New Media, Institutions, and Economic Change:Evidence from the Protestant Reformation
Jeremiah Dittmar and Skipper Seabold∗
Version 0.1 – Preliminary & Incomplete DraftPlease Do Not Circulate
September 27, 2013
Abstract
The printing press was the great information technology revolution of the Euro-pean Renaissance. During the Protestant Reformation of the early 1500s, religiousreformers employed print media to disseminate their ideas. We construct a newmeasure of religious content in the media using data on all known books and pam-phlets printed in German-speaking Europe 1450-1600 (100,000+ publications). Weapply the measure to study the diffusion of the ideas of the Reformation across citiesand time. We document the relationship between the diffusion of religious ideasand the characteristics of local media markets and variations in diffusion inducedby geography. At the city level, we find that Protestant content was associatedwith increases in educational and non-religious print media in high frequency data.We document that cities with high levels of exposure to Protestant ideas were morelikely to adopt the institutions of the Reformation using new data on municipal lawsof the 1500s that established Europe’s first state-run education system. We showthat these legal institutions explain subsequent cross-county variations in literacy.
∗Dittmar: London School of Economics & American University. Email: [email protected]: American University. Email: [email protected]. We thank Russ Gasdia for researchassistance, discussion, and suggestions. Dittmar acknowledges support through the Deutsche BankMembership at the Institute for Advanced Study (Princeton, NJ) and the National Science Foundation.
1
1 Introduction
Sometimes people develop ideas and social movements that radically alter economic in-
stitutions, culture, and beliefs. The use of internet-based communications technologies
in contemporary social movements such as the so-called Color Revolutions, Arab Spring,
and Occupy Wall Street has raised questions about the role of new media in the diffusion
of beliefs, the impact of exposure to new ideas, and institutional change. This research
examines the diffusion of ideas in new media and institutional change over one of the
most significant upheavals in European history – the Protestant Reformation.
The new media used in the Protestant Reformation were print media. Before the in-
vention of the movable type printing press the production of texts had been concentrated
in religious establishments often located outside cities. The printing press was invented
in Mainz, Germany around 1454. Between 1454 and the early 1500s the technology
spread to cities across Europe.1 The price of text data used to store and disseminate
ideas declined. Book prices fell by over 80% between the 1450s and the early 1500s.
The Protestant Reformation was the first mass movement to exploit the new infor-
mation technology and the first religious movement to successfully breach the monopoly
of the Catholic church in Western Europe (Edwards 1994, Brady 1999). In 1517, Mar-
tin Luther posted a set of theses calling for a reform of Church practices in the city of
Wittenberg, in Northeast Germany. In a matter of weeks, Luther’s ideas were being dis-
seminated in print across German-speaking Europe. Over the coming years, Protestant
and Catholic thinkers disseminated their ideas in the new media.
This research addresses three principal questions: What factors determined the diffu-
sion of Protestant ideas in the media? How did the diffusion of Protestant ideas change
media markets more broadly? And what were key impacts of religious and institutional
reform? These two questions reflect religion’s “two-way interaction with political econ-
omy” (McCleary and Barro 2006).
Historians argue that print media were critical to the success of the Reformist chal-
lenge to the religious monopoly of the Catholic church (Brady 1999, Ozment 1981).2
However, previous social science research has not systematically documented the diffu-
sion of Protestant ideas in print in quantitative terms and the existing data on historic
1See Dittmar (2011) for a discussion of the diffusion process, including evidence that the technologydiffused from Mainz rather than from other centers.
2Print media was important despite the fact that literacy levels were low (Edwards 1994). Print mediawas read aloud and the ideas transmitted in print were further circulated in sermons and conversations.On literacy and the impact of print media see Brady (1999) and Scribner (1982) and discussion below.
2
print media do not systematically categorize books or authors by religion.3
A substantial social science literature argues that Protestantism transformed Europe’s
economic landscape, but there is mixed evidence on the economic impact of Protestantism
in historic Germany. Arguments running back to Max Weber (1930) have suggested that
Protestantism fostered a work ethic and social norms conducive to commerce and growth.
Economists have suggested that Protestant beliefs generated a demand shift for literacy
that explains variations in technology adoption and economic outcomes observed across
Prussian counties by the mid-1800s (Becker and Woessman 2009; Becker, Hornung, and
Woessman 2011). However, city growth was historically associated with economic dy-
namism and there is no evidence that cities in territories where rulers adopted Protes-
tantism in the 1600s enjoyed any subsequent growth advantage (Cantoni 2010).
The first contribution of this research is to construct a new, high frequency measure of
religious content in print media at the city and firm levels. We assemble new data on all
known books and pamphlets printed in German-speaking Europe (100,000+ publications
printed in 200+ cities). We use historical sources to identify 500+ leading Protestant
and Catholic authors – authors responsible for 18% of the books printed in Germany
1450-1600. We identify the language characteristic of Protestant and Catholic authors in
long, historical book titles using statistical models from natural language processing.4 We
then generate a measure of the religious content of print media at the firm- and the city-
level that captures how similar the language of their output is to the language of known
Protestant and Catholic writers.5 In contrast to our high frequency measure of content
3For example, the Universal Short Title Catalogue (USTC 2012) which we discuss below simply clas-sifies book subject matter as “religious” or not. The work that comes closest to the spirit of this paperis Edwards (1994), which examines a sample of pamphlets from the first half of the 1500s that com-prised 3,183 pamphlets authored by Martin Luther and 1,763 pamphlets authored by Catholic activists.Consistent with the evidence presented below, Rubin (2011) finds that Protestantism was more likelyto be adopted in cities that had a printing press in 1500. However, the data in Rubin (2011) do notdocument the spread of Protestant media and do not untangle the supply and demand puzzle aroundwhether Protestantism was adopted due to the supply of religious media content or whether printingmay have been associated with growth and the emergence of social groups that had a higher demandfor (were more receptive to) religious innovation. Similarly, the existing data and previous work do notidentify the firms producing content.
4In the data section below, we describe these titles. The median title is 21 words long (the mean titleis 20.4 words). The median title has 153 characters (mean 149.6). For comparison, twitter messages areno longer than 140 characters. As discussed below, estimation strategies similar to ours are widely usedto classify the content of tweets, spam email, and other short texts.
5For similar estimation strategies see Taddy (2012) and Gentzkow and Shapiro (2010), which measuresthe “slant” (ideology) of US newspapers by determining whether they employ language similar to thelanguage used in speeches by Democratic or Republican representatives in the US Congress. Details ofthe estimation strategy are discussed below. The data we construct identify the author, city and year ofpublication, and the printing firm for each book. We identify the firm producing each individual bookfrom the printer information contained on the front pages of historical books (see below and Dittmar2012 for details on the construction of firm-level data).
