9The median title in our data has 21 words and 153 characters. For comparison, twitter messagesare no longer than 140 characters. Estimation strategies similar to ours are widely used to classify thecontent of tweets, spam email, and other short texts. We discuss the data and estimation in detail below.

10Our estimation strategy builds on Gentzkow and Shapiro (2010), which measures the “slant” ofUS newspapers by determining whether they employ language similar to that used by Democratic orRepublican members of the US Congress. A large literature uses high-dimensional estimation techniquesto document political sentiment in news media and the diffusion of ideas in social media such as twitter“tweets” (Gentzkow and Shapiro 2010; Taddy 2013b;c; Pak and Paroubek 2010; Bollen et al. 2011). Wetake these estimators to the data from the Reformation.

11We also construct an alternate measure of firms based on the names of printers listed on the frontpages of historic books and pamphlets. Our results hold in both cases. We provide further discussionbelow and in the appendix.

12We construct comprehensive data on the timing of printers’ deaths by tracking every individualprinter active in the early 1500s. The data we study here differ fundamentally from the data in Dittmar(2015), which studies printer deaths as a source of variation in media markets across Europe, but onlygathers evidence on manager deaths in firms that do not exit following these shocks. We discuss ourdata in detail below and in the appendix.

6

2 History

In October 1517, Martin Luther circulated his famous hand-written theses criticizing

church corruption. Luther’s theses notably criticized the Catholic Church’s practice

of selling indulgences which were believed to secure the release of dead relatives from

purgatory. The fees charged for indulgences were classed as administrative charges to

skirt theological prohibitions on the sale of sacred things. But indulgence transactions

were jointly coordinated by the Church, local rulers, and bankers, acting in recognizable

business partnerships and sharing the revenues (Brady 2009; Whaley 2011). The Church

used revenues from indulgences to finance investments (e.g. the basilica of St. Peter in

Rome) and consumption.13 Elites used cash flows on future indulgences to secure loans

that financed the purchase high Church offices.14 Luther circulated his theses criticizing

these and other practices in letters to three correspondents. Within months they were

printed in multiple German-speaking cities.15 The dissemination of Luther’s ideas in the

media provoked a public controversy.

Two key features distinguish Luther’s intervention and the emergence of the Protes-

tant Reformation from previous attempts to challenge church institutions and practices.

First, when the Catholic Church attacked Luther and other protesting clergy, the re-

formers responded by developing and disseminating their ideas in print media. Second,

politically active laymen adopted and adapted reformist ideas and pressed them on gov-

erning elites at the local level (Cameron 1991). Luther and the other early Protestant

reformers were clerics and scholars who criticized and then challenged the ideological

and institutional monopoly of the Catholic Church. The reformers argued that biblical

authority was paramount over and above the authority of existing church institutions

(Brady 1978), called for moral renewal within cities (Moeller 1972), and were often anti-

clerical (Dykema and Oberman 1993). These ideas were developed in print by critical

churchmen and lay activists. The Protestant program that emerged involved the aboli-

tion of the Catholic rite mass, the establishment of safeguards against church corruption,

increased public goods provision in health and education, and moves to eliminate clerical

13In his 86th thesis, Luther asked: “Why does the pope, whose wealth today is greater than the wealthof the richest Crassus, build the basilica of Saint Peter with the money of poor believers rather thanwith his own money?” Crassus was arguably the wealthiest individual in Roman history and among therichest of all time. Crassus famously became wealthy via real estate speculation, the slave trade, andinsider political expropriations, and played a role in Rome’s transformation from Republic to Empire.

14Joachim I the Elector of Brandenburg borrowed money from the Fugger bankers to purchase theoffice of Archbishop of Mainz for his brother. The loan was guaranteed by future revenue on indulgences.

15This research focuses on the urban reformations because of the key role cities played as incubatorsof Reformation ideas and institutional innovations (Hamm 1994). However, the Reformation was notrestricted to cities. Religious ideas were central to the so-called peasants’ war of 1525 (Blickle 1981).

7

tax exemptions and to regulate or eliminate religious monopolies.16

Print media played a central role in the diffusion of Protestant ideas (Edwards 1994;

Brady 2009; Eisenstein 1980). Chrisman (1982; p. 29) observes that the influence of

printers derived from the fact that, “their decision to print or not to print a particular

book or tract could have an immediate effect on political and religious events and, in a

time of rapid change, on institutions. The most striking example of their influence can

be seen in the religious publication of the pivotal years of the Reformation.” The flow

of media challenging the Catholic monopoly shifted views and helped solve coordination

problems: “It was the superabundance, the cascade of titles, that created the impression

of an overwhelming tide, an unstoppable movement of opinion” (Pettegree 2005; p. 163).

Printing was a concentrated business in which religion became a dimension for product

differentiation and competition (Reske 2007; Pettegree 2000; Creasman 2012). Historians

observe that competition mattered for the local diffusion of religious ideas: “Where a

market was controlled, the free flow of innovative theological speculation was greatly in-

hibited” (Pettegree 2000; p. 114). The nature of competition varied across cities. Cartels

and strategic anti-competitive behavior were endemic (Reske 2007; Dittmar 2015). But

entry and markets were unregulated, as printing did a Schumpeterian end-run around

guild regulation. Fussel (2005; p. 59) observes that into the 16th century, the business

was, “free to develop without regulation by governments, princely houses or the Church,

nor is there any evidence that any restrictions were imposed by guilds.”17 Some large

cities had no printing, others had 1 firm or several. Prior to the Reformation, the mean

printing city had 3.1 firms and the median 1. Transport costs limited inter-city trade,

diffusion through reprinting was typical, and the key competitive dynamics were within

city (Edwards 1994; Dittmar 2015). A typical pamphlet cost 1/3 of one day’s wages.18

Protestant ideas spread in the media and then orally through a two-part process

(Edwards 1994). First, print media impacted clergy and educated lay “opinion leaders.”19

Opinion leaders then transmitted ideas orally to the broad public and developed popular

movements. The diffusion process, with ideas first disseminating in the media, followed a

“strikingly common pattern” in cities and in towns, in Northern and in Southern Germany

(Brady 2009; p. 161). In Zurich, Reformation activism dates to March 1522, when printer

16Besides exemption from taxes and civic duties, religious orders enjoyed monopolies on priced religiousservices (e.g. funeral services) and on the production of products like beer. In a pamphlet published in1522, Sebastian Meyer provided the following summary to his reader: “Dear layman, it is all done withone purpose in mind and that is your pocketbook. . . They exist by the founding and confirmation of thepope and they help him make off with your goods.” Translation in Ozment (1975; p. 58).

17See also Barbier (2006); Nicholas (2003); Hirsch (1974).18The appendices provide data on prices and estimates of the price-gradient on transport for books.19A number of significant lay reformers were city clerks. For example, Lazarus Spengler and Jorg

Vogeli were clerks of Nurnberg and Konstanz, respectively, and published reformist works in the 1520s.

8

Christoph Froschauer and his workers engaged in civil disobedience by breaking the Lent

fast with pastor Ulrich Zwingli as their witness. City authorities arrested the printers.

Zwingli preached in their defense, and his sermons were disseminated as pamphlets,

opening public debate. In Augsburg, “A wave of religious pamphlets and, from 1520,

the introduction of evangelical preaching, spread the new teaching” (Broadhead 1996; p.

581). In Northern cities, reading groups of the 1520s preceded activism and legal change

in the 1530s (Schilling 1983).

The city-level Reformations were popular movements that emerged without initial

support from oligarchic city governments or territorial lords.20 The constituency for re-

form came from citizens who were excluded from political power by oligarchic elites, typi-

cally lesser merchants and guild members (Ozment 1975; Schilling 1983). Thus Cameron

(1991; p. 240) observes, “As a rule neither the city patricians nor the local princes showed

any sympathy for the Reformation in the crucial period in the late 1520s and early 1530s;

they identified themselves with the old Church hierarchy and accordingly shared its un-

popularity. Popular agitation on a broad social base led to the formation of a ‘burgher

committee’.” This model characterized the Reformation even in its birth place: “It is

undeniable that the Wittenberg movement was borne on a wave of popular enthusiasm.

It outran the city magistrates’ ability to control it, and finally forced them to act even

against the will of the Elector [the territorial ruler], who had prohibited any innovations

in church matters” (Scribner 1979; p. 53).21 While local Reformations were popular

movements, princes did vary in their response to Protestant ideas.22 In the econometric

work below we show that the positive relationship between competition and the diffu-

sion of Protestant ideas in the media holds when we study variation in competition and

Protestant content across cities in the same principality, subject to the same lord.

The institutions of the Holy Roman Empire shaped the diffusion of the ideas of

the Reformation. Political decentralization limited the capacity of central and regional

20The evidence indicates, “no clear causal link between confiscating lands and turning Protestant”(Cameron 1991; p. 296). Several rulers extracted resources from the Church years before definingtheir religious position. The dukes of Bavaria used the threat of concessions to Protestants to extractresources, but remained Catholic. At the city level, expropriated assets were put to new uses, largelythe provision of public goods.

21For example, in Augsburg, the city council was forced to drop its policy of religious neutrality fol-lowing riots in 1524, 1530, and 1534 (Broadhead 1979). In Northern cities, such as Rostock, Stralsund,Griefswald, Lubeck, Braunschweig, Luneberg, Gottingen, and Hanover institutional change led by cit-izens excluded from political power had a coup d’etat quality (Cameron 1991). In Zwickau, Lutheranpublications were printed in 1523; the city council attempted to suppress protests in 1524; the Reforma-tion was formally adopted in 1525 (Scribner 1979). See also Dickens (1979).

22For example, the Dukes of Bavaria were particularly committed opponents of the Reformation. Incontrast, the Dukes of Pomerania were similarly opponents of Protestantism until 1531, when the deathof Duke George I led to a pro-Protestant shift in ruler preferences and the principality-level assembly(landtag) voted to formally allow Protestant preaching. Protestant ideas diffused in both principalities.

9

authorities to regulate media markets (Kapp 1886; Creasman 2012). The Holy Roman

Empire was composed of many semi-autonomous principalities and two main types of

cities. Free cities (Freie und Reichsstadte) were constitutionally outside the jurisdiction

of territorial lords and had city councils with extensive legal autonomy. Free cities have

been identified by historians as playing a leading role in the diffusion of the Reformation

(Moeller 1972; Cameron 1991). Territorial cities (landstadt) fell under the juridiction of

lords, but distance and state capacity put limits on the extent of lords’ control (Whaley

2011).23 In this context, suppression of dissent was costly and slow to be tried. Printers

producing Protestant media were not censored in the early years of the Reformation

and the Edict of Worms (1521) banning Luther’s work was not rushed into print (Brady

1985; p. 153). At the city-level, censorship rules were endogenous, “as much a product

of public opinion as a force acting upon it” (Creasman 2012; p. 227).24

The diffusion of Protestant ideas preceded formal institutional institutional change

at the city level. The institutions Protestant reformers set-up were new laws called

church ordinances (kirchenordnungen) that governed religious practice, established pub-

lic education, and expanded social welfare provision. These laws were passed first by

city magistrates and municipal councils and only later by princes at the territorial level

(Hamm 1994). Starting in the mid-1520s, city councils began to defy the emperor and

pass such laws institutionalizing the Reformation. In 1526, a formal right to reform (ius

reformandi) was passed into law by the Imperial Diet or parliament of the Holy Roman

Empire (Brady 2009; p. 55).

By the mid-1500s, Protestantism in historic Germany acquired the geographic distri-

bution it would maintain for several centuries (Brady 1998; p. 373). The early Reforma-

tion in which printing played a central role thus spanned the period from 1518 to 1554.

Starting in 1518, we observe the initial explosion of Protestant ideas in the media and the

emergence of popular pro-Reform movements.25 In the 1540s, Protestant and Catholic

forces went to war (the Schmalkaldic War). The Peace of Augsburg of 1555 formalized

a legal settlement governing the religious geography of the Holy Roman Empire and the

religious prerogatives of rulers – effectively fixing religious institutions.26

23For example, the feudal lord might appoint a district official (Amtmann) or a jurisdictional officer(Stadtrichter) who would reside in a territorial city located within his domains and assume responsibilityfor local policy and governance alongside a less fully empowered city council. See Dixon (1996). Lordscould be non-resident ecclesiastical (e.g. prince-bishops) or secular (dukes, counts, etc.) rulers.

24Cologne was the exception to this rule (Scribner 1976). See also Chrisman (1982).25In 1525, the peasants’ movement for religious and economic reform was crushed (Blickle 1981).26The Peace of Augsburg set the rule cuius regio, euis religio (whose rule, his religion) with exceptions

for cities where Protestants and Catholics were to share churches and magistracies (Brady 1998; p. 375).

10

3 Data

We restrict analysis to the set of 191 historically German-speaking cities identified as

printing centers by Reske (2007; 2015) and/or for which population data exist at 1500

in Bairoch et al. (1988).27 Figure 1 shows the cities we examine and the boundaries of

principalities in the Holy Roman Empire.

The key unit of analysis in this paper is the book or pamphlet edition, which can be

thought of as a variety produced in a given city-year.28 We study over 100,000 individual

book and pamphlet editions that were printed in German-speaking cities between 1454

and 1600. Our primary source for data on print media is the Universal Short Title

Catalogue (USTC) database of St. Andrews (2012), which is designed as a universal

catalogue of all known books and pamphlets printed in Europe between 1450 and 1600.

We construct data on firms from Reske (2007; 2015), Die Buchdrucker des 16 und

17 Jahrhunderts im deutschen Sprachgebiet, which is the authoritative biographical dic-

tionary of historical printers active in German-speaking Europe.29 Reske (2007; 2015)

records information on the lives, business operations, and dates of operation and death for

historic printers. As a robustness check, we also construct a measure of “firms from pub-

lications” by identifying the printer-publisher named in inscriptions on the front pages of

books. Our results on content diffusion are robust to using either measure, but stronger

using the Reske biographies measure due to measurement error in evidence constructed

from books. We discuss these differences and implications below and in the appendix.

We focus on industrial structure in the ten years before the Reformation (1508-1517).

We use information on the deaths of printers as a measure of shocks to local media

markets. We consider deaths occuring within one year of a printer’s last observed business

activity as the death of an active printer. Other deaths are inactive (retirees). Our data

on the timing of death are comprehensive and cover every firm and all printers active in

27These cities may be considered German-speaking in a qualified sense. Many were characterized bylinguistic diversity. Within German-speaking Europe High and Low German co-existed. The set of citiesin our data includes cities now in Austria, France (e.g. Metz and Strasbourg), Switzerland (e.g. Zurichand Basel), Poland (e.g. Gdansk, historically known as Danzig, and Szcezecin, historically known asStettin), the Czech Republic (e.g. Brno or Brunn), and Russia (Kaliningrad or Konigsberg). Bohemiancities are excluded as their vernacular printing was overwhelmingly in Czech. Our results are robust toalternate definitions of the relevant data, such as expanding the sample to include all German-speakinglocations that appear in Bairoch et al. (1988) even those without population recorded in 1500.

28For a subset of several hundred books we have data on the number of copies printed per edition.These data and evidence from book contracts indicate that the typical print run rose from 400-800 copiesaround 1500 to 1,000-1,400 copies in the later 1500s. See Dittmar (2015).

29Reske (2007; 2015) builds on the biographical catalog produced by Josef Benzing (1982). Benzingwas employed by the Prussian State Library (Preußischen Staatsbibliothek) in Berlin 1934-1945 and ashead librarian at the University of Mainz 1946-1966. Benzing’s work is described in the appendix.

11

Figure 1: German-speaking cities with and without printing pre-1517

#V #V

#V

#V

#V

#V

#V

#V

#V

#V

#V

#V#V

#V#V

#V

#V

#V

#V

#V

#V

#V

#V

#V

#V#V

#V#V#V

#V#V

#V

#V

#V

#V

#V

#V

#V

#V

#V

#V#V

#V

#V

#V

#V

#V

#V

#V #V

#V

#V#V

#V

#V

#V

#V#V

#V

#V

#V

#V

#V

#V#V

#V

#V

#V

#V

#V

#V

#V#V

#V

#V#V

#V

#V

#V

#V

#V#V

#V

#V

#V

#V

#V

#V

#V

#V

#V

#V

#V#V

#V

#V

#V

#V

#V

#V

#V

#V#V

#V

#V#V

#V

#V

#V

#V

#V#V

#V #V

#V

#V

#V

#V

#V

#V

#V#V

#V #V#V

#V

#V

#V#V

#V

#V#V

!(!(

!(

!(

!(

!(!(

!( !(

!(

!(

!(

!(

!(!(

!(

!(

!(

!(

!(

!(

!(

!(!(!(

!(

!(

!(

!(

!(

!(

!(

!(

!(

!(

!(

!(!(

!(!( !(

!(

!(

!(

!(

!(

!(

!(!(

!(

!( !(!(

!(

!(

!(

!(

!(

!(

!( Towns with Printing#V Towns without Printing

This map shows the cities of German-speaking Europe that we study. Circles (in black) designate citieswith any printing before 1517. Triangles (in white and black) are cities without printing before 1517.Principality boundaries as of 1500 designate territories of the Holy Roman Empire.

the ten years before the Reformation. From Reske (2007; 2015), we identify four types of

firms. First are firms that experience a death which leads to their exit. Second are firms

that experience a death but do not exit because a widow or heir becomes owner-manager.

Third are firms who survive without deaths. Fourth are the firms that exit between 1508

and 1517 without a death recorded in this period by Reske. For these 31 printers, we

gather historical data on the timing of their later deaths, inter-city migrations, and

business activities to document that they did not die in our key pre-Reformation period

and that selective data preservation in Reske (2007; 2015) is not driving our results. We

provide detailed evidence on each of these individual printers in the appendix.

We construct data on the religious affiliation of known Protestant and Catholic au-

thors from several sources. Klaiber (1978) provides data on 225 Catholic authors in

German-speaking Europe before 1600. We identify 234 Protestant authors from Mul-

lett (2010), Carey and Lienhard (2000), and Wikipedia’s list of Protestant Reformers.30

Known Catholic authors account for 2,929 titles (3% of books) and known Protestants

account for 15,507 (15% of books).31

30See www.wikipedia.org/wiki/List of Protestant Reformers (downloaded 12/15/2012).31The print media dominance of Protestant reformers has been observed by social historians and in

earlier research based on small samples of historical print media. For instance, Edwards (1994; p. 29)

12

Table 1 presents summary statistics on output in media markets and shows that

almost half of religious media was written by authors whose religion we do not observe.

Table 1: Print media output in German-speaking Europe 1454-1599

Time Cities Total Religious Religious by KnownPeriod Printing Publications Publications Protestants Catholics

1454-1517 59 16966 6635 64 2041518-1554 91 33218 17108 8883 15351555-1599 107 57940 23247 6560 1190

The time periods in this table are pre-Reformation (1454-1517), Reformation before the Peace of Augs-burg (1518-1554), and after the peace (1555-1599). Cities printing records the number of cities printingin the USTC. Total publications records the number of publications. Religious publications records thenumber of publications classified as “Religion” by the USTC. Religious publications by known Protes-tants and Catholics are those written by authors whose religion we identify from historical evidence.

We use the text of book and pamphlet titles to identify the language most characteris-

tic of Protestant and Catholic print media and classify content. Historical titles provided

extended descriptions of content. The median title in our data has 153 characters (mean

171.6). We provide discussion of the text data and our research strategy for classifying

and indexing religious ideas below.32

To identify the institutions of the Reformation we code municipal Reformation laws.

Our principal source is, Die evangelischen Kirchenordnungen des 16 Jahrhunderts (21

volumes 1902-2013).33 Additional sources are described in the appendix.

4 Measuring Religious Ideas in Media

4.1 Methodology

We index religious ideas in the media as follows. We first estimate how the distribution

of language changes with the religious affiliations of 459 Protestant and Catholic authors

we identify from historical sources.34 This allows us determine which features of lan-

guage are important in identifying religious beliefs. We then use the estimated partisan

ideology of language to predict the content of media where authors’ religious affiliations

may not be known. These predictions form our index. Methodologically we build on

finds in a sample of vernacular (German-language) pamphlets that the ratio of works by Martin Lutherto works by Catholic publicists was approximately 5 to 1.

32A large body of natural language processing research performs classification tasks using twittermessages that are no longer than 140 characters.

33We refer to these volumes collectively as Sehling et al. (2009).34We identify 234 Protestant authors and 225 Catholics as described above.

13

Gentzkow and Shapiro (2010) and perform classification using multinomial inverse re-

gression (MNIR) (Taddy 2013b).35 The distinction we maintain between Catholic and

Protestant provides a powerful first model for thinking about religious ideas in the media

during the Reformation. In part this is because later divisions between Lutheran and

Calvinist Protestants were only incipient in the first half of the 1500s.36

We are able to identify how language shifts with religion because historical titles are

long and provide extensive glosses on content. To understand the information in titles, we

first provide two examples of English-language books printed in 16th century Germany.37

An example of a Protestant title is a book written by Martin Luther and printed in Wesel:

The last wil and last Confeßion of martyn Luthers faith concerning the prin-

cipal articles of religion which are in controversy, which he wil defend &

maiteine until his death, agaynst the pope and the gates of hell.