3
in media, previous work has relied on measures of religious belief observed centuries after
the Reformation (Becker and Woessman 2009) or on the binary Protestant-or-Catholic
religious choice made by rulers in the 1600s (Cantoni 2010).6
The second contribution of this paper is to document (1) the relationship between
the diffusion of Protestant print media and the characteristics of local media markets
and (2) variations in diffusion induced by geography. We show that within cities the
firms that were more likely to become producers of Protestant content were younger,
larger, and had prior specialization in producing humanistic content. Across cities, we
find locations where the vernacular (German) share of output was relatively large were
more likely to adopt Protestant content than locations specializing in Latin print media.
We document that distance from Martin Luther’s base in Wittenberg was significantly
and negatively associated with the diffusion of Protestant content. We further estimate
the cost of inter-city trade using new data on 2,000+ individual level book and pamphlet
purchases made in 42 cities over the 1500s.
The third contribution of this paper is to present new evidence on the relationship
between exposure to Protestant media, media markets, institutions, and economic out-
comes at the city level. We document the positive relationship between exposure to
Protestant media and the diffusion of educational and non-religious vernacular print me-
dia 1517-1600. This finding provides quantitative support for narrative evidence on the
relationship between media, reform, and lay culture previously examined by historians in
select cities (Chrisman 1981). We also document the relationship between local diffusion
of religious content in media and the adoption at the city and territory level of the legal
institutions of the reformation. The legal institutions we consider are 2,000+ church or-
dinances (evangelische kirchenordenung) that both stipulated correct religious doctrine
in law and made provision for the organization, funding, and oversight of Europe’s first
public education system. We show that these laws explain cross-county variations in
literacy and development in the mid-1800s.
Section 2 provides a distilled overview of the history of the Reformation. Section
3 describes the data. Section 4 presents the estimator used to measure the religious
content of print media. Section 5 documents the diffusion of religious content in the
media. Section 6 examines the relationship between the diffusion of Protestantism ideas,
institutions, and economic outcomes.
6Basten and Baetz (2013) examine differences in contemporary preferences using a fuzzy regressiondiscontinuity along a border between Protestant and Catholic cantons in Switzerland.
4
2 History
In this section we present a highly distilled history of the Reformation.7 The Reformation
began as a protest of Catholic clerics and scholars against church institutions and their
superiors. It quickly became a mass movement.
Two key features distinguish the Protestant Reformation from previous attempts to
challenge church institutions and practices. First, when the Catholic church attacked
the protesting clergy, the reformers responded developing and dissemminating their ar-
guments in print media. Second, politically active laymen adopted and adapted these
calls for reform – and pressed them on their governors. “This blending coalition – of
reformers’ protests and laymen’s political ambitions – is the essence of the Reformation”
(Cameron 1991, p. 2).
The Reformation was based on a critique of existing church institutions and practices.
Reformers called for moral renewal within cities (Moeller 1962), emphasized their belief
that biblical authority was paramount over and above the authority of existing church
institutions (Brady 1975), and were typically though not always anti-clerical (Goertz
1975). The reforms that were at the heart of the Reformation included the abolition of
the Catholic rite mass, rejection of the argument and rule that clergy should be celebate,
and moves to eliminate church corruption.
The reformers produced and set-up new legal institutions governing religious prac-
tice and education. The legal institutions of the Reformation were church ordinances
(evangelische kirchenordnungen and schulordnungen) passed by magistrates, municipal
councils, and princes at the city and territory level. These ordinances provided for the
first state-run public education system in Europe. The ordinances established schools,
sources of funding for these schools, and oversight mechanisms.8 The oversight mech-
anisms included regular and formal “visitations” in which the physical assets of local
schools would be inspected and both student and teacher performance would be assessed
and recorded.9 As discussed below, we examine data on some 2,000+ ordinances.
7There is a very rich literature on the history of the Reformation to which we only allude here. Weask historically sensitive readers to make allowances for the fact the current research is directed at anaudience of economists. Similarly, our use of techniques from economics and machine learning should beseen as complementing – not substituting for – research based on primary sources and local case studies.
8A small subset of ordinances provide explicit terms of teacher pay.9Formal proclamations were printed and sent to each local authority before each visitation. Typically,
visitations occurred once per year. Local church officials were required to submit a tabulation of revenuesand inventory of relics. Preachers and towns and village people were formally questioned on doctrineand the adequacy of funding for local schools by inspectors who then submitted written reports. SeeStrauss (1981).
5
The Reformation is usually dated to October 1517, when Martin Luther posted a
set of hand-written theses criticizing church corruption in Wittenberg. Luther was a
Dominican monk and professor of theology at the University of Wittenberg. Luther’s
theses notably criticized the Catholic church’s practice of selling indulgences which were
believed to secure the early release of dead relatives from purgatory in the afterlife. The
proceeds from the sale of indulgences were used to finance church investments (e.g. the
basilica of St. Peter in Rome) and consumption.10 Luther circulated his theses in letters
to three correspondents. Within weeks they were being printed throughout German-
speaking Europe.
Cities and urban actors played a central role in the development and diffusion of
Reformation ideas and institutions (Moeller 1962).11 Cities had the concentrations of
people and levels of literacy and cultural sophistication to put the ideas of religious
reformers, preachers, and pamphleteers at the top of the political agenda in the 1520s
(Brady 1998, Ozment 1975). The ideas of the Reformation diffused within and across
cities and were first formalized at the city level in municipal ordinances in the 1520s and
1530s. The adoption of the Reformation at the territory or principality level came later,
after city-level adoption.12
Historians argue that print media played a central role in the diffusion of the Protes-
tant Reformation (Edwards 1994, Brady 1999). In the early years of the movement,
Reformist ideas were innovations with very uncertain prospects that had not yet diffused
through the population. As Chrisman (1980, p. 29) observes, “The importance of the
printers lay neither in birth, wealth, property, nor political power...It lay in their control
of the printed word. Ultimately, 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.” Beyond the impact
of the particular book or tract, the overall shift in ideas and increase in quantity of print
media influenced events. “It was the superabundance, the cascade of titles, that created
the impression of an overwhelming tide, an unstoppable movement of opinion...Pamphlets
and their purchasers had together created the impression of irresistible force.” (Pettegree
2005, p. 163). Tens and even hundreds of thousands of copies of individual Lutheran
pamphlets were printed and Martin Luther became the first best-selling author. As a
10In 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?”