An example of a Catholic title is a book written by John Old and printed in Emden:

A Confeßion of the most auncient and true christen catholike olde belefe

accordyng to the ordre of the .XII. Articles of our comon crede, set furthe in

Englishe to the glory of almightye God, and to the confirmacion of Christes

people in Christes catholike olde faith.

These examples both illustrate the information in historical titles and suggest potential

challenges in analyzing historical data in which spelling is not standardized. However,

the results below show that our estimation strategy accurately predicts the religion of

known religious authors in out-of-sample tests. We also are able to index content in the

majority of output which is by unknown authors, and to index content in cities in which

no publications by known religious authors appeared.

Formally, let a title be denoted X i = [xi1,...,iW ] where xiw represents the number of

times phrase w appears in title i. The W phrases are determined by the estimating vocab-

ulary V . To identify phrases that differentiate publications according to their religious

35Gentzkow and Shapiro (2010) develop a similar estimation strategy to identify the dimensions ofpolitical ideology in Congressional speeches and to measure the political “slant” of US newspapers bydocumenting the extent to which these media use language characteristic of Democrats or Republicans.Taddy (2013b) extends the Gentzkow and Shapiro (henceforth “GS”) approach, and Taddy (2013c)applies a similar strategy to measure political sentiment in twitter “tweets.”

36While a powerful first model, the distinction between Catholics and Protestants does not exhaustthe distinctions one could draw among authors’ religious views. Extensions could distinguish Lutheran,Calvinist, and Zwinglian ideas in Protestant media and different types of Catholic authors. We discusstime-varying features of religious language below.

37In our database, 96% of books and pamphlets are in German and Latin. We provide examples ofGerman titles with translations in the appendix.

14

content r with r ∈ P,C for Protestant and Catholic, we use an estimating vocabulary of

one-, two-, and three-word phrases chosen from all titles using the log-odds ratio with an

informative Dirichlet prior for each phrase’s frequency between denominations (Monroe

et al. 2008). We model the language in the titles as distributed according to a multinomial

distribution that differs across religions r such that

Xr ∼MN(qr,mr) (1)

where qr is the vector of parameterized phrase probabilities and mr =∑

i

∑w xiwr. Ap-

pendix B provides full details of the MNIR model and our estimation strategy, including

how we construct the vocabulary, how our approach compares to Gentzkow and Shapiro

(2010), the performance of the estimator, and questions concerning time-varying hetero-

geneity in language. When predicting religion out-of-sample, we collapse all titles for an

author to obtain a single oeuvre in order to avoid uncertain estimates for shorter titles.38

4.2 Classification Results

Our classifier shows high performance on in- and out-of-sample classification. In-sample,

the estimator correctly predicts the religion of an author 86% of the time.39 This success

rate compares favorably to results in similar prediction problems in the literature.40 To

show that the estimator is not overfitting, we present cross-validation exercises where we

estimate (train) over 80% of the data and predict (test) over the held-out 20%.

Figure 2 shows the performance of the classifier on the held-out test data from a single

draw and highlights how well we predict prominent authors Martin Luther and Johannes

Eck when they are omitted from training. Figure 3 documents the strong out-of-sample

performance of the estimation strategy in terms of the absence of Type I and Type II

errors across repeated cross-validation draws from the data. This figure presents the

distribution of the F1 score, which is harmonic mean of the absence of Type I errors

(precision) and the absence of Type II errors (recall).41 These cross-validation results are

consistent with the estimator’s in-sample performance.

38By collapsing to the author level, we defer a study of within-author variation. The sacrifice here isminimized by the fact that, while first generation Protestant reformers were former Catholics, extremelyfew Protestant authors had any substantial publication record as Catholics. As a result, the “withinauthor” variation available to study is largely within religion.

39Unlike the imbalance in the number of titles by religion, the number of authors by religion is relativelywell-balanced.

40For comparison, Gentzkow and Shapiro (2010) predict the party of US Congress members based onthe text of their speeches, and obtain a correlation of 0.61 between true and predicted affiliation.

41Appendix Identifying Partisan Language B provides discussion of the F1 score and of how the esti-mation strategy performs under different hyperprior specifications to address concerns about overfitting.

15

Figure 2: Out-of-sample classification of religious authors

Protestant Catholic0.0

0.2

0.4

0.6

0.8

1.0

Pr(P

rote

stan

t==1

)

Martin Luther

Johannes Eck

This graph presents the out-of-sample classification performance for a single draw from the data. Themodel is trained on 80% of the data and predicted on the held-out 20%. We present predictions forheld-out Protestants (at left) and for held-out Catholics (at right). We highlight Martin Luther andJohannes Eck as the most important Protestant and Catholic authors in our held-out data.

Figure 3: Absence of type I and type II errors in out-of-sample classification

0

2

4

6

8

10

Den

sity

.7 .75 .8 .85 .9 .95F1 Measure

This graph presents the kernel density of the F1-measure for out-of-sample classification over repeatedsamples from the data. The F1-measure is harmonic mean of the absence of Type I errors and theabsence of Type II errors. We calculate the F1-measure for 100 runs of the model, each trained on arandom 80% of the data and predicted on the held-out 20%. For details see Appendix B.

16

5 The Diffusion of Religious Ideas in the Media

5.1 Measuring Diffusion

Two key facts characterize the diffusion of Protestant ideas in the media – a discon-

tinuous shift following Martin Luther’s intervention in 1517 and considerable city-level

heterogeneity post-1517. Neither of these facts have been previously documented quan-

titatively, to the best of our knowledge.

Figure 4 highlights the discontinuity in media content in 1517 by presenting the mean

of the estimated index of religious content in the media for all of Germany. Figure 4

shows that the religion index was approximately stable in the run up to the Reforma-

tion, and a discontinous shift towards Protestant ideas after 1517. The first panel shows

our estimates for all religious publications. The second panel examines religious publi-

cations in German and suggests a slight pre-1517 trend away from Catholic-type speech

in vernacular media, consistent with the observation that Protestantism was in part a

response to underlying cultural trends (Cameron 1991).42 In German language media,

there is also a relatively much larger increase in the number of publications printed post-

1517, as indicated by the scale of the annual markers. The third panel shows that the

discontinuous shift towards Protestant-type speech also characterizes religious media in

Latin.

Figure 4: Index of religious ideas in the media

0

.25

.5

.75

1

Rel

igio

n In

dex

1475 1500 1525 1550 1575 1600

All Religious Media

0

.25

.5

.75

1

Rel

igio

n In

dex

1475 1500 1525 1550 1575 1600

Religious Media in German

0

.25

.5

.75

1

Rel

igio

n In

dex

1475 1500 1525 1550 1575 1600

Religious Media in Latin

This graph plots the annual mean of the religion index across all religious publications in all historicallyGerman-speaking cities in our data, where Protestant = 1 and Catholic = 0. Marker sizes are scaledto represent the relative number of publications in each year. The vertical line marks 1517, the yearMartin Luther circulated his 95 theses criticizing the Catholic church.

Figure 5 maps the heterogeneity in media output at the city level. In Figure 5, city

42For example, humanistic ideas and ideas about pan-German identity were developing before 1517.

17

markers are scaled to reflect the number of titles produced and shaded to reflect the mean

of the Religion Index for city-level output. Lighter markers indicate more Catholic media.

Darker markers indicate more Protestant media. We present data on media output for

the period between Luther’s intervention in 1517 and the Peace of Augsburg of 1555.

Figure 5: Variation in religious ideas in the media across cities

Protestant IndexLowestLowMidHighHighest

This map presents information on the overall size of media output and on the religious ideas in the mediaat the city level between 1518 and 1554. Markers are shaded to reflect the average value of the religionindex. Shading is by index quintile. Darker markers represent relatively more Protestant ideas. Markersare scaled to reflect the number of publications (in logs). Only cities with positive media output areincluded.

5.2 Determinants of Diffusion

In this section we test the hypothesis that pre-Reformation industrial structure in media

markets had implications for competition that shaped the diffusion of Protestant ideas.

Historical evidence motivates the hypothesis that Protestant ideas diffused more in

more competitive media markets (Pettegree 2000; Creasman 2012). However, the rela-

tionships between printers and city councils provide a particular reason to hypothesize

that the difference between competition and monopoly may have been important. City

councils across Germany were initially opposed to reformist arguments (Cameron 1991;

Dickens 1979; Sehling et al. 2009). City councils were also a source of important work

18

orders for printers. City councils commissioned printers to produce forms, proclama-

tions, and other “jobbing” work. These smaller scale projects could be an important

source of revenue and could be used ensure fixed capital was employed while larger book

projects were set up (Febvre and Martin 1958). Once the Reformation began, orders

from the municipality could be threatened by the decision to print Protestant content

at odds with city council policy (Reske 2007). In cities with one printer, that printer

was de facto if not de jure the city printer. Faced by an adversarial monopolist, city

councils might encourage a new entrant. In cities with multiple printers, one printer was

typically the official city printer (ratsbuchdrucker) and the rest were not. The fact that

known city printers were not early advocates of the Reformation is consistent with the

view that they did not want to endanger official work orders or antagonize city govern-

ments.43 More broadly, there is considerable evidence that individual printers produced

both Catholic and Protestant materials and that competitive pressures induced printers

to produce Protestant materials that were at odds with city council views: “Numerous

studies of book production and readership in early modern Germany have documented

the popularity of controversial religious polemic. . . In the financially risky book business,

it was vital that dealers be able to gauge public interest and respond to market demand

as soon as possible” to avoid “competitive disadvantage” (Creasman 2012).

Motivated by this evidence, we first study the relationship between the number of

firms competing and diffusion. We document the relationship shared across all cities and

the differential relationship between the number of firms competing and the diffusion of

Protestant ideas where formal political freedom was more restricted by cities’ constitu-

tional status. We then show that the big effects in competitive conduct are observed when

we compare cities with two or three firms to cities with monopolies. Finally, we provide

evidence that competition mattered more for Protestant content than for Catholic.

We first study the relationship between the number of firms active in the pre-Reformation

period and the subsequent diffusion of Protestant media at the city level. Our motivating

estimates are obtained from regressions of the form:

protestanti,post = αfirmsi,pre + βXi,pre + εi (2)

43Existing evidence does not enable us to definitively identify all city printers because the existingdata record the printers who produced books and pamphlets but not the forms, bills, and administrativeephemera that were often the source of municipal contracts. However, what we observe supports theview that particular printers were the privileged recipients of city work. For example, in Mainz, allproclamations printed 1508-1517 were produced by Johann Schoffer. In Munich, all proclamationsprinted 1508-1517 were produced by Hans Schobser. In Strasbourg, starting in 1514 all city councilpublications were produced by Matthias Shurer. In Cologne, all but two proclamations authored by thecity council 1508-1517 were printed by Hermann Bungart. In Speyer, Peter Drach III produced bothmunicipal proclamations recorded 1508-1517. In Heidelberg, Jakob Stadelberger similarly printed bothproclamations we observe published in this period.

19

Here protestanti,post measures Protestant content produced between 1518 and 1554 in

city i. Our baseline measure of Protestant content is the logarithm of the number of

Protestant books plus one, but we obtain similar results studying variation in the con-

tinuous measure of predicted “Protestant” and estimating negative binomial regression

models examining the count of Protestant books including zeros. The variable firmsi,pre

is the number of firms active 1508-1517.44 We control for the quantity of city-level print

media output in German and in Latin and the share of city output on religious subjects

between 1454 and 1517, population in 1500, the presence of universities, and distance to

Wittenberg, where Luther was based. Because access to trade networks shaped how ideas

spread, we include indicators for cities that were members of the commercial Hanseatic

League or were on navigable rivers. To account for variations in city institutions we

include indicators for cities subject to lords (landestadt), as opposed to free cities. We

include indicators for cities subject to ecclesiastical rule and for prince-bishoprics, both

of which are subsets of cities subject to lords. We also introduce fixed effects for princi-

palities and for regions to study variation across cities subject to the same territory-wide

laws and ruler or regional preferences. These variables are captured by X.45

Table 2 reports results from regression estimates of the relationship between the num-

ber of firms in media markets on the eve of the reformation and the number of Protestant

media varieties produced between 1518 and 1554. Across specifications an additional firm

was associated with at least a 30% increase in Protestant media.46 The estimated rela-

tionship between competition and diffusion holds within political territories and within

local geographic regions. In column 4 we show that the relationship between ex ante

industrial structure and ex post diffusion is if anything stronger within geographically

defined latitude-longitude grid cells of approximately 210 km by 210 km (2-degree lati-

tude by 3-degree longitude). In column 5, we show that the relationship between firms

44We study the number of firms active in the ten-year period 1508-1517 as a baseline, however similarresults obtain for other relatively short windows pre-1518.

45To test the hypothesis that the presence of religious establishments was a determinant of diffusion,we construct comprehensive city-level data on monasteries in each city as of 1517. We find no significantrelationship between religious establishments and diffusion and that our estimates for competition arenot changed when we control for these establishments. We also find that controlling for pre-reformationdiffusion of Germanic humanism does not change our estimates – using as our measure publications byJohannes Trithemius, Johannes Geiler von Kaisersberg, Sebastian Brant, Jakob Wimpfeling, JohannAventinus, and Conrad Celtis. We also observe whether a city had formal market rights as of 1517 andthe number of market rights granted using data from Cantoni and Yuchtman (2014) that are restrictedto cities located in contemporary Germany. We find no systematic relationship between the marketrights and the diffusion of the Reformation in media or law. We do not report all these specifications.

46We report standard errors clustered by historical principalities (“states” in the regression table).Standard errors are almost unchanged if we do not cluster at the principality level. Historical princi-palities are mapped in Figure 1 and discussed in the appendix. Our results hold when we cluster onalternative geographic units (see below) and when we adjust standard errors with the wild bootstrapmethodology of Cameron et al. (2008).

20

and diffusion holds within 70km grid cells.47 In columns 6 and 7 we restrict the sample to

states with at least two cities. The states we study are the historic territories of the Holy

Roman Empire (Duchies, Margravates, Electorates, etc.).48 In column 6 we estimate a

regression without fixed effects and in column 7 we introduce territory fixed effects. In

this subset of the data, state fixed effects again scarcely shift the point estimates, however

the point estimate is approximately 50%-100% times larger than in the complete data.

This is because the set of places located in principalities with two or more cities excludes

free cities, so we effectively recover a slope for cities subject to feudal lords. This result

points to important heterogeneity between free cities and cities subject to lords.

Table 2: Industrial structure and Protestant ideas

Ln Protestant 1518-1554[1] [2] [3] [4] [5] [6] [7]

Firms 1508-1517 0.66*** 0.61*** 0.36*** 0.46*** 0.31* 0.61** 0.64**(0.07) (0.10) (0.12) (0.11) (0.17) (0.22) (0.29)

Population in 1500 Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesCluster FE Yes Yes YesObservations 191 191 191 191 191 101 101Cluster Definition State State State 210km 70km State StateTerritories All All All All All Multi- Multi-in Sample City City

This table reports OLS regressions of the relationship between Protestant media and pre-Reformationfirms at the city level. The dependent variable is the log of the number of Protestant publications plusone. “Firms 1508-1517” is the number of firms active 1508-1517. Controls include: Latin Media pre-1517 and Vernacular Media pre-1517 measured in hundreds of titles; Religious Media pre-1517 measuredas the share of titles on religious topics; distance to Wittenberg measured in hundreds of kilometers;indicators for Hansa cities, ecclesiastical rule cities, prince bishroprics, cities on navigable rivers, andcities ever printing pre-1517. Population in 1500 is controlled for with fixed effects for bins: unknown(omitted), 1000-5000, 6000-1000, 11000-25000, 26000+. Standard errors clustered on state or map gridcell. The states are the historic principalities of the Holy Roman Empire. Free cities are assigned theirown cluster. The “210km” cluster designates 2 degree × 3 degree grid cells. The “70km” grid celldesignates grid cells of approximately 70km. Columns [6] and [7] restrict to cities in principalities withat least 2 cities, and thus exclude principalities with single cities and free cities. Significance at the 99%,95%, and 90% levels denoted “***”, “**”, and “*”.

Table 3 presents evidence that the relationship between initial industrial structure

and subsequent diffusion of Protestant ideas was stronger in cities subject to territorial

lords. We introduce interaction terms to study the variation specific to cities with lords.

We find that the relationship is twice as strong for cities subject to lords. We also find

47There are 59 70×70 km grid cells, of which 14 contain just 1 city and 28 contain just 2 cities.48The appendix provides the number of towns in each principality and discusses the decentralized

nature of governance within principalities. The 191 cities in the sample are located in 29 grid cells.

21

that diffusion was not lower in cities subject to lords, conditional on industrial structure.

This variation cuts across the stylized fact that free cities were unconditionally more

likely to adopt the Reformation.49 In contrast, we observe that in the subset of cities

that were subject to ecclesiastical rule the diffusion of Protestant media was significantly

lower. This evidence suggests that ex ante industrial structure in media markets mattered

more for subsequent diffusion where municipal rulers were legally subject to authoritarian

rulers, but that where Catholic Church officers were themselves the feudal lords diffusion

was significantly depressed. Social history evidence strongly suggests that ecclesiastical

rule limited the diffusion of Protestant media. But it is also noteworthy that of the 21

ecclesiastical cities in our data, only 5 had printing firms active in the decade before the

Reformation, and only three of these five cities had more than one firm. We find no

evidence of an interaction between industrial structure and ecclesiastical status.

Given our interest in the roles of technology and competitive conduct in the diffusion of

the Reformation, our baseline results raise three natural questions. The first is whether

cities with more firms produced more Protestant content not because of competition,

but because having an extra firm linearly increases Protestant content in expectation.

Because research on concentrated markets suggests important changes in competitive

conduct are nonlinear in the number of firms (Bresnahan and Reiss 1991), it is natural

to wonder whether the underlying relationship in our setting is nonlinear. The second

question is whether Protestant and Catholic content responded in similar ways to vari-

ations in industrial structure. Because Protestant ideas attacked the Catholic Church

and incumbent elites, we hypothesize that competition had a differential impact on the

diffusion of Protestant ideas. The third question is whether the relevant variation is be-

tween printing and non-printing or between more and less competition. If technology was

the key driver of diffusion, we would expect the most salient distinction to be between

cities with and without printing. If competition was fundamental, we would expect the

distinction between cities with monopolies and cities with two or three firms to be salient.

To explore the implications of industrial structure, we first study the relationship

between variations in the number firms and the diffusion of Protestant ideas nonlinearly.

We examine the relationship between content diffusion and initial industrial structure by

studying how residual content – conditional on other observables and state fixed effects –

49In our baseline sample of Reske printing cities, 25 of 43 free cities passed Reformation laws, while56 of 148 territorial cities did. Our econometric results below are robust in an alternate larger samplesof cities that includes cities without printing and without population observed in 1500. Of the 85free imperial cities recorded in the tax register of 1521 (Reichsmatrikel), 73 fall within the geographicbounds of this study and of these 50 ultimately adopted the formal legal reforms of the Reformation.Schmidt (1984) argues that only 69 of the 85 cities listed in the Reichsmatrikel truly met the criteriafor institutional autonomy typically understood to define an imperial city. The results we report belowfollow the Reichsmatrikel classification, but are robust to using the alternative Schmidt classification.