11The Reformation was by no means restricted to cities. Religious ideas were notably central to theso-called peasants war of 1525 (Blickle 1979).
12Cantoni (2010) examines the diffusion of the Reformation at the territorial level.
6
counterfactual to the success of Protestantism in the print age, historians have examined
the failure of pre-print heresies such as Lollardism in England and the Hussite movement
in Bohemia in the late 1300s and early 1400s. The suppression of the Hussite movement
notably involved the burning of some 200 Hussite manuscripts in Prague in 1410 (Cheney
1936).
The decentralization of the Holy Roman Empire limited the emperor’s capacity to
regulate religious and political speech and was an important feature of the environment
in which the Reformation developed and spread. The Holy Roman Empire was composed
of a large number of semi-autonomous principalities, as well as 85 so-called free imperial
cities. The free imperial cities were jurisdictionally free from the control of local territorial
lords and princes and in theory answerable only to the emperor.13 In practice, these
cities had considerable autonomy in governance.14 In this context, suppression of dissent
was relatively costly and in fact 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 itself not rushed into print (Brady 1985, p. 153). By the
mid-1520s, magistrates and urban governments – notably in the free cities – begin to
defy the emperor and institute ordinances institutionalizing the Reformation. In 1526, a
formal magesterial right to reform (ius reformandi) was passed into law by the Imperial
Diet or parliament of the Holy Roman Empire (Brady 1995, p. 55).
The ideas of the Reformation spread through a two-part process and were followed
by staggered institutional change (Edwards 1994). First, print media impacted “opinion
leaders”, notably clergy and educated laymen.15 Opinion leaders then transmitted ideas
orally to the broad public and developed a militant popular movement. At the city level,
the Reformation typically reflected a basic conflict within cities between lower and middle
burghers on the one hand and increasingly plutocratic and oligarchic elites on the other
(Ozment 1975, p. 121-3).
Institutional change came after the diffusion of the ideas of the Reformation as the
last and decisive steps in formalizing the Reformation (Ozment 1975). However, even
the institutional changes were gradual. In Wittenberg, Catholic mass was not abolished
until 1525. In Strasbourg, Catholic mass was restricted to specific churches in 1525 and
13The number of cities with the legal status of free imperial cities fluctuated over time. There were85 such cities listed in the imperial budget (Reichsmatrikel) of 1521.
14The empire as a whole had a representative assembly called the Diet. The Diet provided represen-tation for the different “estates” involved in governance. These included: seven electors (seven princes),the college of imperial princes, and the college of imperial cities.
15A number of significant lay proponents of the Reformation were city clerks. For example, LazarusSpengler was the town clerk of Nurnberg and published a defence of Luther that was printed in Basel in1520. Jorg Vogeli became city clerk of Konstanz in 1524, and defended three local Reformist preachersin a series of publications over the 1520s.
7
abolished only in 1529. In Nurnberg, the legal institutions of the Reformation were
instituted between 1520-1533. In Osnabruck, the process lasted from 1521 to 1542. By
the middle 1500s, Lutheran Protestantism in historic Germany acquired the geographic
distribution it would maintain for several centuries (Brady 1998, p. 373). The Peace of
Augsburg (1555) established the rule of cuius regio, euis religio (whose rule, his religion)
with exceptions for mixed cities where Protestants and Catholics were to share churches
and magistracies (Brady 1998, p. 375).
3 Data
We focus attention in this research on new media and institutional change in the German-
speaking regions of the Holy Roman Empire.
The primary source for data on print media is the Universal Short Title Catalogue
(USTC 2012) database. The USTC is designed as a universal catalogue of all known
books printed in Europe 1450-1600.16 We rely on Reske (2007), which identifies 384
German-speaking cities with printing.17 Figure 1 shows the boundaries of the principal-
ities comprising the Holy Roman Empire and the German-speaking cities where print
media was produced 1450-1600. Our database comprises 101,042 titles that were printed
1457-1600 in these cities.
Figure 2 shows the number of book titles printed each year in German-speaking
Europe. The unit of analysis in this paper is the book title (edition), which can be
thought of as a variety.18 Figure 3 shows the dynamics of book production across Europe
by subject.
We construct data on the religious affiliation of authors from several sources. Klaiber
(1978) provides data on 564 Catholic authors and “controversialists” working in German-
16The closest competitor dataset is the Consortium of European Research Libraries’ Heritage of thePrinted Book (HPB) database. The HPB is essentially a master catalog of the major national researchlibrary catalogs. However, since these catalogs are themselves incomplete for several countries the USTCprovides much more comprehensive coverage. In the Appendix, we present evidence that rules out theleading concern around survivorship bias due to historical events that might distort the data on historicalbooks.
17The set of cities in our data includes cities now in Austria, France (e.g. Metz and Strasbourg),Switzerland (e.g. Zurich and Basel), Poland (e.g. Gdansk, historically known as Danzig, and Szcezecin,historically known as Stettin), the Czech Republic (e.g. Brno or Brunn), and Russia (Kaliningrador Konigsberg). These cities should be considered German-speaking in a qualified sense. Many werecharacterized by linguistic diversity. Within German-speaking Europe High and Low German co-existed.
18For a subset of several hundred books we have data on the number of copies printed per edition.These data and evidence from book contracts and printer’s correspondence indicate that the typicalprint run rose from 400-800 copies around 1500 to 1,000-1,400 copies in the later 1500s. We present thisevidence and further discussion of print runs in the appendix.
8
Figure 1: Cities Producing Books in German-Speaking Europe. Principality boundariesare as of 1500. Cities are German-speaking printing centers from Reske (2007).
0
500
1000
1500
2000
Boo
k T
itles
Prin
ted
1475 1500 1525 1550 1575 1600
Annual 10 Year MA
Figure 2: Book and Pamphlet Titles Printed in German-Speaking Europe
9
0
.05
.1
.15
.2
Sha
re o
f Out
put
1475 1500 1525 1550 1575 1600
Merchants' Manuals Science Education Bibles
Figure 3: The Subject Matter of Print Media
speaking Europe. For each author, the Klaiber data provide a standardized name, a
list of name variants (alternate spellings and aliases), and a list of publications. For
Protestant authors we rely on Mullett (2010), Carey (2000), and Wikipedia’s list of over
900 Protestant Reformers.19 Based on these sources, known Catholic authors account for
2,937 titles (3% of books) and known Protestants account for 15,142 (15% of books).20
The research uses the text of long book and pamphlet titles to identify the language
most characteristic of Protestant and Catholic print media. Historical titles typically
provide an extended gloss on content. Table 1 summarizes the distribution of the word
lengths of these titles. The first column of Table 1 shows that the median title in our data
is 21 words long (the mean title is 23.2 words); at the 5th percentile titles are 3 words long;
at the 95th percentile titles are 51 words.21 To estimate the language characteristic of
religious authors we construct a vocabulary that excludes common “stop words” (words
such as and and the) and extremely rare words. The second column of Table 1 provides
19See www.wikipedia.org/wiki/List of Protestant Reformers.20The 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)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.