22

Table 3: Industrial structure and Protestant ideas under lord rule

Ln Protestant 1518-1554[1] [2] [3] [4] [5]

Firms 1508-1517 0.58*** 0.49*** 0.29** 0.37*** 0.24*(0.05) (0.06) (0.12) (0.10) (0.12)

Firms 1508-1517 × Lord 0.55*** 0.51*** 0.39** 0.49** 0.47***(0.13) (0.14) (0.16) (0.19) (0.16)

Lord -0.40 -0.14 -0.24 -0.15 -0.32(0.30) (0.27) (0.29) (0.53) (0.60)

Ecclesiastical -0.77* -1.00*** -0.96**(0.39) (0.29) (0.37)

Population in 1500 Yes Yes Yes YesControls Yes Yes YesCluster FE Yes YesObservations 191 191 191 191 191Cluster Definition State State State 210km 70km

This table reports OLS regressions of the relationship between Protestant media and pre-Reformationfirms at the city level. The dependent variable is the log of the number of Protestant publications plusone. “Firms 1508-1517” is the number of firms active 1508-1517. “Firms 1508-1517 × Lord” is theinteraction between firms and an indicator for cities subject to lords. “Ecclesiastical” is an indicatorfor cities subject to ecclesiastical rule. Controls include: Latin Media pre-1517 and Vernacular Mediapre-1517 measured in hundreds of titles; Religious Media pre-1517 measured as the share of titles onreligious topics; distance to Wittenberg measured in hundreds of kilometers; indicators for Hansa cities,ecclesiastical rule cities, prince bishroprics, cities on navigable rivers, and cities ever printing pre-1517.Population in 1500 is controlled for with fixed effects for bins: unknown (omitted), 1000-5000, 6000-1000,11000-25000, 26000+. Standard errors clustered on state or map grid cell. The states are the historicprincipalities of the Holy Roman Empire. Free cities are assigned their own cluster. The “210km” clusterdesignates 2 degree × 3 degree grid cells. The “70km” grid cell designates grid cells of approximately70km. Significance at the 99%, 95%, and 90% levels denoted “***”, “**”, and “*”.

varied with the initial number firms using a local polynomial kernel regression. Figure 6

documents these nonlinearities for Protestant and Catholic content separately. There

are three key findings. First, the increase in diffusion is observed when we compare

markets with 2 or 3 competitors to markets with a monopoly. This finding is consistent

with contemporary evidence from concentrated markets documenting that almost all the

variation in competitive conduct occurs with the entry of the second and third firm

(Bresnahan and Reiss 1991; p. 978). Second, the point estimates for the two principal

types of religious media differ. The increase in Protestant media associated with two

to four firms is larger, and notable divergences open up at four and five firms. Third,

the point estimates are highly suggestive from a descriptive perspective, despite the fact

that the differences are clearly not statistically significant. The absence of statistical

significance reflects our sample size and the small number of cities with the same number

of firms active.

23

Figure 6: Nonlinear relationship between religious ideas and the number of firms

-.5

0

.5

1

1.5

Res

idua

l Pro

test

ant M

edia

0 1 2 3 4 5 6Firms Active 1508-1517

Protestant

-.5

0

.5

1

1.5

Res

idua

l Cat

holic

Med

ia

0 1 2 3 4 5 6Firms Active 1508-1517

Catholic

This graph plots the local polynomial regression estimates of the relationship between residual religiouscontent published between 1518 and 1554 and firms active 1508-1517. Residual Protestant content andresidual Catholic content are estimated with OLS, conditional on the complete set of city observables inTable 2, including state fixed effects, except firms active 1508-1517.

The evidence we observe on nonlinearities suggests the contrast between monopoly

and competition was salient. We are interested in how monopoly and competition pre-

dict the diffusion of Protestant content after the new content hits markets, how these

dynamics compare between cities in general and cities subject to lords, and whether

the competitive responses we observe for Protestant media were mirrored for Catholic

content. To formalize these comparisons we estimate regressions of the form:

publicationsi,r = α0 + α1competitioni + α2monopolyi + βXi + εi (3)

Here publicationsi,r is natural logarithm of the count of publications of religion r (Protes-

tant or Catholic) plus one in city i, competitioni is an indicator for cities with more than

one firm active, and monopolyi is an indicator for cities with one firm active. The omitted

category comprises cities with no firms initially active and is relevant because we observe

entrants and output post-1517 in cities that ex ante were without printing.50

Table 4 documents the difference in the way competition shaped the diffusion of

Protestant and of Catholic media by comparing post-1517 output across cities with com-

petition and cities with local monopolies on the eve of the Reformation. Panel A studies

all cities in our data. Columns 1 and 2 present baseline estimates showing how compet-

50The dynamics and implications of this entry are topics we currently defer for future research.

24

itive markets were associated with greater subsequent diffusion of both Protestant and

Catholic content, conditional on a rich set of characteristics. Column 3 shows that the

competition effect was significantly larger for Protestant media. Columns 4-6 extend the

analysis to study the interaction between competition and monopoly and lord rule. It

is also notable that cities with monopolies saw less diffusion than the omitted category

of cities without any firms in the early 1500s, although these estimates are not statisti-

cally significant. When we condition on the lord interactions, the competition effect for

Protestant media is weakly significant and for Catholic cities is insignificant for all cities

(first row). However, the total effect of competition for cities subject to lords is strongly

significant and again significantly larger for Protestant media. Because the results of our

polynomial regression analysis show considerable differences for Protestant and Catholic

media for cities with five and six firms, it is natural to wonder to what extent our findings

here may be driven by differences in the upper tail of cities with many firms. Panel B

restricts analysis to cities with not more than 4 firms. In this subset, the competition

effects are larger in size and highly statistically significant.

The asymmetry in response to competition between Protestant and Catholic re-

flects the competitive environment of the Reformation era. Before the Reformation,

the Catholic church enjoyed something approaching an ideological monopoly.51 Prior to

1517, ideological competition over religion in the European media was minimal or at least

very tightly bounded. The Protestant ideas that diffused post-1517 challenged the ideo-

logical monopoly of the Catholic church and were soon understood to be heretical. As a

result, producers faced incentives to differentiate, but differentiation potentially carried

political economy and business implications.

5.3 Endogeneity

Our baseline analysis takes initial differences in industrial structure as given and examines

the sudden appearance of Reformation ideas as a supply shock. The natural question is

whether variations in initial competitive environments caused subsequent differences in

the diffusion of Protestant ideas, or whether the variations in competition and diffusion

reflect underlying differences in preferences or institutions.

We use the deaths of printers active in the period 1508-1517 as an instrumental

variable (IV) that provides plausibly exogenous variation in competition. By examining

deaths in the period just before the Reformation, we study shocks to competition that are

51This characterization is designed as a high level approximation. The social history literature providesevidence on the heterogeneity of views within Catholicism and the fact that the ideological monopolywas never complete.

25

Table 4: Competition and monopoly as determinants of religious ideas

Panel A: All CitiesProt Cath Diff Prot Cath Diff[1] [2] [3] [4] [5] [6]

Competition 2.30*** 1.01* 1.28*** 1.82** 0.63 1.18***(0.76) (0.57) (0.44) (0.90) (0.60) (0.46)

Monopoly -0.84 -0.71 0.13 -0.58 -0.93* 0.36(0.70) (0.47) (0.42) (0.81) (0.47) (0.53)

Competition × Lord 0.90 0.61 0.29(0.94) (0.69) (0.62)

Monopoly × Lord -0.59 0.28 -0.87(1.11) (0.83) (0.78)

Observations 191 191 — 191 191 —Total Competition 2.71*** 1.24* 1.47**Effect Under Lords (0.81) (0.68) (0.61)

Panel B: Cities with Not More Than 4 FirmsProt Cath Diff Prot Cath Diff[1] [2] [3] [4] [5] [6]

Competition 3.97*** 1.88*** 2.09*** 3.45*** 0.98 2.48***(1.08) (0.63) (0.60) (0.98) (0.62) (0.59)

Monopoly -2.38** -1.53*** -0.85 -1.92** -1.13** -0.79(1.02) (0.56) (0.65) (0.87) (0.54) (0.73)

Competition × Lord 1.12 1.58** -0.47(0.98) (0.71) (0.54)

Monopoly × Lord -1.03 -0.88 -0.15(1.22) (0.87) (0.78)

Observations 183 183 — 183 183 —Total Competition 4.57*** 2.56*** 2.01***Effect Under Lords (1.33) (0.77) (0.74)

This table reports OLS regressions of the relationship between the diffusion of Protestant and Catholicmedia and city-level industrial structure. The dependent variable is the log of the number of publicationsproduced between 1518 and 1554 plus one. Regressions studying Protestant output have column header“Prot”. Regressions studying Catholic output have header “Cath”. “Competition” is an indicator forcities with more than one firm active 1508-1517. “Monopoly” is an indicator for cities with just onefirm. The omitted category is for cities with no firms. “Competition × Lord” interacts indicators forcompetition and rule by a feudal lord. “Monopoly × Lord” interacts indicators for monopoly and feudallord. Column 3 presents the difference between the parameter estimates for Protestant (column 1) andCatholic (column 2). Column 6 presents the differences between estimates for Protestants (column 4)and Catholics (column 5) in the specifications with lord interactions. All regressions use the completeset of controls from Table 2. Standard errors clustered on territorial principality (state). Significance atthe 99%, 95%, and 90% levels denoted “***”, “**”, and “*”.

free from the sorts of endogeneity that could characterize entrance decisions or city level

responses to deaths and potential entrance during the Reformation itself.52 We argue that

52This distinguishes the methodological scope and aims of our study from research that studies incen-

26

these printer deaths are a source of plausibly exogenous variation in local competition

conditional on the total number of printers active in a city and the full set of observable

city characteristics – including the size of the media market and the composition of media

output before the Reformation. More broadly, our identification strategy is motivated by

the fact that the underlying trends and features in media markets are relatively smooth

compared to the lumpy nature of competition between individual firms and in particular

variations in competition induced by shocks. The measure of printer deaths we use is

the number of deaths occurring not more than one year after a printer’s last business

activity. This measure allows us to discriminate between the deaths of active printers

(which matter for competition) and the deaths of retirees, which serve as a sort of placebo.

We provide additional discussion of evidence for the identifying assumptions below.

The deaths of active printers had an impact on competition because concentrated

local media markets were characterized by a small number of firms involved in strate-

gic behavior designed to limit entrance and competition. Cartels and monopolies were

endemic (Reske 2007; Dittmar 2015).53 The social history record contains considerable

evidence on printers coordinating business activities in cartels, engaging in joint opera-

tions, and even signing formal contractual agrements to govern cooperactive activies. For

example, in Augsburg, Johan Otmar (in business from 1502 until death in 1516) coordi-

nated production with Erhard Ogelin and Johann Schonsperger. In Cologne, Eucharius

Cevicornus cooperated with Johannes Prael, Hero Fuchs, Peter Quentel, and Johann

Gymnich (Reske 2007; p. 430). In Frankfurt am Main, Ambrosius Lacher collaborated

with Konrad Baumgarten and Balthasar Murrer collaborated with Nicolas Lamparter. In

Basel, Johannes Petri (active 1502-1507, died 1511) coordinated with Johann Froben and

Johann Amerbach. In early modern European printing, formal joint operations agree-

ments were typically of a set duration of under 10 years. In addition, anti-competitive

strategic behavior was common. Andre Wechel, a printer based first in Paris and then

in Frankfurt am Main, was a leading producer for the lucrative textbook market in the

1500s. Maclean (2009; p. 177) observes that when competitors printed school books,

Wechel used a quantity competition response strategy and unleashed, “a massive and

systematic onslaught. . . aiming at little short of a monopoly. . . by putting into practice

the commercial principle: if a competitor produces an edition, do the same.”

In the concentrated media markets of early 1500s Germany, shocks to incumbents

induced entrance and led to net increases in the number of competitors. In the data

spanning the entire 1500s, we observe that within cities, and even within city-decades,

tives to ideological positioning for entrants (Gentzkow et al. 2014).53The costs of inter-city trade provided a level of protection against inter-city competition and are

discussed in the appendix.

27

entrance increases in the precise years when printers die leading to persistent increases

in the number of firms. Studies examining data from across Europe similarly documents

increases in entrance and persistent increases in the number of firms competing at the

city level in the precise years when printers die (Dittmar 2015).54

We estimate a two-stage least squares instrumental variable regression, where the first

stage estimates the relationship between industrial structure and printer deaths:

firmsi,1508−1517 = αdeathsi,1508−1517 + βfirmsi,1498−1507 + θXi,pre + εi (4)

We then use the variation in industrial structure induced by variation in the number of

deaths to re-estimate our baseline specification from equation (2).

Table 5 presents instrumental variable estimates. Panel A presents first stage esti-

mates. We find that a death was associated with an additional firm.55 We find this

relationship holds as we add controls for population and city characteristics, and within

territories and small grid cells. We also find that the deaths of retirees have no impact

on industrial organization, as shown in column 6. By controlling for the total number

of firms and the deaths of retirees, we provide evidence that suggests the effects are not

driven by differences in the total number of printers. Panel B presents 2SLS estimates

showing how induced variations in competition explain subsequent differences in Protes-

tant media. We find that an extra firm is associated with output effects of 133% to 266%

(0.85 to 1.3 log points).56 For comparison, we found comparable OLS estimates ranging

between 35% (0.3 log points) for all cities and 100% (0.7 log points) for cities subject

to lords. When we studied the difference between competition and monopoly, we found

that cities with competition produced in excess of 600% (2 log points) more Protestant

media than comparable cities characterized by monopoly. Three factors explain these

differences. First, the IV imposes linearity. Second, most deaths did not occur in cities

with monopolies so the variation the IV captures is principally for cities with competi-

tion. Only three cities with deaths 1508-1517 had monopolies in the period 1497-1508.

54Dittmar (2015) assembles panel data for all firms printing in Europe 1454-1600 and documentssignificant increases in entrance in the precise city-years in which printers died, even controlling forbusiness conditions in media markets in a tight windows captured by city-cross-five-year-period fixedeffects. However, Dittmar (2015) is only able to examine the subset of deaths recorded for survivingfirms. In the more complete data for German firms which we examine for this paper – which record alldeaths – identical findings hold: Deaths increase competition at the city-year level in the 1500s.

55We find no systematic evidence suggesting that new entrants were acquiring their printing pressequipment from the firms experiencing the death, in the over 200 individual instances where Reske(2007) records how printers acquired their machinery.

56Table 5 shows that the first stage gets stronger when we introduce controls. The key controls whichdrive this greater precision are the measures of pre-Reformation media output. Controlling for justpopulation and our measures of pre-Reformation output, 2SLS estimate is 1.29 (standard error 0.15)and the first stage F statistic is 59.9.

28

Third, the IV may recover a local average treatment effect as we discuss below.

Table 5: IV estimates of the impact of competition on Protestant ideas

Panel A: First Stage Dependent Variable Firms 1508-1517[1] [2] [3] [4] [5] [6]

Deaths Active 1.22** 1.25*** 0.99*** 1.04** 0.96** 0.98***(0.48) (0.45) (0.14) (0.15) (0.38) (0.15)

Deaths Retired 0.07(0.25)

Population Yes Yes Yes Yes YesControls Yes Yes Yes YesCluster State State State State 70km StateCluster FE Yes YesObservations 191 191 191 191 191 191

Panel B: 2SLS Dependent Variable Ln Protestant Media 1518-1554[1] [2] [3] [4] [5] [6]

Firms 1508-1517 1.37 1.20*** 1.31*** 0.86*** 1.02* 1.29***(0.86) (0.45) (0.25) (0.26) (0.48) (0.57)

Observations 191 191 191 191 191 191F Statistic on IV 6.54 7.83 51.85 44.68 6.44 26.14Population Yes Yes Yes Yes YesControls Yes Yes Yes YesCluster State State State State 70km StateCluster FE Yes YesObservations 191 191 191 191 191 191

Panel A presents first stage regression estimates. The dependent variable is the number of firms active1508-1517. All specifications control for the number of firms active 1498-1507. “Deaths Active” isthe count of deaths of active printers. “Deaths Retired” is the count of deaths of retired printers.“Population” denotes categorical indicators for population in 1500. Controls are as in Table 2. Statefixed effects control for principalities in Euratlas. Free cities are classed following the Euratlas coding asbelonging to “small states” rather than as individual unique territories. Panel B presents 2SLS estimatesof relationship between initial firms and the diffusion of Protestantism, based on first stage predictions.The dependent variable is logarithm of Protestant titles between 1518 to 1554 plus one. The controlsin Panel B match those in Panel A. Standard errors clustered by state except where “70km” grid cell isspecified. Significance at the 99%, 95%, and 90% levels denoted “***”, “**”, and “*”, respectively.

The assumptions on which identification rests here are supported by historical evi-

dence. The first assumption is the conditional exogeneity of deaths: The occurrence of

deaths in a narrow time period on the eve of the Reformation is random conditional on

observables. The observables we condition on include the number of firms in each city

ex ante, as well the size of the local media market, and the extensive set of city level

characteristics we observe. The plausibility of conditional exogeneity is strengthened by

the fact that we also observe the deaths of retirees, which have no impact on competition

or the diffusion of Protestant media post-1517.

29

The second assumption required for identification is that the deaths of printers in the

immediate pre-Reformation period impacted the subsequent diffusion of religious ideas in

the media only through their impact on the industrial structure of local media markets.

The historical evidence overwhelmingly suggests that these deaths must have shaped

diffusion through a media market channel, but it is natural to wonder whether a channel

besides competitive conduct associated with industrial structure could have been active.

One alternate candidate explanation stands out – that printer deaths may have shifted

the composition of output by altering the age distribution of the local printers. One

could imagine that by shifting the mix of printers towards youth, deaths also shifted the

mix to one more open to innovation and risk. To test this idea, we code the age of each

firm, but we find no relationship between either the average firm age, the median firm

age, or the minimum firm age and the subsequent diffusion of Protestant content. These

results are shown in the appendix.57

A related set of questions concern the way in which printer deaths translated into

variations in competition and whether there are sources of heterogeneity in response to

shocks. Our research strategy is designed to address these questions. We study deaths

and entrance before the Reformation because in this period entrance was not shaped

by Reformation-era incentives or concerns relating to new printers producing Protestant

content. It is also important to note that entrance was free before the Reformation

(Fussel 2005; Barbier 2006; Nicholas 2003; Hirsch 1974). Free entrance matters because it

rules out potential forms of confounding variation. We may wonder whether cities where

industrial structure shifted in response to printer deaths (the so-called IV compliers)

differ on unobservable city council openness to innovation. Free entrance implies that

our results are not explained by more open city authorities responding more permissively

to printer deaths. That said, the IV estimates recover a local average treatment effect.

It is possible that deaths of printers occurred selectively – or indeed randomly – in the

sorts of cities that were ex ante predisposed to innovation in religion. While we cannot

rule this out definitively, we observe no relationship between the city-level rate of printer

deaths in earlier periods and death rates in this key period, we see no effects associated

with the deaths of retired printers, and the key identifying assumption is that the timing

of premature death is effectively random conditional on observables.

57We use firm age because birth dates are not known for a substantial number of printers and becauseprinter age and firm age are highly correlated for printers with known dates of birth. The hypothesisthat the age of printers mattered is suggested to us by evidence on the generational differences historiansobserve among academics during the Reformation. For example, at the University of Wittenberg, Lutherquickly gained the support of his fellow junior faculty members, while senior colleagues remained initiallyskeptical and resistant to his arguments (Cameron 1991). The fact that age seems to have mattered inacademia but not in printing invites speculation about differences in the incentives facing professors andprofit-seekers.

30

6 Protestant Ideas and Institutional Change

The diffusion of ideas in the media led to changes in institutions designed to support

Protestant beliefs and ensure the persistence of Protestant reforms. As Ozment (1975; p.

49) observes, the reformist media that we study, “express viewpoints that were later em-

bodied. . . in church ordinances that consolidated the final phase of the Reformation. . . the

pamphlet became a church ordinance. . . the new Protestant institutions persisted.” The

city-level adoption of the Reformation in law was shaped by multiple factors. Here we

provide several pieces of evidence that show how Protestant media predicts institutional

change in the time series, including at the local level.

The institutional innovations formalizing the Reformation were designed to establish

the “Christian commonwealths” that reformers advocated and were first developed at

the city level (Hamm 1994; Cameron 1991). The church ordinances of the Protestant

Reformation formalized doctrinal change, set up anti-corruption safeguards for religious

activities, expanded public goods provision for the poor and needy, and established Eu-

rope’s first large scale experiments with mass public education (Cameron 1991).

Figure 7 shows the number of cities passing their first Reformation law each year

before 1555. In 1555 the Peace of Augsburg established a new religious equilibrium in

law and subsequent city-level institutional change was limited.

In Figure 8 we show the relative intensity of Protestant ideas in the media in cities

that did and did not adopt city-level Reformation laws. Prior to the Reformation, these

cities were producing similar religious media and there are no pre-trends, as measured

by our index. During the first years of the Reformation, Protestant content increases

in all cities, but cities that pass laws produce more. From the 1520s, a gap opens and

stabilizes.

A striking feature of the landscape is the fact that institutional changes varied across

cities at the very local level. Figure 9 maps the geographic distribution of city-level legal

reform and shows the spatial heterogeneity of institutional change.

Regression analysis confirms that local variations in exposure to Protestant ideas

in the media in the early years of the Reformation predict institutional change. We

examine linear probability models in which the passage of a law between 1521 and 1554

is predicted by exposure to Protestant media between 1518 and 1521. We condition on

the complete set of institutional observables and pre-Reformation city characteristics, and

study variation within states and within geographic grid cells. Our baseline estimating

31

Figure 7: The timing of city legal reforms

0

2

4

6

8

Law

s Pa

ssed

1510 1520 1530 1540 1550

This graph shows the the number of cities passing their first Reformation law each year before the 1555Peace of Augsburg.

Figure 8: Religious ideas in cities that did and did not pass legal reforms

0

.2

.4

.6

.8

1

Rel

igio

n In

dex

1475 1485 1495 1505 1515 1525 1535 1545 1555

Cities that Pass Laws Cities that do not Pass Laws

This graph presents the mean of the religion index for religious publication in for cities that did and didnot pass Reformation laws (Protestant = 1, Catholic = 0).