21The median title has 153 characters (mean 171.6). For comparison, twitter messages are no longerthan 140 characters. A substantial literature employs statistical models to identify sentiment in twittermessages. For example, Go, Bhayani, and Huang (2009); Pak and Paroubek (2010); Bollen, Mao, andpepe (2011); Bifet and Frank (2010).
10
Table 1: Number of Words in Titles
Total Words Words in Estimatingincluding Stop Words Vocabulary
5th Percentile 3.0 2.025th Percentile 11.0 7.0Median 21.0 13.075th Percentile 33.0 20.095th Percentile 51.0 32.0Mean 23.2 14.4
summary statistics on titles restricting to the words in the estimating vocabulary.
To document the diffusion of religious content produced at the firm level we construct
annual panel data on the output of 1,577 printing firms operating in historic Germany
1450-1600. We identify the individual firms that produced each publication from the
printer information recorded on the front pages of historic publications. Information
identifying the printing firm is available on 99% of historic books and pamphlets.22
To identify the legal institutions of the Reformation we code over 2,000 reforma-
tion ordinances (kirchenordennung) passed at the municipal and provincial levels. These
ordinances were laws written and passed by local magistrates to institutionalize the ref-
ormation. The ordinances have five key dimensions embodied in provisions governing (1)
religious dogma, (2) public morality, (3) sex, family, and marriage, (4) social welfare and
poor relief, and (5) education (Witte 2002). A key innovation in the new legal framework
was provision for state-run education in the form of city and territorial schools, together
with a system of funding and oversight. The principal source for the data on reformation
ordinances is Sehling, Die evangelischen Kirchenordnungen des 16. Jahrhunderts (21
volumes 1902-1910).23 We supplement Sehling with historical sources described in the
appendix to identify the dates of ordinances in cities outside contemporary Germany.
To document the costs of the inter-city book trade we construct new data on the price
of books and trade routes. The price data are from the purchasing notebooks of Hernando
Colon (Christopher Columbus’ son).24 Over the period 1509-1540, Colon purchased
several thousand books in 42 different cities. Colon’s notebooks record the price paid
for individual books in the local currency and the current exchange rate (Martınez,
22Printers are identified in multiple languages (e.g. Latinized and German variants of the same name),with non-standard spelling, abbreviations, and in some instances aliases. We construct standardizedfirm-level identifiers for each printer and book. For details see Dittmar (2012).
23We refer to these volumes collectively as Sehling (1910).24The data in the notebooks are to my knowledge (1) our best and most comprehensive currently
available source of data on book prices in the 16th century and (2) have not previously been used ineconomic research other than Dittmar (2012). The data are discussed in further detail in Dittmar (2012).
11
Asencio, Wagner 1993; Biblioteca de Huelva 2012). We extract data on book prices in
local historical currencies, the contemporaneous exchange rate, and book characteristics
from the Colon notebooks and the archive catalogue.25 We match these price data to
information on book characteristics. The characteristics include the city and date where
the book was purchased, the city and date where it was printed, the length in pages,
the physical size of the pages, the format (octavo, quarto, folio, etc.), and whether there
are illuminations. The catalogue also provides a subject classification we use to identify
book subject matter.26 We construct a digitized GIS network of historic land and river
trade routes as described below.
We rely on city population data from Bairoch, Batou, and Chevre (1988) and data
from the Prussian Censuses of the 1800s (Becker, Cinnirella, Hornung, and Woessmann
2012). These and other data are described as introduced below and in the appendix.
4 Measuring Religious Ideas in Media
4.1 Overview
We construct an index of the religious content of print media as follows. We first use
historical sources to identify the religious denomination – Protestant or Catholic – of
a subset of 500+ authors. We then document which language (words and phrases) is
important in identifying religious denomination for authors with known religious affili-
ation. The information on partisan language is used to predict the religious content of
media where authors’ religious affiliations are not known and to construct an index of
content. In the economics literature, Gentzkow and Shapiro (2010) use a similar estima-
tion strategy to identify the dimensions of political ideology in Congressional speeches
and to measure the political “slant” of US newspapers by documenting the extent to
which these media use language characteristic of Democrats or Republicans.27
The distinction between Catholic and Protestant authors is stark and revealing. We
argue that this distinction provides a fruitful starting point and powerful first model for
25A typical example of how the notebooks record prices is as follows: “Este libro costo 8 negmit enAnvers a 29 de julio de 1531 y el ducado de oro vale 320 negmit.” In my translation: “This book cost 8negmit in Anvers [Antwerp] on July 29, 1531 when the gold ducat was worth 320 negmit.”
26The classification includes bibles, jurisprudence, philosophy, literature, orations, poetry, theology,medicine, languages, religion, and history and legislation.
27Taddy (2012) extends the Gentzkow and Shapiro (henceforth “GS”) approach by introducing amodel-based framework to estimate such indices. Taddy shows that the GS index is a special case in theInverse Regression (IR) framework which seeks to reduce high-dimensional text data to low-dimensionalsufficient statistics.
12
thinking about media markets and religious change in historic Germany. However, the
distinction between Catholics and Protestants clearly does not exhaust the distinctions
one could draw among authors and literatures.28
Our estimation strategy has traction because historical book titles provide extensive
glosses on the content in books. To fix ideas and get a “feel” for our data examples are
useful. The overwhelming majority of books in our database were written in German and
Latin. Here we provide two examples of English-language books printed in 16th century
Germany that may be useful for English language readers. An example of a Protestant
title is a book written by Martin Luther and produced in Wesel (North Rhine-Westfalia):
The last wil and last Confeßion of martyn Luthers faith concerming 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 produced in Emden
(East Frisia):
A Confeßion of the most auncient and true christen catholike olde belefe
accordyng to the ordre of the .XII. Articles of our co[m]mon 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 illustrate the fact that historical titles convey significant information
about content. They also suggest potential challenges in analyzing historical text data in
which spellings are not yet standardized. The results described below suggest that these
difficulties do not preclude obtaining useful estimates.29
Below we document how our estimation strategy accurately predicts the religious
affiliation of known religious authors in out of sample tests. We also document that our
measures of the local production of religious ideas capture information not present in
the existing bibliographic meta-data – which simply designate book subjects as religious
28Extensions could distinguish between Lutheran, Calvinist, and Zwinglian ideas in Protestant media,between different types of Catholic authors, and between time varying features of religious language.