32

Figure 9: Cities that did and did not pass Reformation laws

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!!

!

!

!

!

!

!

!! !

!

!

!

!

!#I

#I

#I

#I

#I

#I

#I

#I

#I#I

#I#I

#I

#I

#I#I

#I

#I

#I

#I

#I

#I#I

#I#I

#I

#I#I

#I#I

#I

#I

#I

#I

#I

#I

#I

#I#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I#I

#I#I

#I

#I

#I

#I

#I#I

#I

#I#I

#I

#I

#I

#I

#I#I

#I

#I

#I#I

#I

#I#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I#I

#I#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I

#I#I

#I

#I

#I

#I#I

#I

#I

#I

#I#I

#I

#I#I

#I

#I

#I

#I

#I

#I

! Reformation law pre-1555#I No Reformation law pre-1555

This map shows the locations of cities that did and did not pass Reformation laws before 1555. Citiesthat passed Reformation laws are designated with circles, cities that did not are triangles.

equation is:

lawi,1521−1554 = α + β ln protestanti,1518−1521 + θXi + εi (5)

where lawi,1521−1554 is an indicator for cities with Reformation laws, ln protestanti,1518−1521

is the logarithm of Protestant media varieties plus one, and the Xi are city level observ-

ables. Table 6 shows that variations in Protestant ideas in the media predict institutional

change. Our estimates imply that a one standard deviation shift in Protestant content

between 1518 and 1521 is associated with a 6% to 9% increase in the probability a city

adopted the legal institutions of the Reformation (one standard deviation is 1.10). Our

estimates are relatively stable across specifications, and hold within tight geographic

regions including 70 km grid cells.

Because institutional change occurred at different times across cities, it is natural

to wonder how Protestant ideas in the media evolved both in the run-up to and in the

years following institutional change. Figure 10 presents the mean religion index for cities

that adopted Reformation laws over the 10 years before and after institutional change.

Figure 10 shows that the Protestant share of content increases dramatically in the run-up

to legal change and stabilizes afterwards.

To interpret these findings it is important to recognize that print media did travel

33

Table 6: Protestant ideas in print and the passage of Reformation laws

Reformation Law Passed 1521-1554[1] [2] [3] [4]

Ln Protestant 1518-1521 0.09*** 0.06*** 0.06 0.08**(0.03) (0.02) (0.04) (0.04)

Cluster Definition State State 210km 70kmCluster Fixed Effect Yes Yes YesObservations 191 191 191 191Controls Yes Yes Yes Yes

This table presents linear probability model estimates of the relationship between Protestant ideas inprint between 1518 and 1521 and the passage of Reformation laws between 1521 and 1554. The outcomeis a binary indicator for the passage of a law between 1521 and 1554. “Ln Protestant 1518-1521” is thelogarithm of the count of Protestant publications plus one. All specifications control for the complete setof controls in Table 2. Standard errors are clustered on state or on geographic grid cells as indicated. Thestates are the historic principalities of the Holy Roman Empire. Following the Euratlas classification,free cities are assigned to the “Small States of the Holy Roman Empire” state category. The “210km”cluster designates 2 degree × 3 degree grid cells. The “70km” grid cell designates 70km grid cells.Significance at the 99%, 95%, and 90% levels denoted “***”, “**”, and “*”, respectively.

Figure 10: Religious ideas in the media before and after legal reform

.4

.5

.6

.7

.8

Rel

igio

n In

dex

-10 -8 -6 -4 -2 0 2 4 6 8 10Year Relative to Passage of Reformation Law

This graph presents the mean of the religion index before and after the passage of Reformation laws(Protestant = 1, Catholic = 0). We restrict analysis to cities that did pass legal reforms and presentthe mean religion index with time centered at year 0, defined as the year in which each city’s firstReformation law was passed.

and that legal change reflected more than the presence of Protestant content in local

media output. In Schwabach, for example, reformist activity began with the distribution

34

of pamphlets at the church door before Sunday services in September 1524 (Dixon 1996).

No printing firms were active in Schwabach at this time – the pamphlets were printed

in neighboring cities. In our regressions above, we document that local variation in

the media predicts institutional change even in very tight geographic neighborhoods.

However, the Schwabach example supports a broader point: The cities of central Europe

were generically exposed to a wave of Protestant media, and how and whether exposure

to the media translated into social movements and institutional change in part reflected

variations in local demand for the ideas in the Protestant media.58

7 Conclusion

The Catholic Church enjoyed a virtual monopoly in religion in Western Europe that

lasted for centuries. This monopoly was buttressed by institutions and beliefs that se-

cured the overwhelming support of non-Church elites. The Catholic monopoly and these

arrangements were challenged by Protestant reformers in the early 1500s – in an era when

religious freedom and political participation were extremely limited.

The Protestant Reformation of the 1500s is widely considered a canonical example

of the way breakthroughs in information technology may drive large scale social change.

Gutenberg’s printing press technology preceded the Reformation by some fifty years, and

Protestant ideas diffused suddenly in the new media across cities in German-speaking

Europe. We assemble data on the universe of known books and pamphlets printed in

German-speaking Europe and construct a measure of religious ideas using estimators for

high dimensional data. No previous research has documented the diffusion of Protes-

tant ideas in the media in quantitative terms. We use the data to study economic and

institutional dimensions of diffusion.

We find that the diffusion of the Reformation was fundamentally driven by competi-

tion in new media markets, and not by the simple presence of printing press technology.

In cities with media market competition, Protestant ideas diffused dramatically. In cities

with media monopolies, they did not. In the cities where political freedom was most lim-

ited – where authoritarian local rulers enjoyed extensive legal prerogatives – competition

and openness in the media delivered their biggest effects on the diffusion of radical ideas.

The Protestant ideas that diffused in the media provided the outlines of a new social

order. We find that the local diffusion of Protestant ideas in the media predicts the legal

changes that institutionalized the Reformation at the municipal level.

58We defer a more focused study of variations in the demand for institutional change during theReformation to on-going and future research.

35

The Reformation is arguably one of the most important transformations in the market

for ideas, culture, and institutions in European history. It is also a canonical example

of the ways in which economic competition can drive political change in circumstances

where avenues for democratic participation are not available and where political freedom,

voice, and inclusion are severely limited.

References

Daron Acemoglu and James Robinson. Economic Origins of Dictatorship and Democracy.

Cambridge University Press, 2006.

Daron Acemoglu and James Robinson. Why Nations Fail: The Origins of Power, Pros-

perity, and Poverty. Crown Publishing Group, 2012.

Daron Acemoglu, Simon Johnson, and James Robinson. The colonial origins of com-

parative development: An empirical investigation. American Economic Review, 91(5):

1369–1401, 2001.

Daron Acemoglu, Davide Cantoni, Simon Johnson, and James Robinson. The conse-

quences of radical reform: The french revolution. American Economic Review, 101:

3286–3307, 2011.

Maja Adena, Ruben Enikopolov, Maria Petrova, Veronica Santarosa, and Ekaterina Zhu-

ravskaya. Radio and the rise of the nazis in germany. Quarterly Journal of Economics,

pages 1885–1939, 2015.

Paul Bairoch, Jean Batou, and Pierre Chevre. Population des villes europeennes de 800

a 1850. Droz, 1988.

Frederic Barbier. L’Europe de Gutenberg: le livre et l’invention de la modernite occiden-

tale, XIIIe-XVIe siecle. Belin, 2006.

Sascha Becker and Ludger Woessmann. Was weber wrong? a human capital theory of

protestant economic history. The Quarterly Journal of Economics, 124(2):531–596,

2009.

Josef Benzing. Die Buchdrucker Des 16. Und 17. Jahrhunderts Im Deutschen Sprachge-

biet. Otto Harrasowitz, 1982.

Peter Blickle. The Revolution of 1525: The German Peasants’ War from a New Per-

spective. Johns Hopkins University Press, 1981.

36

Johan Bollen, Huina Mao, and Alberto Pepe. Modeling public mood and emotion: Twit-

ter sentiment and socio-economic phenomena. In ICWSM, 2011.

Thomas A. Brady. Ruling Class, Regime and Reformation at Strassbourg, 1520-1555. E.

J. Brill, 1978.

Thomas A. Brady. Turning Swiss: Cities and Empire, 1450-1550. Cambridge University

Press, 1985.

Thomas A. Brady. Communities, Politics, and Reformation in Early Modern Europe.

Brill, 1998.

Thomas A. Brady. German Histories in the Age of Reformations, 1400-1650. Cambridge

University Press, 2009.

Timothy F. Bresnahan and Peter C. Reiss. Entry and competition in concentrated

markets. Journal of Political Economy, pages 977–1009, 1991.

Asa Briggs and Peter Burke. Social History of the Media: From Gutenberg to the Internet.

Wiley, 2010.

Philip Broadhead. Popular pressure for reform in augsburg, 1524-1534. In W. Mommsen,

editor, Stadtburgertum und Adel in der Reformation, Stuttgart. Ernst Klett, 1979.

Philip Broadhead. Guildsmen, religious reform and the search for the common good: the

role of the guilds in the early reformation in augsburg. The Historical Journal, 39:

577–597, 9 1996.

A. Colin Cameron, Jonah B. Gelbach, and Douglas L. Miller. Bootstrap-based improve-

ments for inference with clustered errors. Review of Economics and Statistics, 90(3):

414–427, 2008.

Euan Cameron. The European Reformation. Clarendon Press, 1991.

Davide Cantoni. Adopting a new religion: The case of protestantism in 16th century

germany. Economic Journal, 122(560):502–531, 2012.

Davide Cantoni. The economic effects of the protestant reformation: testing the weber

hypothesis in the german lands. Journal of the European Economic Association, 2014.

Davide Cantoni and Noam Yuchtman. Medieval universities, legal institutions, and the

commercial revolution. Quarterly Journal of Economics, 129(2), 2014.

P.W. Carey and J.T. Lienhard. Biographical Dictionary of Christian Theologians. Green-

wood Press, 2000.

37

Miriam U. Chrisman. Lay Culture, Learned Culture: Books and Social Change in Stras-

bourg, 1480-1599. Yale University Press, 1982.

Allyson Creasman. Censorship and Civic Order in Reformation Germany, 1517–1648:

‘Printed Poison & Evil Talk’. Ashgate, 2012.

Stefano DellaVigna and Ethan D. Kaplan. The fox news effect: Media bias and voting.

The Quarterly Journal of Economics, 122(3):1187–1234, 2007.

A.G. Dickens. Intellectual and social forces in the german reformation. In W. Mommsen,

editor, Stadtburgertum und Adel in der Reformation. Ernst Klett, 1979.

Jeremiah Dittmar. Information technology and economic change: the impact of the

printing press. The Quarterly Journal of Economics, 126(3):1133–1172, 2011.

Jeremiah Dittmar. New media, competition, and growth: European cities after guten-

berg. 2015. URL www.jeremiahdittmar.com/research.

C. Scott Dixon. The german reformation and the territorial society. German History,

14:123–140, 1996.

Peter A. Dykema and Heiko A. Oberman. Anticlericalism: in late medieval and early

modern Europe. E.J. Brill, 1993.

Mark U. Edwards. Catholic controversial literature, 1518-1555: Some statistics. Archiv

fur Reformationsgeschichte, 79, 1988.

Mark U. Edwards. Printing, Propaganda, and Martin Luther. University of California

Press, 1994.

Elizabeth L. Eisenstein. The Printing Press as an Agent of Change. Cambridge University

Press, 1980.

Lucien Febvre and Henri-Jean Martin. L’Apparition du livre. Evolution de l’humanite.

A. Michel, 1958.

John Flood. The book in reformation germany. In J.F. Gilmont, editor, The Reformation

and the book, St. Andrews studies in Reformation history. Ashgate, 1998.

Stephan Fussel. Gutenberg and the Impact of Printing. Ashgate, 2005.

Matthew Gentzkow and Jesse Shapiro. What drives media slant? evidence from u.s.

daily newspapers. Econometrica, 78(1):35–71, 2010.

38

Matthew Gentzkow, Jesse Shapiro, and Michael Sinkinson. The effect of newspaper entry

and exit on electoral politics. American Economic Review, 101(7):2980–3018, 2011.

Matthew Gentzkow, Jesse Shapiro, and Michael Sinkinson. Competition and ideological

diversity: Historical evidence from us newspapers. American Economic Review, 104

(10):3073–3114, 2014.

Luigi Guiso, Paola Sapienza, and Luigi Zingales. People’s opium? religion and economic

attitudes. Journal of monetary economics, 50(1):225–282, 2003.

Berndt Hamm. The urban reformation in the holy roman empire. In T.A. Brady, H.A.

Oberman, and J.D. Tracy, editors, Visions, Programs and Outcomes. Brill, 1994.

Rudolf Hirsch. Printing, selling and reading, 1450-1550. Harrassowitz, 1974.

Friedrich Kapp. Geschichte des deutschen buchhandels bis in das siebzehnte jahrhun-

dert. In Geschichte des Deutschen Buchhandels, Leipzig. Verlag des Brsenvereins der

Deutschen Buchhandler, 1886.

Wilbirgis Klaiber. Katholische kontroverstheologen und reformer des 16 jahrhunderts.

Reformationsgeschichtliche Studien und Texte, 136(116), 1978.

Ian Maclean. Learning and the Market Place : Essays in the History of the Early Modern

Book. Brill, 2009.

Bernd Moeller. Imperial Cities and the Reformation: Three Essays. Fortress Press, 1972.

Burt Monroe, Michael Colaresi, and Kevin Quinn. Fightin’ words: Lexical feature selec-

tion and evaluation for identifying the content of political conflict. Political Analysis,

16(4):372–403, 2008.

Michael Mullett. Historical Dictionary of the Reformation and Counter-Reformation.

Scarecrow Press, 2010.

David Nicholas. Urban Europe, 1100-1700. Palgrave Macmillan, 2003.

University of St. Andrews. Universal short title catalogue, 2012. URL http://arts.

st-andrews.ac.uk/~ustc/index.php#.

Steven Ozment. The Reformation in the Cities: The Appael of Protestantism to Sixteenth

Century Germany and Switzerland, volume 357. Yale University Press, 1975.

Steven Ozment. The Age of Reform, 1250-1550: An Intellectual and Religious History

of Late Medieval and Reformation Europe. Yale University Press, 1980.

39

Alexander Pak and Patrick Paroubek. Twitter as a corpus for sentiment analysis and

opinion mining. In LREC, 2010.

Andrew Pettegree. The Reformation World. Routledge, 2000.

Andrew Pettegree. Reformation and the Culture of Persuasion. Cambridge University

Press, 2005.

Christoph Reske. Die Buchdrucker des 16. und 17. Jahrhunderts im deutschen Sprachge-

biet: auf der Grundlage des gleichnamigen Werkes von Josef Benzing. Harrassowitz,

2007.

Christoph Reske. Die Buchdrucker des 16. und 17. Jahrhunderts im deutschen Sprachge-

biet: auf der Grundlage des gleichnamigen Werkes von Josef Benzing. Harrassowitz,

2015.

Bernd Roeck. Health care and poverty relief in counter-reformation catholic germany. In

O.P. Grell, A. Cunningham, and J. Arrizabalaga, editors, Health Care and Poor Relief

in Counter-Reformation Europe. Routledge, 1999.

Jared Rubin. Printing and protestants: an empirical test of the role of printing in the

reformation. Review of Economics and Statistics, 96(2):270–286, 2014.

Gerard Salton and Michael McGill. Introduction to modern information retrieval. 1986.

Heinz Schilling. The reformation in the hanseatic cities. The Sixteenth Century Journal,

14(4):pp. 443–456, 1983.

Georg Schmidt. Der Stadtetag in der Reichsverfassung. Franz Steiner, Stuttgart, 1984.

Robert W. Scribner. Why was there no reformation in cologne? Bulletin of the Institute

of Historical Research, 49(120):217–241, 1976.

Robert W. Scribner. The reformation as a social movement. In W. Mommsen, editor,

Stadtburgertum und Adel in der Reformation, Stuttgart. Ernst Klett, 1979.

Robert W. Scribner. For the Sake of Simple Folk: Popular Propaganda for the German

Reformation. Oxford University Press, Incorporated, 1994.

Emil Sehling, Institut fur Evangelisches Kirchenrecht, and Lutheran Church. Die Evan-

gelischen Kirchenordnungen des XVI. Jahrhunderts. O. R. Reisland, 2009.

Matt Taddy. The gamma lasso. arXiv preprint arXiv:1308.5623, 2013a.

40

Matt Taddy. Multinomial inverse regression for text analysis. Journal of the American

Statistical Association, 108(503):755–770, 2013b.

Matt Taddy. Measuring political sentiment on twitter: Factor optimal design for multi-

nomial inverse regression, 2013c.

Nico Voigtlander and Hans-Joachim Voth. Persecution perpetuated: the medieval origins

of anti-semitic violence in nazi germany. Quarterly Journal of Economics, 127(3):1339–

1392, 2012.

Joachim Whaley. Germany and the Holy Roman Empire: Volume I: Maximilian I to the

Peace of Westphalia, 1493-1648. OUP Oxford, 2011.

David Yanagizawa-Drott. Propaganda and conflict: Evidence from the rwandan genocide.

The Quarterly Journal of Economics, 129(4):1947–1994, 2014.

41

Appendices

42

A Data

A.1 Identifying Firms

In this section we describe how we construct evidence on the number of firms active in

local media markets.

We construct our preferred, baseline measure of firms from the authoritative bio-

graphical dictionary of printers in German-speaking Europe: Reske (2007; 2015), Die

Buchdrucker des 16 und 17 Jahrhunderts im deutschen Sprachgebiet: auf der Grundlage

des gleichnamigen Werkes von Josef Benzing.59 We call this our baseline or “biograph-

ical” measure. To identify firms from biographical data we manually code information

from Reske (2007; 2015). Reske (2007; 2015) is a biographical dictionary that has the

objective of providing an entry on every single printer active in German-speaking Europe

over our time frame. Reske provides the dates printers were active in different cities and,

for a large subset, the dates that they died. We manually code this information into a

database to identify (i) the firms active in different cities in different periods, (ii) the

dates printers died, and (iii) whether printers died while still active in business or when

retired.60

For robustness, we also construct a measure of firms from inscriptions on historical

books. We call this our “publications” measure. To construct a measure of firms from

publications data we identify firms producing from inscriptions on historical books and

pamphlets. Information identifying firms is available on 98% of historic media but in

highly non-standardized form. Printers are identified in multiple languages (e.g. La-

tinized and German variants of the same name), with non-standard spelling, abbrevia-

tions, and in some instances aliases.61 As a result, the data on firms from publications

includes a fringe of small, short-lived printers who do not appear in Reske (2007; 2015).

Table 7 provides an example of the underlying publication-level data and how we con-

struct standardized firm identifiers from publications data. Table 7 presents inscriptions

59Reske (2007; 2015) builds on the research in Josef Benzing’s Die Buchdrucker des 16 und 17 Jahrhun-derts im deutschen Sprachgebiet (Verlag Harrassowitz, 1963). Josef Benzing’s research was conductedwhile he was employed by the Preußischen Staatsbibliothek in Berlin 1934-1945 and subsequently whenBenzing was head librarian at the University of Mainz 1946-1966. Benzing was a member of the NationalSocialist party and during World War II coordinated both the cataloging and the theft of historic bookcollections in territories conquered by the Nazis.

60The Reske evidence on printer deaths is unique in the historiography of printing. Dittmar (2015)documents a selected set of printer deaths using inscriptions on books that indicate when printing is firstdone by “the widow of...” or “the heirs of...” This sort of book level primary source data are valuablebut only identify the deaths of printers whose firms survived to be managed by widows and heirs.

61We research each name and printer and construct standardized firm identifiers as described inDittmar (2015) and the appendix.

43

for 8 individual publications printed in Augsburg and provides a microcosm of the coding

we conduct for the full set of publications in our data.

Table 7: Example of how firms are identified from book-level inscriptions

Printer Inscription Firm Name #1 Firm Name #2(1) (2) (3)

Johann Schonsperger Johann Schonsperger

Johann Schonsperger & ThomasRuger

Johann Schonsperger Thomas Ruger

Gedruckt und volendt von AnnaRugerin in der keyserlichen statAugspurg

Thomas Ruger

Heinrich Steiner von Augsburg Heinrich Steiner

Heinrich von Augsburg Steiner Heinrich Steiner

Heinrich von Augsburg Steiner &haer. Erhard Oeglin

Heinrich Steiner Erhard Oeglin

excudebat Heinrich von AugsburgSteiner

Heinrich Steiner

apud Heinrich von Augsburg Steiner Heinrich Steiner

This table presents illustrative printer inscriptions on eight individual books published in Augsburg.