29According to one ideal, we would want to identify spelling variants as the same word. However, it isnot clear that introducing standardizations or model features to capture orthographic variations mightimprove the precision of our estimates. Because spelling conventions reflect both regional and ideologicalinfluences, orthographic standardization could entail the loss of potentially valuable information. Forthis reason, and because standardizing orthography is a difficult computational problem exacerbated bythe fact that we have nearly 70 different languages and language combinations in the texts, we analyzetexts “as is”.
13
or non-religious, without capturing the denominational nature of the content. Similarly,
we are able to estimate the religious content of print media in 136 cities while books
produced by our set of known religious authors appeared in only 104 cities.
4.2 Estimator
After we code a subset of the titles to document whether the author is a Protestant or a
Catholic, we use supervised learning techniques to extract low-dimensional features of the
text that indicate whether the author is likely Catholic or Protestant. We then project
these features onto texts for which the author’s religious affiliation is not known to get a
measure of similarity to the texts of known Protestants and Catholics. The estimation
strategy used is multinomial inverse regression (MNIR) introduced by Taddy (2012).
The idea of MNIR is to reduce the dimensionality of the documents to a summary
statistic while preserving meaningful sentiment and/or content information. The key
assumption for MNIR and in the text analysis literature more broadly is that the order
of the words or phrases is relatively unimportant in classifying the content of a text. Fol-
lowing the exchangeability assumption a text is treated as a “bag” of words or phrases
(Blei et al. 2005; Taddy 2012).30 The exchangeability assumption allows the econome-
trician to treat a document as a multinomial random variable in which the words or
phrases are the categories. Let Xi be a document. Any such document may be denoted
Xi = [xi1, xi2, . . . , xiW ] where xiw represents the number of times token w appears in
document i for each of the W words in a given vocabulary V . The vocabulary V is
analogous to the possible categories of a multinomial variable and must be chosen be-
fore estimation. We discuss how we construct our vocabulary below. In this setting,
we are interested in ascribing to these summary statistics a religious affiliation r, for
Protestant P and Catholic C. Assuming that Protestant and Catholic documents are
independent, we may collapse the counts for each document such that Xr =∑
iXi|ri = r
for r ∈ {Protestant, Catholic}. Following Taddy (2012), the basic MNIR model of each
document is given:
Xr ∼MN(qr,mr) where
qrw =exp [αw + rϕw]∑Wl=1 exp [αl + rϕl]
for w = 1, . . . ,W, r ∈ (P,C)(1)
where each Xr is a W−dimensional multinomial variable with size mr =∑
imi, mi =
30One might choose to use the word token in place of word. The unit of analysis can be a word or aphrase as we show below.
14
∑Ww xiw, and probabilities qr = [qr1, . . . , qrW ] that are a linear function of r transformed
by the logistic link function. The sufficient reduction (SR) score for the token frequencies
fi = ximi
is31:
zi = ϕ′ → ri⊥⊥xi,mi|xi (2)
Given the sufficient reduction projection of ϕ′ onto the frequencies, identifying the
relationship between documents and religious affiliation reduces to a separate univariate
logistic regression, called the forward regression:
Pr(ri = r|zi) =1
1 + exp [β0 + β1zi](3)
Since both fat-tailed and sparse multinomials arise naturally in the context of text
data analysis, independent Laplace priors with unknown variance are placed on the in-
dividual regression coefficients ϕw. The unknown rate parameter λw accounts for our
uncertainty as to how much variable-specific regularization is appropriate. The rate
parameter 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. (4)
Estimation of the likelihood implied by the multinomial distribution in (1) and the
prior (4) takes place via the gamma-lasso algorithm to maximize the joint posterior over
coefficients and their prior scale (Taddy 2012).32
4.3 Identifying Partisan Vocabulary
To prepare the titles for use in the classifier described above, it is necessary to identify
the words which serve as the support for the multinomial distribution. This support will
be refered to as the vocabulary.
We construct the vocabulary as follows. First all generic titles are removed.33 We
clean the remaining text by making the titles lower-case and removing punctuation,
numbers and roman numerals, words that occur fewer than four times, and German and
31The conditions under which the sufficient reduction holds are detailed in Taddy (2012).32This algorithm outperforms similar algorithms in precision and computational efficiency as docu-
mented in Taddy (2012).33Titles that do not contain words illuminating what the works are about such as Theil, an alternate
spelling of the German teil meaning part of a whole, and Samml, an abbreviation of the German forcollection, are removed.
15
Table 2: Distribution of Known Catholic Title Lengths
Total Words Words in Estimatingincluding Stop Words Vocabulary
5th Percentile 6.0 4.025th Percentile 14.0 9.0Median 23.0 15.075th Percentile 34.0 21.095th Percentile 50.0 33.0Mean 25.1 16.0
Table 3: Distribution of Known Protestant Title Lengths
Total Words Words in Estimatingincluding Stop Words Vocabulary
5th Percentile 4.0 2.025th Percentile 10.0 6.0Median 17.0 10.075th Percentile 27.0 15.095th Percentile 43.0 26.0Mean 25.1 16.0
Latin stop words.34 The impact of removing stop words on the title length distribu-
tions can be seen in the second column of table (1). As can be seen in tables (2) and
(3) the sample-wide distribution is indicative of the sub-populations of Protestants and
Catholics.
We identify “Protestant phrases” and “Catholic phrases” as those phrases that are
used significantly more by one group than the other. To do this, we obtain counts of the
individual words, bigrams, and trigrams used in each unique cleaned title.35 We then
place the word and phrase counts into a contingency table and compare the incidence of
usage by the identified Protestants against that of the identified Catholics – normalizing
the marginal totals for overall phrase usage within the religious camps. We use the
Pearson’s χ2 statistic as in Gentzkow and Shapiro (2010). The χ2 measure is:
χ2wn =
(fwncf∼wnp − fwnpf∼wnc)2
(fwnc + fwnp) (fwnc + f∼wnc) (fwnp + f∼wnp) (f∼wnc + f∼wnp)(5)
34German 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.