There are differences between the measure of firms from biographical data and from

publications evidence. The principal difference is that the publications data include a

fringe of small printer-publishers who are absent in Reske (2007; 2015). The publications

data record 167 printers-publishers active 1508-1517; the Reske biographical data records

115 printers. The printers in the publications data who are not recorded in Reske appear

to have left a minimal historical record. It is possible that printers named in inscrip-

tions produced under aliases that we incorrectly class as independent printers and that

occasional individuals brought out books under their name while actually working with

or using the equipment of an established printer. Brady (1998) observes that, “Almost

all the printers were investors, organizers, and managers of their firms.” While this is

true to a first approximation, Reske (2007) documents that occasionally printers pro-

duced in the workshops – and using the printing press equipment – of other printers.

Reske also documents that the occasional title would be produced under the name of a

printer-publisher who was not and did not emerge as a principal player in local media

markets. It is possible that some of the printers we observe in publications data but not

in Reske (2007; 2015) are of this type. In addition to these cases, there are a few printers

recorded in Reske who do not appear in our data because no evidence of their output

survives. When we compare our sources, we face generic economic questions about how

44

we define production units and where the relevant variation lies for thinking about mar-

ket structure, as well as challenges specific to data from the 1500s. Our two measures

of firms provide a rich characterization of the actors in, and the operation of, historic

media markets. In appendix E.2, we show that our results are robust to using either the

baseline (biographical) or alternate (publications) measure of firms.

Our principal measure of the number of firms competing in a media market counts the

number of independent firms active in a city-decade. The Reske data identifies printers

who operate independently directly. In the book-level data, we define as an independent

firm any printer who prints at least one item independently. This means that we do not

class as an independent firm, for example, a printer who only co-produced works as the

second named printer on a book front page. For example, in Augsburg in the 1508-1517

period Jorg Nadler produced as both a first-named and second-named printer. So our

baseline classification counts him as “independent.” In the same period Johannes Mader

only appears on a single book as the second-named printer. Our baseline classification

does not count Johannes Mader as an independent printer. However, our results are

robust to using every named printer recorded on any publication to form a maximal

measure of all possible, locally active printers. Summary statistics for both measures are

provided in section A.5.

A.2 Identifying Printer Deaths

This section describes how we construct our measure of shocks due to the deaths of active

printers. The fundamental point is that our measure of shocks is based on tracking every

printer active 1508-1517 in order to construct a comprehensive record of firms and cities

where active printers died.

To understand how we track all printers and their firms, it is important to understand

the nature of the historical data we construct. There are four types of firms in our baseline

data (Reske 2007; 2015).

1. There are firms active 1508-1517 that experience a death which leads to their exit

in this period. Deaths are observed from Reske (2007; 2015).

2. There are firms active 1508-1517 that experience a death but do not exit. These

firms survive because management passes to other family members (e.g. a widow

or other heirs). Deaths are observed from Reske (2007; 2015).

3. There are firms active 1508-1517 that do not experience a death over this period

and who do not exit. Survival without death is observed from Reske (2007; 2015).

45

4. There are firms that exit 1508-1517 without a death recorded in this particular

period. We call these the “untreated who exit.”

Our key questions about measurement and measurement error center on the last

group: firms that exit just before the Reformation without recorded deaths in this period.

The concern is two-fold. First, it could be that these exits were actually caused by

unobserved deaths. Second, it could be that deaths are recorded more frequently in

dynamic and sophisticated cities that were receptive to innovation in thought and belief.

Were this the case, the variation in competition induced by observed deaths could be

an artefact of the data preservation process, and the relationship between the induced

variations in competition and subsequent religious media could just reflect underlying

differences in cities – as opposed to exogenous shocks to competition.

We document that our data are free of this sort of underlying selection by providing

detailed information on the timing of death and circumstances of exit for each individual

printer who makes an exit that is not associated with a death (the “untreated who exit”).

Tables 8-11 document that unobserved deaths do not account for the exits by providing

individual-level data on the exit circumstances and life histories of every printer who

ceased production in some city over the period 1508-1517. Tables 8-11 show that we are

able to definitively document that 25 of these 31 exits were not caused by deaths by ex-

amining evidence on: (i) relocation decisions, (ii) the closure or sales of business , (iii) the

timing of other subsequent professional activities, and (iv) the timing of deaths observed

after printers exited production, and (v) other historical evidence from university and city

records. The remaining 6 printers where we are unable to pin down the timing of death

definitively cannot account for our results. Of these 6 printers, all were located in cities

that became Protestant. If we were to assume that these exits were caused by deaths – or

that some arbitrary subset of these exits were caused by deaths – our baseline results do

not change qualitatively and in fact because slightly stronger quantitatively. It is highly

unlikely that these 6 exits were actually caused by deaths. In the case of 2 of these 6

ambiguous exits the evidence very strongly suggests that death had no role in the close

of business. We class these as unlikely. The remaining 4 cases we conservatively class as

uncertain. This classification is conservative in the sense that it remains highly unlikely

that the death of an active printer would fail to be recorded in historical evidence given

(i) the importance of both printers and printer deaths to the local organization of an in-

dustry that had such an important place in urban culture and the dissemination of ideas

and (ii) the fact that printer deaths frequently initiated complicated legal proceedings

associated with estate management and outstanding obligations.

46

Name of Printer City Year Exit Death Evidence and Sources(1) (2) (3) (4) (5)

Georg Stuchs Nuernberg 1517 No Stuchs died in 1520. See Reske(2007).

HermannGuitschaiff

Cologne 1517 No Guitschaiff’s business operationsended when he was caught byCologne city authorities printinga Reformist (Protestant) tract in1517. This is recorded in cityarchives, “Im Libermalefactorum der Stadt Koeln.”See Hoffmann (2003),“Entreregionisierung im KolnerBuchdruck in den erstenJahrzehnten des 16.Jahrhunderts?” For discussion ofCologne, see Scribner (1972)cited in text and the mainbibliography.

Johann vonSolingen

Cologne 1517 No Von Solingen lived until at least1520. See Benzing (1982,Volume 2, p. 236).

KonradKachelofen

Leipzig 1517 No Kachelofen died in 1528. SeeReske (2007).

Konrad Kerner Strassburg 1517 No Kerner is observed in historicalrecords on the city of Rottenbergfrom the 1520s. See Febvre andMartin (1979, p. 293). Onpossible evidence that Kernerbriefly worked in Rothenburg abTauber in 1523, see Reske (2007,p. 879).

JakobStadelberger

Heidelberg 1516 Unlikely Stadelberger’s output wasobserved to be of strikingly poorquality [“auffallend schlechterQualitaet”] for the four yearsbefore exit. See Reske (2007, p.356)

Klosterdruckereider KaratauseSt. Barbara

Cologne 1516 No This was the printing house of areligious establishment.

MatthiasHupfuff

Strassburg 1516 No Hupfuff died in 1522. See Reske(2007).

Table 8: Printers exiting without death observed 1508-1517.

47

Name of Printer City Year Exit Death Evidence and Sources(1) (2) (3) (4) (5)

ThomasAnshelm

Tubingen 1516 No Anshelm moved from Tubingento Hagenau in 1516. See Reske(2007).

Klosterdruckereiwith LukasZeissenmair

Wessobrunn 1515 No This was the printing house of areligious establishment outsideour set of printing cities.

Konrad Hist Speyer 1515 No Hist died in 1531. See Reske(2007).

Privatepressedes HermannBarkhausen

Rostock 1515 No Barkhausen died in 1527. SeeReske (2007).

WilhelmSchaffner

Lahr 1515 No Schaffner moved from Lahr toStrasbourg in 1515. Reske(2007).

GregorBartholomaus

Basel 1515 Uncertain No historical information found

Wolfgang Huber Nuernberg 1514 No Huber’s business was taken overby Jobst Gutknecht in 1514(Reske 2007, p. 662)

JohannWeissenburger

Nuernberg 1513 No Weissenberger moved fromNuernberg to Landshut in 1513.See Reske (2007).

Nikolaus Keibs Durlach 1513 Uncertain See Gisela Moncke: “Zum fruhenDurlacher Buchdruck: NikolausKeibs,” in Gutenberg-Jahrbuch(1994).

Privatpresse desUlrich Pindermit DruckerFriedrichPeypus

Nuernberg 1513 No Peypus died in 1519. See Reske(2007).

SebaldusStriblita

Erfurt 1512 Uncertain See Martin von Hase, “SebaldusStriblita in Erfurt, der erstdeutsche Kursivdrucker (1510)”,Gutenberg-Jahrbuch, Bd. 11(1936) S. 94-97

Table 9: Printers exiting without death observed 1508-1517.

48

Name of Printer City Year Exit Death Evidence and Sources(1) (2) (3) (4) (5)

FreidrichHeumann

Mainz 1512 No Heumann lived until 1535. SeeFalkenstein (1840, p. 150)Geschichte derBuchdruckerkunrst in ihrerEntstehung und Ausbildung

Kolsterdruckereides Kartauserk-losters

Rostock 1512 No This was the printing house of areligious establishment.

Reinhard Beckd. A

Baden-Baden 1511 No Beck moved from Baden-Badento Strasbourg in 1511. See Reske(2007).

ThomasAnshelm

Pforzheim 1511 No Anshelm moved from Pforzheimto Tubingen in 1511. See Reske(2007).

Sixt Murner Strassburg 1510 Unlikely Murner is recorded as the printerof a single book written by hiscelebrated brother, the authorThomas Murner in 1510.

Hans Borchardwith ThomasBorchard

Hamburg 1510 Uncertain Hans [Johannes] Borchardproduces solo 1505-1510 (Reske,p. 332). Deutsche Biographieindicates that nothing more isknown after 1510:http://www.deutsche-biographie.de/sfz5287.html.Thomas Borchard only producesin 1491.

Nikolaus Kessler Basel 1510 No Kessler died in 1519. See Reske(2007).

BartholomausKistler

Strassburg 1510 No Kistler died in 1525. See Reske(2007).

Rudolf Spot Cologne 1509 No Spot lived until at least 1521.See Deutsches Jahrbuch furVolkskunde, Volume 6 (1960).

Table 10: Printers exiting without death observed 1508-1517.

49

Name of Printer City Year Exit Death Evidence and Sources(1) (2) (3) (4) (5)

Georg Richolffd. A furLaurenzBornemann

Muenster 1509 No Richolff moved from Muenster toLuebeck in 1509. See Reske(2007).

KonradBaumgarten

Frankfurt ander Oder

1509 No Baumgarten closed hisoperations by selling plant andequipment to Johann Hanau in1509. Johann Hanau operatedhis new business 1509-1540. SeeDeutsche Biographie:http://www.deutsche-biographie.de/sfz5223.html.

BalthasarMurrer withNikolausLamparter

Frankfurt ander Oder

1509 No Lamparter left Frankfurt forBasel in 1507. Murrer continuedas sole proprietor until he soldthe business to Johann Hanau in1510. See Reske (2007, p. 268).

AmbrosiusLacher

Frankfurt ander Oder

1508 No Lacher is observed engaged inacademic training and workingas academician from 1510sonwards for several decades. Forexample receives Lizentiat(degree) in Medicine in 1522 andworks as Rektor at theUniversity of Frankfurt. Dies in1540. See Deutsche Biographie,online: Deutsche Biographie:http://www.deutsche-biographie.de/sfz47283.html

Table 11: Printers exiting without death observed 1508-1517.

50

A.3 Text Data

Table 12 summarizes the distribution of the lengths of titles in our data, as is and

dropping “stop words” (words such as and and the).

Table 12: Summary statistics on the length of historical titles

Number of Words in TitlesMean 5% 25% 50% 75% 95%

Titles with stop words 23.2 3.0 11.0 21.0 33.0 51.0Titles without stop words 14.2 2.0 7.0 13.0 20.0 32.0

This table presents the mean and select percentiles of the title length distribution.

In the main body of the paper we present two examples of the titles of English-

language publications. In our complete data, 96% of publications are in German and

Latin. The long and content-revealing nature of the titles which we study in the original

can also be illustrated with examples. A first example is a pamphlet by the little known

Heinrich Spelt, published in 1523, under the title:

Ain ware declaration oder Erklaerung der profession Gelubten und leben

so die gemalten falschen gaystlichenn wider alle Ewangelische freyhayt und

Christliche lyeb thun und wie sy solche halten auch ursach varum sy den

selben gleyssenden Hailigenschein nit vlassen

A rough translation of this title is:

A True Explanation of the Profession, Vows and Life which the Coloured,

False Priests Pursue Against Evangelical Freedom and Christian Love and

how they behave is the reason why they do not leave the same Blinding Halo

A second example is a booklet by the little known Johann Schwann, also published in

1523, under the title:

Ein Sendbriff. Darinne er anzeigt außder Bibel und schryfft, Warumb er

Barfusser orden des er etwan ym Kloster zu Baßell gewest, verlassen

A rough translation of this title is:

An epistle, in Which He Shows from the Bible and Scripture Why He Left

the Barefoot [i.e. mendicant Franciscan] Order in Whose Cloister at Basel

He Formerly Was

51

We provide additional and detailed discussion of how we treat and classify language,

and which figures of speech are most associated with different religious views, in the

classification appendix below.

A.4 Legal Institutions

Our principle source on the laws (Kirchenordnungen) of the Reformation is the multi-

volume collection of Protestant church ordinances Die Evangelischen Kirchenordnungen

Des XVI. Jahrhunderts, originally edited by Emil Sehling. The complete list of vol-

umes we code is as follows. Emil Sehling editor, Volume I Sachsen und Thuringen nebst

angrenzenden Gebieten (1902) (Leipzig: O.R. Reisland). Emil Sehling editor, Volume

II Sachsen und Thuringen nebst angrenzenden Gebieten (1904) (Leipzig: O.R. Reis-

land). Emil Sehling editor, Volume III Brandenburg, Ober- und Niederlausitz, Schle-

sien (1909). Emil Sehling editor, Volume IV Preußen, Polen, Pommern (1911). Emil

Sehling editor, Volume V Baltische Lander, Mecklenburg, Lubeck, Lauenburg, Hamburg

(1913). Institut fur evangelisches Kirchenrecht der EKD, Volume VI/1 Niedersachsen

(1955). Institut fur evangelisches Kirchenrecht der EKD, Volume VI/2 Niedersachsen

(1957). Institut fur evangelisches Kirchenrecht der EKD, Volume VII/1, Niedersach-

sen (1963). Institut fur evangelisches Kirchenrecht der EKD, Volume VII/2, Nieder-

sachsen (1980). Institut fur evangelisches Kirchenrecht der EKD, Volume VIII Hessen

I: Landgrafschaft bis 1582 (1965). Institut fur evangelisches Kirchenrecht der EKD,

Volume XI, Franken, (1961). Institut fur evangelisches Kirchenrecht der EKD, Vol-

ume XII, Schwaben (1963). Institut fur evangelisches Kirchenrecht der EKD, Volume

XIII Altbayern (1966). Institut fur evangelisches Kirchenrecht der EKD, Volume XIV

Kurpfalz (1969). Institut fur evangelisches Kirchenrecht der EKD, Volume XV Baden-

Wurttemberg I: Hohenlohe (1977). Heidelberger Akademie der Wissenschaften, Volume

XVI Baden-Wurttemberg II: Wurttemberg, Baden u. a. (2004). Heidelberger Akademie

der Wissenschaften, Volume XVII/1; XVII/2 Baden-Wurttemberg III/IV: Reichsstadte

(2007/09). Heidelberger Akademie der Wissenschaften, Volume XVIII Rheinland-Pfalz

I: Zweibrucken, Veldenz, Sponheim u. a. (2006). Heidelberger Akademie der Wis-

senschaften, Volume XIX Rheinland-Pfalz II: Wild- und Rheingrafschaft, Leiningen,

Wied u. a. (2008). Heidelberger Akademie der Wissenschaften, Volume XX/1 Elsass

I: Straßburg (2011). Heidelberger Akademie der Wissenschaften, Volume IX Hessen II:

Landgrafschaft ab 1582, Waldeck, Solms, Frankfurt u. a. (2011). Heidelberger Akademie

der Wissenschaften,Volume X Hessen III: Nassau, Hanau-Munzenberg, Ysenburg (2012).

Heidelberger Akademie der Wissenschaften, Volume XX/2 Elsass II: Hanau-Lichtenberg,

Colmar, Mulhausen, Weißenburg u.a. (2013).

52

For cities not covered in the Die Evangelischen Kirchenordnungen Des XVI. Jahrhun-

derts collection, we identify ordinances from additional historical sources. For Zurich:

Campi, Emidio and Philipp Walchli, Zurcher Kirchenordnungen 1520-1675 (Zurich:

Theologischer Verlag Zurich, 2011). For Basel: Campi, Emidio and Philipp Walchli,

Philippe, Basler Kirchenordnungen 1528-1675 (Zurich: Theologischer Verlag Zurich,

2012). For Brno (Brunn): Thomas A. Brady, Heiko Augustinus Oberman, James D.

Tracy, Handbook of European History 1400 - 1600: Late Middle Ages, Renaissance and

Reformation (Leiden: Brill, 1994). Haguenau: Joseph Guerber, Haguenau et la Reforme

(Lyon: Le Roux L.F., 1861). Kostrzyn (Kostschin): Jurgen Sanowsky, “Vorgeschichte

und Anfange der Reformation in der Ballei Brandenburg des Johanniterordens”, in Jo-

hannes Mol et al. eds., Military Orders and the Reformation: Choices, State Build-

ing, and the Weight of Tradition (Amersfoort: Uitgeverij Verloren, 2006). Montbeliard

(Mompelgard): Elise Dermineur, “Rural Communities and the Reformation: the process

of confessionalization in Montbeliard, 1524-1660” (working paper www.academia.edu).

Mulhouse: Peter Blickle, Communal Reformation: The Quest for Salvation in Sixteenth-

Century Germany (Leiden: Brill, 1992). Schaffhausen: Cameron (1991). St. Gallen:

http://www.sg.ch/home/kultur/stiftsarchiv/geschichte.

Our principal source on the constitutional status of cities is the 1521 tax register

(Reichsmatrikel) of the Holy Roman Empire. The Reichsmatrikel lists the cities con-

stitutionally designated as free and imperial cities (Freie und Reichsstadte). For on-line

list see: http://de.wikisource.org/wiki/Reichsmatrikel von 1521 (downloaded December

2012). We use Cantoni (2014) to identify ecclesiastic cities and Wikipedia to identiy a list

of historic prince-bishops of the Holy Roman Empire (https://en.wikipedia.org/wiki/Prince-

bishop, downloaded February 2014).

We locate cities in historic territories as of 1500 using GIS maps from Euratlas (2008),

Periodical Historical Atlas of Europe. These data are: Copyright 2008, Christos Nussli,

Euratlas www.euratlas.com, utilisation licence of 2009. The geography of these territories

is shown in Figure 1 in the main text. In the econometric analysis, we cluster standard

errors at the Euratlas territory (“holder”) level. It should be understood that the Euratlas

territories capture geographic proximity, but are not a direct measure of local institutions.

Complicated and heterogeneous institutional arrangements in some cases applied even

to cities sharing the status as landstadt (i.e. not a free city) within a given territory.

See Whaley (2011) for a review. Table 13 summarizes the distribution of cities across

political units.

53

Table 13: The distribution of cities across principalities and territories

Principality or Territory in 1500 Cities[1] [2]

Archbishopric of Salzburg 1County of East Friesland 1County of Ruppin 1Crown of Bohemia in Personal Union with Kingdom of Hungary 10Duchies of Juelich and Berg 4Duchy of Brunswick-Calenberg 2Duchy of Brunswick-Lueneburg 2Duchy of Brunswick-Wolfenbuettel 6Duchy of Cleves and County of Mark 7Duchy of Lower Bavaria-Landshut 4Duchy of Mecklenburg 3Duchy of Pomerania 6Duchy of Saxony 14Duchy of Upper Bavaria-Munich 3Duchy of Wuerttemberg 6Electoral Palatinate 4Electorate of Brandenburg 6Electorate of Cologne 2Electorate of Mainz 1Electorate of Saxony 14Electorate of Trier 2Habsburg Monarchy 8Kingdom of Denmark 5Kingdom of Poland 2Landgraviate of Hesse 2Lordships of the House of Nassau 1Margravate of Baden 2Monastic State of the Teutonic Knights 2Prince-Bishopric of Passau 1Principality of Anhalt 3Small States of the Holy Roman Empire 57Swiss Confederacy 8Upper Palatinate 1

This table present the territorial classification of cities in the Euratlas data.