35Bigrams and trigrams are two- and three-word phrases, respectively. We do not stem words toreduce them to a common root due to the lack of orthographic standardization demonstrated above.As shown below, we obtain meaningful results without stemming. For the computation of counts usedto construct measures of denominational language, we temporarily drop titles that may be reprints andwould therefore skew the usage of certain words simply by the number of times a given title appearedin print.
16
Table 4: Most Protestant Phrases
single words translationchristi Christsermon sermonchristen Christianwider againstgottes Godtwo-word phrases translationjesu christi Jesus Christmartini lutheri Martin Lutertheologiae doctore Doctor of Theologyphilippi melanthonis Philip Melanthonisdoctore professore Doctor and Professorthree-word phrases translationtheologiae doctore professore Professor and Doctor of Theologypraeside jacobo heerbrando President Jacob Heerbrandunsers jesu christi of our Lord Jesus Christsacrosanctae theologiae doctore Doctor of Sacred Theologyunsers jhesu christi of our Lord Jesus Christ
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. 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 both groups to obtain our vocabulary. This gives 1,335 unigrams, 2,000
bigrams, and 686 trigrams for each religious group. Tables 4 and 5 list the top n-length
phrases for Catholics and Protestants.
When predicting, we collapse all of the titles for one author to obtain a single oeuvre.
We do this for two reasons. First, it yields a single predicted probability for each author,
Catholic or Protestant, rather than a prediction for each title. Second, it makes the
multinomial observations less sparse, improving accuracy of classification. After elimi-
nating words that are not in the vocabulary and collapsing across authors, we have on
average 366 words per author in the training set in which we know the author’s religious
affiliation and 30 words on average for the data set used for prediction in which the
religious affiliation of the author is unknown ex ante.
17
Table 5: Most Catholic Phrases
single words translationfidei faithpetri Petergeorgii Georgeadversus againstmissae postedtwo-word phrases translationsacrament altars Sacrament altarspetri mosellani Petrus Mosellanus (Peter Schade)martinum lutherum Martin Luthersanctorum patrum Holy Fatherursachen warumb Causes whythree-word phrases translationhochwirdigen sacrament altars Blessed Sacrament of the Altarsrhetorica philippi melanchthonis In the Rhetoric of Philip Melanchthonispaedologia petri mosellani Paedologia, by Petrus Mosellanuslaurentium surium carthusianum Carthusium Laurentius Surius (Lorenz Sauer)petri mosellani protegensis Petrus Mosellanus Protegensis
4.4 In and Out of Sample Tests
To assess the performance of the estimator we present within-sample predictions and the
document the estimator’s performance under different prior specifications on held-out
data from our training sample (i.e. the data coded with authors’ religious information).
Figure 4 presents in-sample estimates and shows that the estimator predicts correctly
in-sample approximately 90% of the time.36
To document the out-of-sample performance, we estimate the model 100 times using
different random samples containing 60% of the available training data, then predict over
the remaining data not used for estimation. In the process, we vary the shape parameter
α of the gamma hyperprior which controls the sparsity of the estimates. Increasing α
can reduce problems with overfitting. Figure 5 presents the results of this exercise and
shows that α = 1 has the best out-of-sample performance.37
36Our estimates have a correlation of XYZ with true religious affiliation. For comparison, Gentzkowand Shapiro (2010) use a similar strategy to predict party affiliation of members of Congress based onthe text of their speeches and obtain a correlation of 0.61 between true and predicted affiliation.
37One of the attractive properties of MNIR is its relative speed compared to other estimators, whichcan take orders of magnitude longer to converge. Wall times for the estimator are given in the Appendix.
18
Protestant Catholic0.0
0.2
0.4
0.6
0.8
1.0
Pr(C
atho
lic==
1)
Figure 4: In-sample performance for MNIR estimator over all known Protestant andCatholic authors. The estimator correctly predicts 89% of the authors’ religious affilia-tions. Incorrect predictions are in red.
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
% M
iscl
assi
fied
s = 0.01 s = 0.10 s = 1.00 s = 10.00Prior for shape
0
10
20
30
40
50
60
Wal
l Tim
e (s
econ
ds)
Figure 5: Out-of-sample performance and wall times under various specifications forhyperprior shape parameter s. λj prior if Γ(s, .5). 100 runs of the model trained on 60%of the data and predicted on the held-out 40%.
5 The Diffusion of Religious Ideas
In this section we first present our measures characterizing the diffusion of Protestant
content. We then document the characteristics of local media markets associated with
diffusion and variations induced by geography.
19
5.1 The Pattern of Diffusion
Our estimation strategy delivers a measure of religious content. In this section we first
present our results in graphical form. We aggregate book-level estimates to obtain sum-
mary measures of religious content at the firm, city, and regional level.
Figure 6 presents the estimated index of religious content in print media for all of
Germany 1475-1600. Figure 6 shows that the relative intensity of Catholic-type speech
was approximately stable in the run up to the Reformation, at which point there is a
discontinous shift towards Protestant-type speech. The first panel shows our estimates
for all religious books. The second panel examines just religious publications in German.
In German language media, there is a positive pre-1517 trend away from Catholic-type
speech. In German language media, there is also a relatively much larger increase in the
number of publications printed in the post-1517 period as indicated by the scale of the
annual markers. This evidence of a move away from Catholic-type media in the vernacular
before Luther’s 1517 intervention is consistent with the observation that Protestantism
was in part a response to underlying cultural trends.38 The third panel documents that
the discontinuous shift towards Protestant-type speech also characterizes religious media
published in Latin.
0
.25
.5
.75
1
Rel
igio
n In
dex
1475 1500 1525 1550 1575 1600
All Religious Books
0
.25
.5
.75
1
Rel
igio
n In
dex
1475 1500 1525 1550 1575 1600
Religious Books in German
0
.25
.5
.75
1
Rel
igio
n In
dex
1475 1500 1525 1550 1575 1600
Religious Books in Latin
Figure 6: Religion index in historic Germany. Protestant = 1. Catholic = 0. Markerspresent the annual mean of the religion index across all publications. Marker sizes arescaled to represent the relative number of publications in each year.
Figure 7 maps the evolution of media output at the city level. In Figure 7, city
markers are scaled to reflect the number of titles produced and shaded to reflect the
38Thus Cameron (1991, p. 175) observes that the majority of first-generation reformers had beeninfluenced by Renaissance humanism before they became advocates of religious reform.
20
average value of the Religion Index for city-level output. Cities with lighter markers
produce Catholic-type print media. Cities with darker markers produce Protestant-type
print media. Panel A presents data for the period up to 1517, when Martin Luther posted
his Theses. Panel B presents data for the period 1518-1539, at which point Duke Georg
of Saxony died. Panel C presents data for 1540-1555, the period between Duke Georg’s
death and the Peace of Augsburg. Panel D presents data from 1556-1600, the period
following the Peace of Augsburg.
Figure 8 compares predicted and observed Protestant output across 226 firms and 127
cities producing 1517-1600.39 Four key facts stand out. First, while there is considerable
variation, the majority of cities and firms were net producers of Protestant-type print
media. Second, there is a set of smaller firms that did not print any known religious
authors for which we find variation in predicted Protestantism These are firms for which
the observed Protestant Index is 0.5. Third, a number of firms with observed protestant
index P ∈ (0.5, 0.6) have predicted content P < 0.5. Fourth, for many firms with
observed Protestant share P ∈ (0.7, 0.8), the predicted Protestant index exceeds what
is observed on the basis of known authors (i.e. P > P ). These facts raise the question:
on average, does the estimator predict more or less Protestant media than is directly
observed on the basis of the circumscribed set of known religious authors?
Figure 9 shows that we systematically estimate more Protestant print media than is
directly observed based on coded authors. The ratio of predicted to observed Protes-
tantism provides a parsimonious measure of the relative magnitude of predicted and
observed religious output. Figure 9 plots the distribution of this ratio at the city- and
firm-level. The fact that our estimation strategy captures otherwise unobserved Protes-
tant output is evident in the fact that the distributions have their mass concentrated in
values greater than one.
5.2 Determinants of Diffusion
In this section we document the determinants of the diffusion of religious print media.
We first document city-level variations in media markets and geography associated with
Protestant media production post-1517. We then document the association between firm
characteristics prior to 1517 and the Protestant share of output post 1517.
39The observed output shares are calculated by indexing Catholic authors with 0, Protestant authorswith 1, and authors with unknown religious affilation with 0.5.
21
A: Media Output 1501-1517 B: Media Output 1518-1539
Page 1 of 1
2/21/2013file:///E:/Dropbox/share_maps_mean_period1.svg
Page 1 of 1
2/21/2013file:///E:/Dropbox/share_maps_mean_period2.svg
C: Media Output 1540-1555 D: Media Output 1556-1600
Page 1 of 1
2/21/2013file:///E:/Dropbox/share_maps_mean_period3.svg
Page 1 of 1
2/21/2013file:///E:/Dropbox/share_maps_mean_period4.svg
Figure 7: City-level print media output. City markers are scaled to reflect the numberof book titles produced. Markers are shaded to reflect the average value of the ReligionIndex for city-level output. Cities with lighter (whiter) cities produce Catholic-type printmedia. Cities with darker (blacker) cities produce Protestant-type print media.
5.2.1 Cities
At the city level, historians and economists observe that Protestant ideas diffused from
Wittenberg (Pettegree 2010; Becker and Woessman 2009). Figure 10 plots predicted
religious content at the city level against distance from Luther’s base at Wittenberg.
Figure 10 is consistent with the diffusion argument in Becker and Woessman (2009),
which argues that Protestantism diffused from Wittenberg and exploits this distance as
an instrument for Protestantism.40
It is natural to wonder how non-geographic characteristics of local media markets
40This argument is also supported by historical evidence on Wittenberg’s transformation from a minorprovincial town to a hub in the media network. See for instance Pettegree (2010).
22
0
.25
.5
.75
1
Pre
dict
ed P
rote
stan
t
0 .2 .4 .6 .8 1Observed Protestant Index
Predicted Protestant Fitted values
City-Level
0
.25
.5
.75
1
Pre
dict
ed P
rote
stan
t
0 .2 .4 .6 .8 1Observed Protestant Index
Predicted Protestant Fitted values
Firm-Level
Figure 8: Index of predicted Protestant print media versus index of observed Protestantoutput. At left, city-level output. At right, firm-level output. Markers are scaled bytotal output at the city or firm level. The index is constructed with Protestantism = 1and Catholicism = 0. See text for details.
0
.5
1
1.5
Den
sity
0 .5 1 1.5 2Ratio of Predicted to Observed Protestant
kernel = epanechnikov, bandwidth = 0.1100
City-Level
0
.5
1
1.5
Den
sity
0 .5 1 1.5 2Ratio of Predicted to Observed Protestant
kernel = epanechnikov, bandwidth = 0.0941
Firm-Level
Figure 9: Ratio of predicted Protestant print media to index of observed Protestantoutput. At left, city-level output. At right, firm-level output.
varied with religious ideas, and whether cities specialized in particular types of print
media were more receptive to religious innovation. A key dimension for variation in
local media markets concerned the extent to which they produced German-language
(as opposed to Latin) content before the Reformation. In general, Latin output served
scholarly and religious consumers who were arguably less receptive to religious innovation.
23
Wittenberg
Leipzig
MagdeburgErfurt Nuernberg
Frankfurt am Main
Koln
AugsburgStrassburg
Basel
0
.2
.4
.6
.8
1
Pre
dict
ed R
elig
ion
0 100 200 300 400 500 600 700Distance to Wittenberg in Km
Figure 10: Predicted Protestant print media at the city level. Marker scales reflect overallnumber of titles produced in each city.
In addition, historians observe that the Protestant Reformation was characterized by an
increase in the use of German vernacular.
Figure 11 examines the data on Protestant print media and the intensity of vernacular
(German-language) printing before and after the Reformation. Figure 11 shows that
cities that became centers of Protestant printing post-1517 were places in which the
share of print media in German vernacular was high before 1517. This finding qualifies
an influential claim in the literature, namely that Protestantism in Germany was drove
growth in access to vernacular print media (Brady 1999, Becker and Woessman 2009,
Rubin 2010). As with our evidence on the pre-1517 trend in the Religion Index away
from Catholic-type content, this finding indicates that pre-Reformation characteristics
were associated with subsequent adoption of religious content.
While the characteristics of pre-1517 media markets were associated with subsequent
diffusion of Protestant ideas, Figure 12 shows that the relationship between distance
to Wittenberg and Protestantism is significant even conditional on pre-1517 vernacular
printing. Figure 12 plots residual Protestantism (1517-1600) controlling for the share
of pre-1517 books printed in German against distance from Wittenberg to document
that Protestant content was relatively high in cities close to Wittenberg and declining in
24
distance.