A.5 City Characteristics and Institutional Change

The baseline sample comprises 191 cities. We include German-language printing cities

identified in Reske (2007). We also include cities in German-speaking regions of the Holy

Roman Empire for which population in 1500 is recorded in Bairoch et al. (1988). Our

54

baseline data includes Konstanz, Landshut, and Oppenheim which had printers active

1508-1517 but no population recorded in 1500. Our baseline sample also includes cities

such as Aachen and Bremen which both had 18,000 inhabitants in 1500 but no printing as

of 1517. Other notable cities without printing in 1517 included Schwaz, Goslar, Stralsund,

Osnabruck, Hildesheim, Elblag (Elbing), Szczecin (Stettin), Dortmund, and Berlin.

The analysis in the body of the paper examines the relationship between the institu-

tional change outcome and several features of cities, including variations in the structure

of their pre-Reformation media markets. Table 14 presents summary statistics on city

characteristics.

55

Table 14: Summary statistics on city characteristics

All Cities Cities Printing pre-1518No Law pre-1555 Law pre-1555

Mean SD Mean SD Mean SD[1] [2] [3] [4] [5] [6]

Firms 1508-1517 0.60 1.93 1.23 2.21 2.52 3.57Firms 1498-1507 0.59 1.78 1.12 1.97 2.45 3.24Printing pre-1517 0.31 0.46 1.00 0.00 1.00 0.00Religious Share pre-1517 0.13 0.24 0.36 0.22 0.50 0.24Latin Publications pre-1517 0.63 2.86 1.43 4.20 2.51 5.39German Publications pre-1517 0.23 1.04 0.35 0.58 1.08 2.29Indicator: Lord Rule 0.77 0.42 0.85 0.37 0.52 0.51Indicator: Ecclesiastical 0.14 0.35 0.36 0.49 0.07 0.27Indicator: Hansa 0.26 0.44 0.08 0.27 0.33 0.48Indicator: University 0.08 0.27 0.35 0.49 0.18 0.39Indicator: Prince-Bishopric 0.16 0.36 0.38 0.50 0.30 0.47Indicator: River 0.30 0.46 0.42 0.50 0.45 0.51Population 1500: Unrecorded 0.25 0.43 0.31 0.47 0.15 0.36Population 1500: 1-5k 0.41 0.49 0.19 0.4 0.15 0.36Population 1500: 6-10k 0.19 0.40 0.31 0.47 0.24 0.44Population 1500: 11-25k 0.12 0.33 0.15 0.37 0.36 0.49Population 1500: 26k+ 0.02 0.14 0.04 0.20 0.09 0.29Distance to Wittenberg 309.95 160.07 404.02 84.31 298.19 168.71Protestant Publications 63.96 255.02 26.04 52.32 326.42 544.10Reformation Law pre-1555 0.34 0.47 0.00 0.00 1.00 0.00Printer Deaths 0.07 0.33 0.08 0.27 0.33 0.69Observations 191 26 33

This table presents summary statistics on city-level variables. Columns 1 and 2 present summary statis-tics for all cities. Columns 3-6 present summary statistics for cities with any printing pre-1518. Columns3 and 4 present summary statistics for printing cities that did not adopt Reformation law before 1555.Columns 5 and 6 present statistics for printing cities that did adopt Reformation law before 1555. Vari-able definitions are as follows. “Firms 1508-1517” and “Firms 1498-1507” are the measure of printersactive coded from Reske (2007; 2015). “Indicator: Printing per-1517” is an indicator for any printingbefore 1517. “Religious Media pre-1518” is the share of titles (varieties) on religious subjects as classifiedby the USTC. “Latin Publications pre-1517” and “German Publications pre-1517” are the number ofLatin and German publications pre-1517 in units of hundreds. “Indicator: Lord Rule” is an indicatorfor landstadt cities subject to lords (i.e. cities that were not “free”). “Indicator: Ecclesiastical” isan indicator for cities subject to ecclesiastical lords. “Indicator: University” is an indicator for citieswith universities before 1517. “Indicator: Prince-Bishopric” and “Indicator: River” are indicators forthese characteristics. Population variables are indicators that record city population categories derivedfrom Bairoch et al. (1988). “Population 1500: Unrecorded” is an indicator for cities with populationunrecorded. “Population 1500: 1-5k” is an indicator for cities with populations 1,000-5,000. Otherpopulation variables are defined similarly. “Distance to Wittenberg” is in kilometers. “Protestant Pub-lications” is the count of predicted Protestant varieties 1518-1554. “Reformation Law pre-1555” is anindicator for cities adopting Reformation law (kirchenordnung) before 1555. “Printer Deaths” is thenumber of printer deaths.

56

B Identifying Partisan Language

This appendix describes the strategy we use to identify partisan language and index

religious ideas in the media. We describe the estimator and discuss how we select the

vocabulary used to study how language shifts with religion. We then compare the per-

formance of the estimator in our preferred strategy with alternative approaches in the

literature.

B.1 Estimator

Using the subset of titles for which we code whether the author is Protestant or Catholic,

we apply an estimation strategy that first provides a low-dimensional representation of

each text that preserves the religious sentiment and second classifies this low-dimensional

representation according to the religion of the authors. We are then able to find low-

dimensional representations of religious content those texts for which we do not know

the author’s religious affiliation and to predict religion. The estimation framework is the

multinomial inverse regression (MNIR) model introduced by Taddy (2013b).62

Formally, let a document be denoted X i = [xi1,...,iW ] where xiw represents the number

of times phrase w appears in document i for each of the W words in the the vocabulary

V . We are interested in identifying the features that allow us to classify the documents

according to their religious content r with r ∈ P,C for Protestant or Catholic. Since the

distribution of a sum of multinomial draws from the same distribution is multinomial,

we are able to pool the observations into (at least) two classes P and C such that

Xr =∑

i xiwr for r ∈ P,C. Our model for documents is

Xr ∼MN(qr,mr) where

qrw =exp [αw + rϕw]∑Wj=1 exp [αj + rϕj]

for w = 1, . . . ,W, r ∈ P,C(6)

Each Xr is a W−dimensional multinomial variable with size mr =∑

imir with

mir =∑

w xiwr and probabilities qr = [qr1, . . . , qrW ]. The estimated factor loadings ϕ

allow us to compute a sufficient reduction (SR) score zi for document word frequencies

62The estimation strategy follows the literature in high-dimensional estimation and machine learning,assumes that the order of phrases or words within a document is relatively unimportant in classifying thecontent of the text, and views documents as a bag of words (Salton and McGill 1986). This assumptionallows documents to be treated as as multinomial random variables in which the phrases or words are thecategories and the support is called the vocabulary. We thus increase efficiency by making an assumptionabout the functional relationship between text and sentiment (Taddy 2013c).

57

f i = xi

mi. The identifying assumption is that the sufficient reduction score zi is a scalar

containing all relevant sentiment information in document i independent of the full X i

and its length.

zi = ϕ′f i ⇒ ri ⊥⊥X i,mi | zi (7)

This assumption allows us to ignore the full-dimensional X i and model the classification

problem as a univariate regression problem:

Pr(ri = r | zi) =1

1 + exp [β0 + β1zi](8)

To estimate the religious content of texts, we take the projection of the factor loadings

estimated in (6) onto the frequencies of these out-of-sample texts to obtain SR scores zi.

We then use the coefficients from (8) to infer the religious content of the texts.

To address possible implications of fat-tailed and sparse distributions, independent

Laplace priors with unknown variance are placed on the factor loadings ϕw. The un-

known rate parameter λw accounts for our uncertainty as to how much variable-specific

regularization is appropriate and is given a gamma hyperprior Γ(α, β) such that:

Pr(ϕw, λw) =λw2e−λw|ϕw| β

α

Γ(α)λα−1w e−βλw , α, β, λw > 0. (9)

Estimation of the likelihood implied by the multinomial distribution in (6) and the prior

(9) takes place via the gamma-lasso algorithm to maximize the joint posterior over coef-

ficients and their prior scale (Taddy 2013b;a).

B.2 Vocabulary Selection

To prepare the titles for use in the classifier described in Section B.1, it is desirable to

identify first the words which will serve as the support for the multinomial distribution.

This is the vocabulary.

We construct the vocabulary as follows. First all generic titles are removed.63 We

clean the remaining text by making the titles lower-case and removing punctuation, num-

bers and roman numerals, and words that occur fewer than three times. Futhermore,

we experiment with both removing German and Latin stop words and preserving them

63Titles that do not contain words illuminating what the works are about such as Theil, an alternatespelling of the German teil meaning part of a whole, and Samml, an abbreviation of the German forcollection, are removed.

58

Table 15: Distribution of Known Catholic Title Lengths

Total Words Words in Estimatingincluding Stop Words Vocabulary

5th Percentile 7.0 3.025th Percentile 15.0 7.0Median 23.0 12.075th Percentile 33.0 18.095th Percentile 49.0 30.0Mean 25.1 13.7

in estimation.64 We do not stem words. Stemming involves using an algorithm to re-

duce words to a common root. It is common practice in text analysis to stem words.

However, we choose not to stem words for several reasons. First, orthographic variation

may in some instances reflect ideological differences in authors’ religions that we wish

to preserve. Second, the lack of orthographic standardization significantly complicates

stemming. Third, we demonstrate meaningful results without stemming. We discuss this

again in Section B.6.

For the computation of counts used to construct measures of denominational language,

we temporarily drop titles that may be reprints and would therefore skew the usage of

certain words simply by the number of times a given title appeared in print.

The impact of removing stop words on the title length distributions can be seen in

the second column of table (12). As can be seen in tables (15) and (16) the sample-wide

distribution is indicative of the sub-populations of Protestants and Catholics.

In selecting the most relevant features of the language for classifying phrases as Protes-

tant or Catholic, we examine three strategies. First, we use the χ2 measure used by

Gentzkow and Shapiro (2010) to select features. Second, we use the log-odds ratio pro-

posed by Monroe et al. (2008). Third, we require no prior selection of the features, relying

only on the model-based feature selection via regularization inherent in the estimator.

For each strategy, we implement two variants: one in which stop words are removed and

one in which stop words are preserved.65

To compare the performance of these six strategies, we perform out-of-sample cross

64German stop words include words such as da and dass, der and die, and und and unde. These arewords for this, the, and and in both modern and Old High German.

65The intuition here is that stop words may contain relevant information for classification. Considerthe distinction between the phrase “the Church” and a phrase like “our church.”

59

Table 16: Distribution of Known Protestant Title Lengths

Total Words Words in Estimatingincluding Stop Words Vocabulary

5th Percentile 4.0 1.025th Percentile 10.0 4.0Median 17.0 8.075th Percentile 26.0 12.095th Percentile 42.0 20.0Mean 19.1 8.8

validation. To implement the out-of-sample cross-validation, we train the model on 80%

of the data and test it on a held-out 20%.66 As discussed below, we find that the log-odds

ratio performed best when retaining stop words and use this as our baseline strategy, as

discussed below.

B.3 The χ2 Measure

The χ2 measure used in Gentzkow and Shapiro (2010) can be written as:

χ2wn =

(fwncf∼wnp − fwnpf∼wnc)2

(fwnc + fwnp) (fwnc + f∼wnc) (fwnp + f∼wnp) (f∼wnc + f∼wnp)(10)

where fwnc and fwnp denotes the total number of times phrase w of length n is used

in a title by a Catholic or Protestant writer, respectively. Whereas, f∼wnc and f∼wnp

denotes the total number of times a length n phrase that is not w is used by Catholics

and Protestants, respectively.

There are two challenges associated with the χ2 measure. The first challenge is

that there is nothing to differentiate “Catholic” phrases from “Protestant” phrases.

The χ2 measure classes any phrase that has more counts in “Catholic” documents as

“Catholic,” and vice versa. Figures 11-16 show the results of applying the Gentzkow-

Shapiro χ2measure to the texts. These figures illustrate how Catholic words seem to

be much more distinguished by the χ2 measure than the language used by Protestants.

The terms do conform to our expectations about what distinguishes Catholic terms from

Protestant terms. However, in Figures 11-16, the χ2 values for the Catholics have been

negated. This potential shortcoming is further discussed in Monroe et al. (2008).

66Simulations training the estimator on 65% and 90% deliver virtually identical results.

60

100 101 102 103 104

Frequency of Words

0.004

0.003

0.002

0.001

0.000

0.001

0.002

0.003

0.004

χ2 w

doctore

doctore

pauli

pauli

professore

professore

academia

academia

jacobo

jacobo

praeside

praeside

et

et

commentarii

commentarii

epistolam

epistolam

theologiae

theologiae

heerbrando

heerbrando

coena

coena

doctrinae

doctrinae

tubingensi

tubingensi

hora

hora

davidis

davidis

libri

libri

propositionespropositiones

conabitur conabituratqueatque

catholicae

catholicae

adversus

adversus

catholica

catholica

emser

emser

eckii

eckii

sanctorum

sanctorum

ordinis

ordinis

eundem

eundem

nauseae

nauseae

beroaldi

beroaldi

sacris

sacris

epigramma

epigramma

haereticos

haereticos

episcopi

episcopi

virginis

virginis

ferum

ferum

pascha

pascha

feriferi

metropolitanaemetropolitanaelutheranos

lutheranos

Gentzkow-Shapiro - Latin

Figure 11: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. Latin unigrams only.

Table 17: The number of phrases left after applying the feature selection method ofGentzkow and Shapiro. Stop words are not removed.

Language unigrams bigrams trigramsProtestant Catholic Protestant Catholic Protestant Catholic

German 2000 1975 2000 2000 2000 2000Latin 2000 1011 2000 2000 2000 2000

The second challenge with the χ2 metric is that there is no natural significance cut-off

for selecting partisan phrases. The scale of the values obtained from (10) do not appear

to follow from any parameterized χ2 distribution. Gentzkow and Shapiro (2010) select

the top 2,000 each of unigrams, bigrams, and trigrams. We follow the same heuristic

with one difference. After obtaining the measure for each phrase for both Protestants

and Catholics, we sort each list of the n-length phrases for each group from largest to

smallest and select out the smaller of 2,000 or the number of total phrases of length n

from each groups to obtain our vocabulary. We do this for both Latin texts and German

texts separately so that we do not bias the results due to choice of publication language.

Tables 17 and 18 presents the results of this selection.

61

100 101 102 103

Frequency of Words

0.003

0.002

0.001

0.000

0.001

0.002

0.003

χ2 w

in academia

in academia

theologiae doctore

theologiae doctore

disputatio de

disputatio de

praeside jacobo

praeside jacobo

jacobo heerbrando

jacobo heerbrando

pauli ad

pauli ad

de coena

de coena

et professore

et professore

doctore et

doctore et

coena domini

coena domini

academia tubingensi

academia tubingensi

in epistolam

in epistolam

cum praefatione

cum praefatione

jacobo andreae

jacobo andreae

martini lutheri

martini lutheri

praeceptore suo

praeceptore suo

davidis chytraei

davidis chytraei

respondere conabiturrespondere conabitur

de coniugio de coniugiode quibus

de quibus

ad eundem

ad eundem

philippi beroaldi

philippi beroaldi

communium adversus

communium adversus

enchiridion locorum

enchiridion locorum

catholicae ecclesiae

catholicae ecclesiae

de tempore

de tempore

joannis feri

joannis feri

friderici nauseae

friderici nauseae

nauseae blancicampiani

nauseae blancicampiani

per joannem

per joannem

joannem ferum

joannem ferum

societatis jesu

societatis jesu

catholicae fidei

catholicae fidei

de sanctis

de sanctis

ex sacris

ex sacris

epistolas evangeliaepistolas evangelia

examen ordinandorum examen ordinandorumde usingende usingen

surium carthusianum

surium carthusianum

per laurentium

per laurentium

Gentzkow-Shapiro - Latin

Figure 12: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. Latin bigrams only.

Table 18: The number of phrases left after applying the feature selection method ofGentzkow and Shapiro. Stop words are removed

Language unigrams bigrams trigramsProtestant Catholic Protestant Catholic Protestant Catholic

German 2000 1932 2000 2000 2000 2000Latin 2000 982 2000 1997 2000 1719

B.4 Log-Odds Ratio with Informative Prior

Monroe et al. (2008) defines several variants of a log-odds ratios to use for feature selection

in classification problems. We present here the results from using the so-called log-odds

ratio with an informative Dirichlet prior.

The general idea is that when using the multinomial-based model described in (6),

we can compute the parameter vector q for the multinomial in a Bayesian framework.

Using the conjugate Dirichlet prior for q, enables us to obtain the log-odds ratio for each

phrase. By virtue of having a probabilistic model, we are able to compute the variance

for each log-odds ratio and, in turn, a Z-score. This gives us a natural way to determine

which phrases help us separate the models.

To implement the log-odds ratio, we start from the idea that the quantity of interest

is the observed “odds” of word w being used by either Protestants or Catholics. We

62

100 101 102

Frequency of Words

0.003

0.002

0.001

0.000

0.001

0.002

0.003

χ2 w

doctore et professore

doctore et professore

praeside jacobo heerbrando

praeside jacobo heerbrando

de coena domini

de coena domini

in academia tubingensi

in academia tubingensi

in epistolam pauli

in epistolam pauli

sacrosanctae theologiae doctore

sacrosanctae theologiae doctore

praeside jacobo andreae

praeside jacobo andreae

theologiae doctore et

theologiae doctore et

jacobo heerbrando doctore

jacobo heerbrando doctore

theologiae doctore professore

theologiae doctore professore

in aula nova

in aula nova

epistolam pauli ad

epistolam pauli ad

sacrae theologiae doctore

sacrae theologiae doctore

professore in academia

professore in academia

et professore in

et professore in

heerbrando doctore et

heerbrando doctore et

academiae tubingensis praeposito

academiae tubingensis praeposito

locos doctrinae christianae

locos doctrinae christianae

theologiae in academia theologiae in academiachristophori pezelii sacraechristophori pezelii sacrae

locorum communium adversus

locorum communium adversus

enchiridion locorum communium

enchiridion locorum communium

friderici nauseae blancicampiani

friderici nauseae blancicampiani

summa doctrinae christianae

summa doctrinae christianae

per laurentium surium

per laurentium surium

laurentium surium carthusianum

laurentium surium carthusianum

fe re ad

fe re ad

epistola tetrastichon ad

epistola tetrastichon ad

beroaldi fe re

beroaldi fe re

philippi beroaldi fe

philippi beroaldi fe

tetrastichon ad eundem

tetrastichon ad eundem

ducem epistola tetrastichon

ducem epistola tetrastichon

philippi beroaldi ab

philippi beroaldi ab

epigramma aliud ex

epigramma aliud ex

ex sententia musoniiex sententia musonii

libello continentur ad libello continentur adusque ad adventum usque ad adventumper joannem viaper joannem via

sententia musonii philosophi

sententia musonii philosophi

puerorum epistolaria progymnasmata

puerorum epistolaria progymnasmata

Gentzkow-Shapiro - Latin

Figure 13: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. Latin trigrams only.

define this as:

Orw =frw

1− frw(11)

The odds ratio between Protestants and Catholics is θ(P−C)w = OPw

OCw. It is common to

work with the log-odds ratio on the grounds of symmetry:

log θ(P−C)w = logOPw − logOCw (12)

The naive log-odds ratio has the undesirable property of the results being determined

by the sampling variation. The variance decreases as words are used more frequently

so that the measure promotes obscure words as helping to classify the religious denom-

ination. Furthermore, it is unclear what to do when a word is used only by one group.

In this case, the naive log-odds ratio is not well defined. We choose to overcome these

problems through one of the model-based approaches of Monroe et al. (2008).

Introducing the assumption that the documents are multinomial random variables,

we have for each group that the documents:

Xr ∼ MN(qr,mr) (13)

Omitting the subscripts for the denomination and following the notation introduced

in Section B.1, the maximum likelihood solution for the probabilities q is simply the

63

100 101 102 103 104

Frequency of Words

0.004

0.003

0.002

0.001

0.000

0.001

0.002

0.003

0.004

χ2 w

warheit

warheit

alten

alten

biß

biß

oder

oder

sermon

sermon

kirchen

kirchen

predigen

predigen

unde

unde

gelesen

gelesen

christi

christi

wirdt

wirdt

unnd

unnd

heiligen

heiligen

eine

eine

zeit

zeit

auß

auß

priester

priester

ob

ob

vanvan

lutherischen lutherischen

catholischen

catholischen

catholischer

catholischer

evangelischer

evangelischer

catholische

catholische

postill

postill

catholisch

catholisch

predigbuchs

predigbuchs

altars

altars

rechter

rechter

bern

bern

ertzdhomstifft

ertzdhomstifft

secten

secten

löblichen

löblichen

sommertheils

sommertheils

wicelii

wicelii

ecclesiae

ecclesiae

gezieret

gezieret

sontaeglichensontaeglichen

alter altersanctissanctis

Gentzkow-Shapiro - German

Figure 14: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. German unigrams only.

frequencies of the words:

qMLE = f (14)

This frequency-based solution, however, tends to overemphasize high-frequency words

with high-sampling variability as belonging to one group or another when their semantic

information in isolation is highly questionable. A natural way to alleviate this poten-

tially problematic feature of the maximum likelihood solution is by adopting a Bayesian

approach. In the Bayesian approach, we choose a sensible prior for q and obtain more

salient results. Due to its conjugacy with the multinomial distribution, this prior will be

the Dirichlet:

q ∼ Dirichlet(α) (15)

where for the Dirichlet, α is a vector of size W with αi > 0∀i = 1, . . . ,W . This prior has

a tidy interpretation in terms of the solution of the model. Each term in the data is now

treated as if it was observed another αw − 1 times. Taking advantage of the conjugacy

of the Dirichlet to the multinomial, and defining α0 =∑

w αw, the full Bayesian solution

to the model is:

q =x+α

m+ α0

(16)

We use these point-estimates to interpret the log-odds ratio in a probabilistic setting.

64

100 101 102 103 104

Frequency of Words

0.004

0.003

0.002

0.001

0.000

0.001

0.002

0.003

0.004

χ2 w

biß auff

biß auff

gelesen werden

gelesen werden

der heiligen

der heiligen

auß den

auß den

ist ein

ist ein

auff die

auff die

sermon von

sermon von

eyn sermon

eyn sermon

der hailigen

der hailigen

wie er

wie er

auch wie

auch wie

auch von

auch von

ob der

ob der

diser zeit

diser zeit

so von

so von

christlich und

christlich und

waren christlichen

waren christlichen

der zeit

der zeit

das istdas ist

der alten der alten

evangelischer warheit

evangelischer warheit

rechter catholischer

rechter catholischer

evangelia so

evangelia so

heiligen gottes

heiligen gottes

predigbuchs evangelischer

predigbuchs evangelischer

oder predigbuchs

oder predigbuchs

postill oder

postill oder

der catholischen

der catholischen

des altars

des altars

lehr uber

lehr uber

zu bern

zu bern

catholischen kirchen

catholischen kirchen

der postill

der postill

catholischen glaubens

catholischen glaubens

ertzdhomstifft zu

ertzdhomstifft zu

catholischer lehr

catholischer lehr

und rechterund rechter

so vom so vomdie festdie fest

lieben heiligen

lieben heiligen

Gentzkow-Shapiro - German

Figure 15: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. German bigrams only.

First we obtain the log-odds ratio by plugging (16) in to (12):

δ(P−C)w = log[

xPw + αrwmP + αP0 − xPw − αPw

]− log[xCw + αCw

mC + αC0 − xCw − αCCw] (17)

Next we obtain the approximate variance for these point estimates:

σ2(δ(P−C)w

)=

1

xPw + αPw+

1

mP + αP0 − xPw − αPw+

1

xCw + αCw+

1

mC + αC0 − xCw − αCw

(18)

This approximation assumes that the observed counts are much larger than the cor-

responding terms in the prior. Using the point-estimate and its variance we can compute

z-scores of the log-odds ratios:

z(P−C)k =

δ(P−C)w√

σ2(δ(P−C)w )

(19)

For the results presented here, we choose Z < Φ−1(.05) ≈ −1.64 and Z > Φ−1(.05) ≈ 1.64

to be the thresholds for inclusion.

Finally, we must choose an appropriate prior for our terms. Monroe et al. (2008)

makes several suggestions from a fully uninformative prior where αw = .01 for each term

to a Laplace prior which acts as an L1 regularization penalty and shrinks estimates to

65

100 101 102 103

Frequency of Words

0.004

0.003

0.002

0.001

0.000

0.001

0.002

0.003

0.004

χ2 w

predig von den

predig von den

das gantz jar

das gantz jar

so in der

so in der

von den heiligen

von den heiligen

ein christliche predig

ein christliche predig

evangelien von den

evangelien von den

oder außlegung der

oder außlegung der

von der beicht

von der beicht

auff ostern in

auff ostern in

ostern in der

ostern in der

jesu christi unsers

jesu christi unsers

sermon von dem

sermon von dem

postilla das ist

postilla das ist

ersten sontag nach

ersten sontag nach

eyn sermon von

eyn sermon von

vom advent an

vom advent an

uber das gantze

uber das gantze

das ist christliche

das ist christliche

sermon von der

sermon von der

durch das gantz durch das gantz

oder predigbuchs evangelischer

oder predigbuchs evangelischer

predigbuchs evangelischer warheit

predigbuchs evangelischer warheit

postill oder predigbuchs

postill oder predigbuchs

rechter catholischer lehr

rechter catholischer lehr

lehr uber die

lehr uber die

sacrament des altars

sacrament des altars

sontags evangelia so

sontags evangelia so

und rechter catholischer

und rechter catholischer

catholischer lehr uber

catholischer lehr uber

lieben heiligen gottes

lieben heiligen gottes

die fest der

die fest der

auff die fest

auff die fest

der catholischen kirchen

der catholischen kirchen

der kurtzen postill

der kurtzen postill

uber alle epistelnuber alle episteln

evangelia so vom evangelia so vombegreiffend die predigenbegreiffend die predigen

theil begreiffend die

theil begreiffend die

evangelischer warheit das

evangelischer warheit das

nach rechter catholischer

nach rechter catholischer

Gentzkow-Shapiro - German

Figure 16: Gentzkow-Shapiro measure vs. Frequency of words in log scale. The top 20words for Protestants and Catholics are listed. German trigrams only.

zero. We take a middle ground here since we will be doing subsequent modeling. We

choose an informative Dirichlet prior. We use what we know about the expected usage of

words in the non-religious texts to give us a sense of what words distinguish Protestant

and Catholic language. Our prior is:

αw = α0qMLE = fα0 (20)

where the frequencies f come from the non-religious texts. To avoid taking the log of

zero, for any word that is in the support for the religious texts but only appears in either

Protestant or Catholic texts and does not appear elsewhere in the non-religious titles,

we assign a prior count of 1 before normalization. That is, we assume this term to be

very rare but avoid taking the log of zero. α0 determines how strong the prior is and,

therefore, how much shrinkage occurs. Given that we observe many more counts for

Protestant authors than Catholic authors, some care needs to be taken in choosing α0.

The estimation uses as a base value 15% of the total number of n-grams for each n-

gram for both Catholics and Protestants. This is equivalent to saying that for each n-gram

for each denomination, we observe a sample that is 15% larger in which the frequencies

of the language used is that of non-religious texts. This approach helps us separate out

those features of the language that are particular to religious texts. Furthermore, by

doing this for both German and Latin texts the analysis ensures that each prior used is

conditional on the relative number of words for Protestants and Catholics and the choice

of language. This approach guards against bias due to having relatively larger Protestant

66

Table 19: The number of phrases left after using the log-odds ratio with an informativeprior for feature selection. Stop words are not removed

Language unigrams bigrams trigramsProtestant Catholic Protestant Catholic Protestant Catholic

German 277 785 144 403 41 91Latin 169 1013 68 1325 12 674

Table 20: The number of phrases left after using the log-odds ratio with an informativeprior for feature selection. Stop words are removed

Language unigrams bigrams trigramsProtestant Catholic Protestant Catholic Protestant Catholic

German 259 710 69 165 21 36Latin 131 947 13 627 2 290

author output and/or the fact that Catholic author’s tend to produce more Latin text

than German text.

The results of applying this metric can be seen in Tables 19 and 20. The words

chosen by each measure can be seen in Figures 17-22. These results strongly suggest that

the log-odds measure is more judicious than the χ2 alternative. To assess this question

formally we study predictive performance in the next subsection.

B.5 Performance

Figures 23, 24, and 25 provide common performance benchmarks for classification algo-

rithims for the experiments described above. The measure contained in each figure is the

F1-measure. The Fβ−measure is defined

Fβ = (1 + β2)precision · recall

(β2 · precision + recall)(21)

where for the F1−measures β = 1, giving the harmonic mean between precision and

recall.67 Precision is analagous to the absence of Type I errors and is defined as

precision =true positive

true positive + false positive(22)

67Higher values of β may be used if one wants to weight recall more than precision.

67

100 101 102 103 104

Frequency of words

15

10

5

0

5

10

15

ln(O

(P)

w/O

(C)

w)

doctore

doctore

pauli

pauli

et

et

professore

professore

theologiae

theologiae

jacobo

jacobo

academia

academia

doctrinae

doctrinae

praeside

praeside

epistolam

epistolam

coena

coena

davidis

davidis

hora

hora

propositiones propositionestubingensi

tubingensipars

pars

commentarii

commentarii

andreae

andreae

continens

continens

annotationes

annotationes

adversus

adversus

catholicae

catholicae

catholica

catholica

sanctorum

sanctorum

eckii

eckii

sacris

sacris

eundem

eundem

ordinis

ordinis

episcopi

episcopi

joannem

joannem

super

super

sanctis

sanctis

cochlaei

cochlaei

concionatorem concionatoremperper

utriusque

utriusque

libri

libri

ss

ss

evangelicae

evangelicae

fidei

fidei

Log-Odds Ratio - Informative Dirichlet Prior - Latin

Figure 17: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. Latin unigrams only.

while recall is analagous to the absence of Type II errors and is defined as

recall =true positive

true postitive + false negative(23)

The values for (22) and (23) come from the confusion matrix (contingency table)

where

Predicted Class0 1

Actual 0 true negative false positiveClass 1 false negative true positive

In terms of performance, we prefer the log odds measure. While the Gentzkow-

Shapiro measure has good performance on average, this is certainly the regularization

performance of the estimator. The GS measure simply gives us more words to work with.

We have the potential for a better estimate, but we also see some poor performance in

the outliers. This suggests overfitting. The log-odds measure, on the other hand, leads

to good average estimates and small variance. The model with no vocabulary selection

does not perform as well on averge due to overfitting.

68

100 101 102 103

Frequency of words

8

6

4

2

0

2

4

6

8

ln(O

(P)

w/O

(C)

w)

in academia

in academia

theologiae doctore

theologiae doctore

disputatio de

disputatio de

pauli ad

pauli ad

praeside jacobo

praeside jacobo

martini lutheri

martini lutheri

doctore et

doctore et

et professore

et professore

jacobo heerbrando

jacobo heerbrando

in epistolamin epistolam

coena domini coena dominicum praefationecum praefatione

de coena

de coena

academia tubingensi

academia tubingensi

praeceptore suo

praeceptore suo

respondere conabitur

respondere conabitur

de quibus

de quibus

oratio de

oratio de

ad romanos

ad romanos

epistolam pauli

epistolam pauli

ad eundem

ad eundem

per joannem

per joannem

de tempore

de tempore

de sanctis

de sanctis

ex sacris

ex sacris

catholicae ecclesiae

catholicae ecclesiae

evangelia quae

evangelia quae

locorum communium

locorum communium

totius anni

totius anni

societatis jesusocietatis jesu

johannis eckii johannis eckiiecclesiae catholicaeecclesiae catholicae

ss theologiae

ss theologiae

summa doctrinae

summa doctrinae

philippi beroaldi

philippi beroaldi

ecclesia catholica

ecclesia catholica

sacris literis

sacris literis

johannis cochlaei

johannis cochlaei

epistolas evangelia

epistolas evangelia

de observantia

de observantia

Log-Odds Ratio - Informative Dirichlet Prior - Latin

Figure 18: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. Latin bigrams only.

B.6 Heterogeneity of Language

One possible question is whether the language used shifts over the time period we

study. This could happen, for example, if the language of the Reformation and Counter-

Reformation changed as the agendas of the camps changed or if spelling norms were

evolving.

Figures 26-34 provide heatmaps of word occurence to document whether and how

religious language evolved. The words are first sorted according to the first time they

appear in use in a title. If the language was not evolving, we would expect the column

associated with the first period to be large and the subsequent row to be randomly

populated. While this is not the case, if the words were truly shifting, we would see a

banded structure. Instead what we observe is simply an expanding vocabulary. In the

first 10 years about half of all unigrams in the whole sample are used once and continue

to be used.

Classifying the nature of this change of language use is the subject of ongoing re-

search. However, our sampling in our cross-validation experiments in Sections B.3 and

B.4 is undertaken without respect to time. Our results indicate that we are still able

to predict well despite the non-standard orthographies of German and Latin in the six-

teenth century. Therefore, while we do think there are interesting findings to be made in

studying the change of language, the results of the present study are not likely to depend

69

100 101 102

Frequency of words

4

2

0

2

4

ln(O

(P)

w/O

(C)

w)

doctore et professore

doctore et professore

praeside jacobo heerbrando

praeside jacobo heerbrando

in academia tubingensi

in academia tubingensi

de coena domini

de coena domini

in epistolam pauli

in epistolam pauli

sacrosanctae theologiae doctore

sacrosanctae theologiae doctore

praeside jacobo andreae

praeside jacobo andreae

epistolam pauli adepistolam pauli ad

in aula novain aula nova

theologiae doctore ettheologiae doctore et

jacobo heerbrando doctore jacobo heerbrando doctoretheologiae doctore professore

theologiae doctore professoresacrae theologiae doctore

sacrae theologiae doctore

heerbrando doctore et

heerbrando doctore et

academiae tubingensis praeposito

academiae tubingensis praeposito

theologiae in academia

theologiae in academia

in evangelia dominicalia

in evangelia dominicalia

pauli ad romanos

pauli ad romanos

theologiae doctore ac

theologiae doctore ac

heerbrando theologiae doctore

heerbrando theologiae doctore

ex sacris literis

ex sacris literis

per laurentium surium

per laurentium surium

laurentium surium carthusianum

laurentium surium carthusianum

hoc libello continentur

hoc libello continentur

usque ad adventum

usque ad adventum

per totum annumper totum annum

in ecclesia christi in ecclesia christiepistolas evangelia quae epistolas evangelia quaeex sacris scripturis

ex sacris scripturispartim etiam ex

partim etiam ex

ex tomis aloysii

ex tomis aloysii

optima fide collectis

optima fide collectis

etiam ex egregiis

etiam ex egregiis

egregiis manuscriptis codicibus

egregiis manuscriptis codicibus

manuscriptis codicibus quarum

manuscriptis codicibus quarum

aloysii lipomani doctissimi

aloysii lipomani doctissimi

complectens sanctos mensium

complectens sanctos mensium

partim ex tomis

partim ex tomis

quarum permultae antehàc

quarum permultae antehàc

ex egregiis manuscriptis

ex egregiis manuscriptis

Log-Odds Ratio - Informative Dirichlet Prior - Latin

Figure 19: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. Latin trigrams only.

on these issues.

70

100 101 102 103 104

Frequency of words

15

10

5

0

5

10

15

ln(O

(P)

w/O

(C)

w)

sermon

sermon

christi

christi

unde

unde

eine

eine

am

am

herrn

herrn

fuer

fuer

christen

christen

de

de

van

van

auslegung

auslegung

dat

dat

sol

sol

lutheri

lutheri

predigten predigtenepistelepistel

psalm

psalm

jhesu

jhesu

unterricht

unterricht

jungen

jungen

catholischen

catholischen

evangelischer

evangelischer

postill

postill

catholische

catholische

catholischer

catholischer

warheit

warheit

altars

altars

rechter

rechter

catholisch

catholisch

altenalten

secten sectenoderoder

kirchen

kirchen

biß

biß

predigen

predigen

gelesen

gelesen

bern

bern

heiligen

heiligen

hohen

hohen

wirdt

wirdt

Log-Odds Ratio - Informative Dirichlet Prior - German

Figure 20: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. German unigrams only.

100 101 102 103 104

Frequency of words

10

5

0

5

10

ln(O

(P)

w/O

(C)

w)

sermon von

sermon von

ain sermon

ain sermon

jhesu christi

jhesu christi

fuer die

fuer die

martini lutheri

martini lutheri

ein sermon

ein sermon

an die

an die

des herrn

des herrn

auslegung der

auslegung der

der episteln

der episteln

aus dem

aus dem

die jungen

die jungen

der evangelien

der evangelien

vor de

vor de

zu sachsenzu sachsen

des evangelii des evangeliijungen christen

jungen christen

epistel pauli

epistel pauli

gottes wort

gottes wort

wie man

wie man

des altars

des altars

postill oder

postill oder

heiligen gottes

heiligen gottes

gelesen werden

gelesen werden

der heiligen

der heiligen

der catholischen

der catholischen

und rechter

und rechter

biß auffbiß auff

so vom so vomder kurtzen der kurtzenan biß

an bißsontags evangelia

sontags evangelia

der postill

der postill

zu bern

zu bern

auß den

auß den

der lieben

der lieben

christlichen und

christlichen und

sommertheil der

sommertheil der

ein gar

ein gar

auff die

auff die

Log-Odds Ratio - Informative Dirichlet Prior - German

Figure 21: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. German bigrams only.

71

100 101 102 103

Frequency of words

10

5

0

5

10

ln(O

(P)

w/O

(C)

w)

sermon von dem

sermon von dem

sermon von der

sermon von der

in der kirchen

in der kirchen

ain sermon von

ain sermon von

der episteln und

der episteln und

die jungen christen

die jungen christen

des leibs und

des leibs und

bis auff ostern

bis auff ostern

herrn jhesu christi

herrn jhesu christi

fuer die jungen

fuer die jungen

aus gottes wort

aus gottes wort

der epistel pauli

der epistel pauli

unsers herrn jhesu

unsers herrn jhesu

des ehrwirdigen herrn

des ehrwirdigen herrn

pauli an die

pauli an die

der kirchen gelesen

der kirchen gelesen

leibs und blutsleibs und bluts

ein sermon von ein sermon vongantze jar mit

gantze jar mitwie die in

wie die in

sacrament des altars

sacrament des altars

die sontags evangelia

die sontags evangelia

biß auff ostern

biß auff ostern

an biß auff

an biß auff

auff die fest

auff die fest

biß auff den

biß auff den

unser lieben frawen

unser lieben frawen

von den vier

von den vier

das ist ein

das ist ein

so von ostern

so von ostern

mit sonderm fleißmit sonderm fleiß

auff den advent auff den adventpredigen uber diepredigen uber die

wie er von

wie er von

auch von den

auch von den

er von der

er von der

zu halten sey

zu halten sey

nach ordnung der

nach ordnung der

auch wie er

auch wie er

predig von den

predig von den

Log-Odds Ratio - Informative Dirichlet Prior - German

Figure 22: Log-odds ratio with informative Dirichlet prior vs. Frequency of words in logscale. The top 20 words for Protestants and Catholics are listed. German trigrams only.

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

0.0

0.2

0.4

0.6

0.8

1.0

F-M

easu

re

Includes stop words

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

Stop words removedF-measure

Figure 23: F1-score for 80:20 train:test split of data using Log-odds measure.

72

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

0.0

0.2

0.4

0.6

0.8

1.0

F-M

easu

re

Includes stop words

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

Stop words removedF-measure

Figure 24: F1-score for 80:20 train:test split of data using Gentzkow-Shapiro χ2 measure.

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

0.0

0.2

0.4

0.6

0.8

1.0

F-M

easu

re

Includes stop words

s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape

Stop words removedF-measure

Figure 25: F1-score for 80:20 train:test split of data with no feature selection.

73

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f W

ord

in V

oca

bula

ry

All years - All language

Figure 26: Heat map of word usage for each year in any language. Unigrams only.

74

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f W

ord

in V

oca

bula

ry

All years - German

Figure 27: Heat map of word usage for each year in German. Unigrams only.

75

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f W

ord

in V

oca

bula

ry

All years - Latin

Figure 28: Heat map of word usage for each year in Latin. Unigrams only.

76

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f B

igra

m in V

oca

bula

ry

All years - All language

Figure 29: Heat map of word usage for each year in any language. Bigrams only.

77

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f B

igra

m in V

oca

bula

ry

All years - German

Figure 30: Heat map of word usage for each year in German. Bigrams only.

78

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f B

igra

m in V

oca

bula

ry

All years - Latin

Figure 31: Heat map of word usage for each year in Latin. Bigrams only.

79

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f Tri

gra

m in V

oca

bula

ry

All years - All language

Figure 32: Heat map of word usage for each year in any language. Trigrams only.

80

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f Tri

gra

m in V

oca

bula

ry

All years - German

Figure 33: Heat map of word usage for each year in German. Trigrams only.

81

1517 1527 1537 1547 1557 1567 1577 1587 1597

Index o

f Tri

gra

m in V

oca

bula

ry

All years - Latin

Figure 34: Heat map of word usage for each year in Latin. Trigrams only.

82

C Market Prices

In the early 1520s, a typical pamphlet of 32 pages cost approximately 1/3 of the daily

wage for an artisan, equivalent to the price of a hen or 1 kg of beef (Edwards 1994).

A typical book cost under 1 day of worker’s wages (Dittmar 2015). Table 21 presents

transaction prices for a number of Protestant publications in our data.

Purchase Price in Price inAuthor Title Year Pfennig Daily Wage

(1) (2) (3) (4) (5)

Staupitz, Johann von Ein nutzbarliches büchlein von der entlichen volziehung ewiger fuersehung

1517 8 0.33

Luther, Martin Eyn sendbrieff an den Bapst Leo. den czehenden 1521 3 0.13

Wittenberg - Rat Newe ordnung der stat Wittenberg. M.D.XXII. jar. Des newen Bischoffs zu der Lochaw verhor und disputation vor dem Bischof von Meissen

1522 3 0.13

Anonymous Ernstlich handlung der Universitet zu Wittenberg an den durchleüchtigsten hochgebornen Churfürsten und Herren. Hertzug Friderich von Sachsen, die Mesz betreffendt

1522 4 0.17

Hutten, Ulrich von Ein clagschrift des hochberuemten und eernuesten Herrn Ulrichs von hutten gekroeneten poeten und orator an alle Stend Deütscher nation

1520 3 0.13

Luther, Martin Warum des Papstes und seiner Jünger Bücher verbrannt

1521 3 0.13

Rhegius, Urbanus Anzaygung dasz die Romisch bull mercklichen schaden in gewissin Manicher menschen gebracht hab und nit Doctor Luthers leer

1522 8 0.33

Luther, Martin Von den newen eckischen bullen und luegen 1521 3 0.13

Bugenhagen, Johann Ein sendbrief uber eyne frage vom Sacrament. Item eyn unterricht von der beycht und Christlichen absolution

1525 1 0.04

Sickingen, Franz von Ein sendbrieff so der edel und ernuest Franciscus von Sickingen seinem schweher dem edlen unnd ernuesten iuncker diethern von henschuchßheym zu underrichtung etzlicher artickel Christliches glaubens kürtzlingen zugeschickt hadt.

1522 3 0.13

Hutten, Ulrich von Vormanung an die freien uund Reich Stette Teutscher nation

1522 2 0.08

Anonymous Der Ritterschafft, brüderliche vereynigug, geselschafft oder verstentnuß jüngst zu Landaw fürnemlich Gott zu lob unnd dann folgendt merung gemeynes nutz auch fürderung fridens und rechtenß uffgricht.

1522 2 0.08

Anonymous Artickel und ursprung der waldenser: und der armen von Lugdun auch Joannis wicleffen und Joannis hussen.

1523 3 0.13

Table 21: Market prices for Protestant media from Martin Brecht, “Kaufpreis und Kauf-daten Einiger Reformationsschriften,” Gutenberg Jahrbuch (1972, pp. 169-173). Follow-ing Edwards (1994) we take 24 Pfennigs as the daily wage of an artisan.

83

D Transport Costs

In this section we present evidence on the transport costs characterizing inter-city trade.

Dittmar (2015) provides similar but more detailed evidence.

The inter-city trade in print media faced significant cost barriers. Transport was

generically costly in early modern Europe. Print media were in addition both heavy and

susceptible to damage due to water and humidity (Dittmar 2011).

Historians observe that due to the magnitude of transport costs, local production pro-

vides a measure of local exposure to content (Edwards 1988; 1994). Transport costs were

sufficiently high that print media typically spread through reprinting rather than inter-

city trade (Edwards 1994; p. 8).68 Outside printing cities, information on the range of

available print media was incomplete and many books were not offered for sale, implying

high shadow prices. Flood (1998; p. 55) observes that during the Reformation, “Outside

the towns where books were printed or which were main centers of the burgeoning book

trade the public were dependent on what itinerant traders offered them and on word of

mouth.” During the reformation, pamphlets in particular were typically printed in the

market where they were to be sold and not for export (Pettegree 2000).

The existing literature on the book trade in Renaissance Europe has not directly

estimated or measured the cost of transport.

To document the quantitative relationship between transport and book prices we

study data on book purchases made and recorded in an accounting ledger by Christopher

Columbus’ son, Hernando Colon, in 42 cities across Europe 1510-1540 (see Dittmar (2015)

for further discussion of these data). Hernando Colon’s explicit and recorded aim was to

establish a universal and representative library of works in print. Colon’s purchases were

financed with income derived from forced labor in the Americas and an annuity granted

by the Spanish crown to support his library project. The data from Colon’s purchases

are used to estimate the price gradient associated with distance between the city where

a book was produced and the city where it was purchased, controlling for a rich set of

observable book characteristics and the overall level of book prices in the city-year where

each individual purchase was made.

Figure 35 maps the cities in which Hernando Colon bought books and the cities that

68The evidence evidence in our database is consistent with this observation. For example, in the pricedata above we observe a publication by Von Sickengen from 1522 (previous appendix section). Thispublication was independently printed in 1522 in both Strasbourg and Bamberg. We also observe in theprice data a publication by Bugenhagen from 1525. In 1525, this publication was independently printedin Erfurt, Altenburg, Wittenberg, Nurnberg, Speyer, and Augsburg.

84

produced books that were exported and purchased by Colon elsewhere. Figure 36 shows

the distribution of distances between printing and purchase city for imported books.Books Purchased by Hernando Colon and the Cities They are From

Figure 35: Cities in which Hernando Colon purchased books 1510-1539 (blue markers)and cities producing books that were exported and purchased by Colon elsewhere (redmarkers). City markers scaled to reflect the number of purchases (exports) at eachlocation.

To document the price gradient, we present regressions that estimate the relationship

between book prices and import distance. The basic estimating equation is:

lnPijkt = α ln δjk + θjt + βXi + εi

Here Pijkt is the price of book i purchased in city j, printed in city k, and purchased at

time t. Import distance between city j and city k is δjk and the parameter of interest is

α. The regressions control for city-year fixed effects θjt and book-specific characteristics

Xi, including the size and number of pages of the book, the presence of illustrations, and

the book’s subject matter.

Table 22 documents the price gradient associated with book import distances. All

regressions control for the general level of prices in each city-year with city cross year

85

0

.5

1

Den

sity

12.5 25 50 100 200 400 800 1600Distance from Production City to Purchase City (Km)

Figure 36: The distribution of import distances between purchase city and productioncity for imported books bought by Hernando Colon 1509-1539.

fixed effects, so the identifying variation is differences in import distances among books

purchased in a given city-year. Column (2) presents the baseline specification where the

price gradient (elasticity with respect to distance) is 0.38, controlling for book length,

the number of pages, the presence of illuminations, and subject matter. Column (3)

shows that there was a differential and larger price gradient for pamphlets, defined as

books of 32 pages or less.69 This evidence shows that while pamphlets were far cheaper,

the price of pamphlets rose more quickly with export distance than did the price of

non-pamphlet print media. This is confirmed in column (3), which shows both the price

gradient associated with import distance conditional on controlling for the interaction

between log distance and log pages, which is negatively and significantly associated with

prices.

69Edwards (1994) suggests that 32 pages (eight folded printed sheets) is an appropriate cut-off fordistinguishing pamphlets from other print media that more closely conforms to our contemporary ideaof what constitutes a book.

86

Regression Variables Ln Price Ln Price Ln Price(1) (2) (3) (4)

Ln Distance 0.383*** 0.178** 0.720***(0.090) (0.080) (0.216)

Ln Distance x Pamphlet 0.238***(0.074)

Pamphlet -2.337***(0.442)

Ln Distance x Ln Pages -0.105**(0.044)

Ln Pages 0.227*** 0.003 0.853***(0.037) (0.037) (0.279)

Ln Page Dimension 3.301*** 3.257*** 3.286***(0.513) (0.477) (0.508)

Book Age -0.007*** -0.008*** -0.008***(0.002) (0.002) (0.002)

Illuminated 0.348*** 0.327*** 0.342***(0.094) (0.100) (0.098)

Bibles 0.320*** 0.268*** 0.322***(0.091) (0.093) (0.098)

Orations -0.687*** -0.570*** -0.685***(0.112) (0.107) (0.109)

Poetry -0.570** -0.547*** -0.551**(0.239) (0.203) (0.233)

Theology 0.280** 0.262** 0.280**(0.114) (0.106) (0.121)

City x Year FE Yes Yes YesObservations 1542 1542 1542

Table 22: Regressions documenting the price gradient associated with import distance.Distance measured as the natural logarithm of distance in kilometers between the pur-chase city and the printing city. Standard errors are clustered at the city-year level. Allregressions control for city-year fixed effects. In addition to the controls presented in thetable, all regressions control for the following subject matter classifications: law books,philosophy, literature, medicine, books on languages, history and legislation. Page di-mensions are measured in centimeters. Book age is measured in years. Illuminated is anindicator variable capturing the presence of illuminations (illustrations). Significance atthe 99%, 95%, and 90% levels denoted “***”, “**”, and “*”, respectively.

87

E Robustness and Additional Evidence

E.1 Protestant Content versus Martin Luther’s Work

Our classification of content provides significant new information on the diffusion of

Protestant ideas. As discussed in the text, our classification of content provides evidence

on religion in the media (i) in cities where no religious publications were produced by

leading authors and (ii) for the majority of titles, which are by unknown or little known

authors. In addition, the classification provides substantive evidence for our interpre-

tation of the relationship between competition and diffusion that appears less precisely

when we look just at leading known authors.

To illustrate, we present a comparison of (i) our baseline estimates of the relationship

between market structure and protestant media, and (ii) the relationship between market

structure and the diffusion of Martin Luther’s work. Table 23 compares regressions where

the outcome is the diffusion of Luther’s works to our baseline estimates where the outcome

is the diffusion of all Protestant content. The relationship for all Protestant content is

larger in magnitude and more precisely estimated.

Table 23: The diffusion of Luther versus all Protestant content

Martin Luther’s Work Classified Protestant[1] [2] [3] [4]

Firms 1508-1517 0.18 0.13 0.36*** 0.29**(0.14) (0.14) (0.12) (0.12)

Firms 1508-1517 × Lord Rule 0.24 0.39**(0.18) (0.16)

Observations 191 191 191 191Controls Yes Yes Yes Yes

This table presents regression estimates of the relationship between firms active 1508-1517 and twomeasures of Protestant content 1518-1554. In columns [1] and [2], the dependent variable is the logarithmof the count of Luther’s works published. In columns [3] and [4], the dependent variable is the logarithmour measure of Protestant publications classified (indexed) based on language in text data. In both caseswe study the logarithm of Luther (or Protestant) titles 1518-1554 plus one. All specifications controlfor the complete set of controls in Table 2. Standard errors reported in parentheses are clustered at thestate level. Significance at the 99%, 95%, and 90% levels denoted “***”, “**”, and “*”, respectively.

88

E.2 Robustness to Alternate Measure of Firms

Our baseline evidence on firms is from (Reske 2007). In this section we present results

using the alternate measure of firms constructed from publications data. These estimates

show that our results are robust to using the alternate measure of firms. Table 24 shows

that the relationship between initial 1508-1517 the number of firms and the subsequent

diffusion of Protestantism strengthens when we use the alternate biographical measure as

the explanatory variable or as an instrumental variable (IV) for the measure of firms from

publications. Because the publications data includes a fringe of very small producers on

whom biographical data does not exist, this suggests that the important variation is not

induced by minor printers who appear in just a few book inscriptions, but by variation

in the number of more substantial firms. The measure of firms from publications is likely

to be noisy for two principal reasons. First, there are a few people named on the front

pages of individual books whom it seems unlikely were ever truly independent printers.

Second, because ambiguities in naming variations may lead us to treat as distinct printers

differently named individuals. (Further discussion in main text and data appendix.)

Table 24: Comparing baseline and alternate measure of firms

Ln Protestant 1518-1554[1] [2] [3] [4]

Firms 1508-1517 – Baseline 0.36*** 0.46***(0.09) (0.11)

Firms 1508-1517 – Evidence from Books 0.13** 0.17***(0.05) (0.06)

Observations 191 191 191 191Controls Yes Yes Yes YesGrid Cell Fixed Effects Yes Yes

This table presents regressions using our baseline measure of firms from biographical data (Reske 2007;2015) and an alternate measures of firms constructed from printer inscriptions on historical books. Allregressions use the complete set of controls from Table 2. Columns [2] and [4] introduced grid cell fixedeffects for 3 × 2 degree grid cells. Standard errors clustered on grid cells. Significance at the 99%, 95%,and 90% levels denoted “***”, “**”, and “*”.

E.3 Firms and Output Thresholds

Results in the paper document the mean effects of how variations in industrial structure

on the eve of the Reformation predict the diffusion of religious output. It is natural to also

wonder whether and if variations in initial industrial structure shaped output along the

distribution. To address this question, Table 25 presents estimates from linear probability

89

model regressions that document whether or not city-level diffusion of Protestant media

crosses different thresholds. Panel A presents OLS regressions. In column 1, the binary

outcome is an indicator for cities producing any Protestant religious content. In column 2,

the binary outcome is an indicator for 25+ publications classified as Protestant. Columns

3 and 4 examine cut offs of 50 and 100 publications. The OLS estimates in Panel A shows

that an extra firm was associated with an 8% increase in the probability of the city-level

diffusion of Protestant content be at least 25 or at least 50 publications. In contrast, Panel

A shows that there is a small and imprecisely estimated positive relationship between the

number of firms and the probability of any Protestant content (column 1) and of 100+

Protestant publications (column 4). Panel B presents IV regression estimates that use

printer deaths to isolate exogenous variation in initial industrial structure, and suggests

a strong and statistically significant relationship between induced variations in industrial

structure and subsequent diffusion across the distribution of Protestant content. The

IV estimates suggest relatively strong and significant effects in the upper tail. An extra

firms is associated with a 14%-18% in the probability of Protestant media crossing output

thresholds from 1+ (any) publications to 50+ publications, but with a 40% increase in

the probability of 100+ Protestant publications.

E.4 Diffusion in the New Institutional Equilibrium

We emphasize the role of printing and competition in the media on the diffusion of

Protestant ideas 1518-1554, following the history literature. In the mid-1500s, war ar-

rested the diffusion of Protestantism and a new institutional equilbrium was settled in

law. Historical research suggests that in the mid-1500s the religious geography of central

Europe was settled in a form that would be maintained for several centuries, and that the

relationship between religious competition and the diffusion of Protestant ideas changed

(Brady 2009). But this hypothesis has not been tested using data on printing.

Our data allow us to test the hypothesis. Consistent with this hypothesis, Table 26

shows that the number of firms competing on the eve of the Schmalkaldic War (1546-7)

does not predict variations in Protestant output under the new institutional equlibrium

established with the Peace of Augsburg (1555). We estimate these specifications control-

ling for previous period Protestant media.70 We emphasize that these regressions present

suggestive correlations. The number of firms in the 1508-1517 period is potentially en-

dogenous for reasons discussed in the main body of the text. The number of firms in

the 1538-1547 period is potentially endogenous for similar reasons and because of factors

70In the regressions examining output 1518-1554, this means controlling for publications we index as“Protestant” that appear pre-1517.

90

Table 25: Protestant media crosses thresholds

Panel A: OLS Regressions

Protestant Protestant Protestant Protestant1+ 25+ 50+ 100+[1] [2] [3] [4]

Firms 1508-1517 0.04 0.08** 0.08** 0.03(0.03) (0.03) (0.03) (0.03)

Observations 191 191 191 191

Panel B: IV Regressions

Protestant Protestant Protestant Protestant1+ 25+ 50+ 100+[1] [2] [3] [4]

Firms 1508-1517 0.15** 0.14* 0.18** 0.41***(0.05) (0.07) (0.06) (0.07)

Observations 191 191 191 191

Panel A presents linear probability model regressions in which the outcome is a binary indicator variablefor cities with Protestant content crossing specified thresholds. “Firms 1508-1517” is the count offirms active. In column 1 the outcome is an indicator for cities with at least one classified Protestantpublication. In columns 2 the outcomes are indicator for at least 25, 50, and 100 Protestant publications,respectively. All regressions use the complete set of controls from Table 2. Panel B present IV regressionsfor the same linear probability model, using printer deaths as the IV for “Firms 1508-1517”. The firststage regressions and first stage F statistics are as reported in the main body of the paper. Standarderrors clustered at the state level as in the baseline IV regressions in the main text. Significance at the99%, 95%, and 90% levels denoted “***”, “**”, and “*”.

specific to the Reformation itself.

E.5 Printer Deaths as Shocks to Competition

Our baseline analysis examines the deaths of printers as a source of exogenous variation

in competition that can be used to examine the relationship between industrial structure

in media markets on the eve of the Reformation and subsequent diffusion.

The exclusion restriction for identification is that printer deaths shifted competition

and only through this channel impacted the subsequent diffusion of Protestant ideas in

the media. As noted in the main body of the paper, a natural question is whether printer

deaths actually impacted the diffusion of Protestant ideas by shifting the age composition

of printers. In particular, we wondered whether printer deaths shifted age mix towards

younger and/or more risk taking owner-managers who might have been more likely to

print Protestant content.

We test the hypothesis that printer deaths impacted the diffusion of ideas via their

91

Table 26: Diffusion of Protestant ideas before and after the mid-1500s

Protestant 1518-1554 Protestant 1555-1599[1] [2] [3] [4]

Firms 1508-1517 0.24** 0.19*(0.12) (0.11)

Firms 1508-1517 × Lord 0.32**(0.15)

Firms 1538-1547 0.13 0.19(0.14) (0.19)

Firms 1538-1547 × Lord 0.17(0.15)

Observations 191 191 191 191Controls Yes Yes Yes Yes

This table presents regression estimates of the relationship between initial number of firms and subse-quent Protestant output before and after the Schmalkaldic War of 1546-1547 and the Peace of Augsburgof 1555. The dependent variable is the logarithm of the count of Protestant publications plus one. Allspecifications control for lagged Protestant content and the complete set of controls from Table 2. Stan-dard errors clustered on territorial principality. Significance at the 99%, 95%, and 90% levels denoted“***”, “**”, and “*”.

impact on the age composition of firms, by constructing information on the distribution

of firm age and testing whether variations in firm age across cities are associated with

variations in content. We consider how several measures of firm age respond to printer

deaths, and find no evidence that changes in the age distribution induced by deaths

mattered for diffusion. For example, we calculate the minimum and mean firm age at

the city level and then document that induced variations in these measures of age have

relatively weak and statistically insignificant relationship to the diffusion of Protestant

ideas, and that in a 2SLS setting the first stage is insignificant.

The baseline relationship we report in the main body of the paper is cross-sectional.

However, the competitive dynamics and responses to shocks that we observe in the cross

section are also observed in more finely grained city-year level data across the entire 16th

century. The relationship between shocks and the number of firms competing is in fact

observed across Europe in this era.71

To provide additional evidence on the nature of this relationship we examine the

city-year level data. In our complete data (1454-1600) we observe statistically significant

increases in entrance and in the number of firms competing in the precise city-year in

which a manager death is observed. This relationship holds controlling for the overall

business environment in a given city in that time period as absorbed in city-decade fixed

71Dittmar (2015) documents this relationship in city printing industries across European economies.

92

Table 27: IV estimates of the impact of firm age on Protestant content

Ln Protestant Media 1518-1554[1] [2] [3] [4] [5] [6]

Change in Minimum Firm Age 0.84 0.72 0.87(0.87) (0.97) (0.86)

Change in Mean Firm Age 1.07 0.90 1.33(0.71) (1.35) (1.36)

Observations 191 191 191 191 191 191F Statistic on IV 0.83 0.49 1.21 1.93 0.38 0.88Controls Yes Yes Yes YesGrid Cell Fixed Effects Yes Yes

This table presents 2SLS estimates of relationship between initial firms and the diffusion of Protestantism,using printer deaths as the instrument for measure of firm age. The dependent variable is logarithmof Protestant titles 1518-1554. “Change in Minimum Firm Age” and “Change in Mean Firm Age”are the changes in minimum firm age between 1498-1507 and 1508-1517. All specifications control forthe number of firms active 1498-1507 and for a complete set of categorical indicators for population in1500. Additional “Controls” are as in Table 2. Standard errors reported in parentheses are clustered atthe state level in columns [1], [2], [4], and [5]. Free cities are classed following the Euratlas coding asbelonging to “small states” rather than as individual unique territories. Columns [3] and [6] introducegrid cell fixed effects are for the 3 × 2 degree cells, and cluster standard errors at the grid cell level.Significance at the 99%, 95%, and 90% levels denoted “***”, “**”, and “*”, respectively.

effects. Figure 37 plots regression estimates of the relationship between firms competing

and leads and lags of printer deaths within-city, controlling for time and city fixed effects.

Figure 37 shows that in the precise year in which deaths are observed we see the number

of firms in a city shift up, become significantly larger than the city mean, and stabilize

at that higher level.

93

-.5

-.25

0

.25

.5

.75

1

1.25

1.5

Cur

rent

Impa

ct o

n C

ity-L

evel

Firm

s

-3 -2 -1 0 1 2 3 4 5Year Relative to Firm-Level Shock

Figure 37: Regression estimates of the relationship between the number of firms active atthe city-year level and leads and lags of within-city printer deaths, controlling for city andtime fixed effects. The basic specification is: firmsit = αdeathit

∑5s1=1 βs1deathi,t−s1 +∑5

s2=1 βs2deathi,t+s2 + θi + δt + εit, and uses publication data for firms.

94