Basel
Koeln
NuernbergWittenberg
0
.2
.4
.6
.8
1
Pre
dict
ed P
rote
stan
t 151
7-16
00
0 .2 .4 .6 .8 1Share German Language 1450-1517
Protestant 1517-1600 vs. Pre-German
Basel
Koeln
NuernbergWittenberg
0
.2
.4
.6
.8
1
Pre
dict
ed P
rote
stan
t 151
7-16
00
0 .2 .4 .6 .8 1Share German Language 1517-1600
Protestant 1517-1600 vs. Post-German
Figure 11: Share of books printed in German and post-1517 Protestantism.
-.4
-.2
0.2
.4R
esid
ual P
rote
stan
tism
0 100 200 300 400 500 600 700Distance to Wittenberg in Km
Figure 12: Distance to Wittenberg and residual Protestant print media 1517-1600 con-ditional on share of pre-1517 printing in German. Marker scales reflect overall numberof titles produced in each city.
25
Table 13 shows that the intensity of Protestant print media 1517-1600 and city-level
pre-characteristics is predicted in the cross section by pre-1517 printing of vernacular
texts, the presence of a university and distance to Wittenberg. In all specifications we
also control for the number of monasteries and convents at the city level as well as
indicators capturing whether each city was an Imperial Free City and/or a member of
the Hanseatic league.41Protestant 1517-1600 on Covariates as of 1517 (Cross Section)
Variable Ln Protestant Ln Protestant Ln Protestant Ln Protestant Ln Protestant Ln Protestant
(1) (2) (3) (4) (5) (6) (7)
Books pre-1517 0.07 0.02 0.05 0.03 0.05 0.03
(0.05) (0.04) (0.05) (0.04) (0.04) (0.04)
Books in German pre-1517 0.39** 0.59*** 0.50*** 0.55*** 0.51*** 0.55***
(0.16) (0.13) (0.14) (0.12) (0.14) (0.11)
University in 1517 1.65** 2.47*** 2.14*** 2.22*** 1.86*** 1.76***
(0.80) (0.89) (0.58) (0.79) (0.53) (0.58)
Distance to Wittenberg -0.28** -0.66***
(0.13) (0.24)
Ln Distance to Wittenberg -0.66*** -1.24***
(0.16) (0.24)
Principality Fixed Effects Yes Yes Yes
Observations 118 118 118 118 118 118
Figure 13: Cross sectional determinants of Protestant print media 1517-1600. Books aremeasured in hundreds. Specifications with principality fixed effects cluster standard er-rors at the principality level. All specifications control for imperial free city and HanseaticLeague status and the number of monasteries and convents at the city level in each ofthe Augustinian, Benedictine, Carmelite, Cistercian, Dominican, and the Brethren of theCommon Life religious orders.
Figure 14 exploits data from the pre-1517 period to estimate the difference-in-differences
impact of distance from Wittenberg on Protestant-type print media. 14 treats all texts
using the characteristic language of Protestantism as “like Protestant” to test whether
this language was more or less frequently observed closer to or farther from Wittenberg
before and after 1517.
5.2.2 Firms
Figure 15 shows that firms that were younger in 1517, and firms that were larger in
1517, produced more intensively Protestant-type print media 1518-1528. The dependent
variable for the regressions in Figure 15 is the firm-level religion index for output produced
41We control separately for the number of monasteries and convents in each of the Augustinian,Benedictine, Carmelite, Cistercian, Dominican, and the Brethren of the Common Life religious orders.
26
Protestant Difference in Differences
Variable Share Protestant Share Protestant Ln Protestant Ln Protestant
(1) (2) (3) (4) (5)
Ln Distance to Wittenberg 0.00 . -0.14 .
(0.01) . (0.22) .
Ln Distance to Wittenberg x Post-1517 -0.06*** -0.06** -0.61*** -0.53**
(0.02) (0.03) (0.19) (0.25)0.0% 0.0%
Post-1517 0.55*** 0.55*** 6.59*** 6.22***
(0.11) (0.16) (1.05) (1.31)
City Fixed Effects Yes Yes
Observations 100 100 94 94
Figure 14: Distance to Wittenberg and Protestant-Type Media Before and After 1517.The sample comprises the balanced panel of German-speaking cities with print mediapublished before 1450-1516 and 1517-1600. Standard errors clustered at the city level.
1518-1528.42 Columns (2) to (4) show the correlations between firm age and firm size, on
the one hand, and religious output, on the other, where the Religion Index is constructed
using only known authors. In columns (5) to (7) the dependent variable is the predicted
Religion Index estimated above, and the correlations are larger in magnitude but less
precisely estimated. Columns (3), (4), (6), and (7) introduce city fixed effects, so that
the identifying variation is across firms in a given city. The finding that firm age is
negative correlated with adoption of innovations in religious thought is consistent with
historians’ observations on the relationship between youth and adoption of innovations.43
42Values close to 0 tell us that at the firm level, content is characterized by (most closely resembles)the language of Catholic activitsts. Values close to 1 tell use that firm output uses the language ofProtestant Reformers.
43Cameron (1991, p. 186) observes that Reformation ideas were typically rejected by the scholars andthinkers who had taught the reformers, and that almost all academics in Wittenberg who were older thanLuther resisted the reformers’ innovations. Cole (1981, p. 141) similarly notes: “During the early yearsof the sixteenth century, the readiness of many people to accept and use the new medium of printingdepended in part on one’s chronological age.”
27
Observed Religion Index Predicted Religion IndexProtestant Protestant Protestant Protestant Protestant Protestant
Regression Variables Share Share Share Share Share Share(1) (2) (3) (4) (5) (6) (7)
Firm Age at 1517 -0.004* -0.006** -0.006** -0.005 -0.008* -0.010(0.002) (0.002) (0.002) (0.003) (0.004) (0.008)
Ln Books Lagged 0.027* 0.038** 0.042 0.025* 0.058*** 0.036(0.014) (0.015) (0.025) (0.014) (0.020) (0.065)
City FE Yes Yes Yes YesFirm Product Mix Yes YesObservations 88 88 88 88 88 88
Figure 15: Firm-level regressions documenting correlates of Religion Index 1518-1528.The dependent variable is the firm-level average of the Religion Index (1= Protestant,0=Catholic). Analysis restricted to firms that produced in the 10 year period beforeLuther posted his theses (i.e. 1507-1517) and in the 10 year period afterwards (1518-1528). Standard errors are clustered at the city level. “Firm Product Mix” controls forthe share of output 1507-1517 in each of 37 subject categories.
6 Outcomes
In this section we document how the diffusion of Protestant ideas was associated with
economic outcomes. Specifically, we document the diffusion of non-religious ideas in print
media, institutional change, and religious and economic outcomes observed in the 1800s.
6.1 Protestant Media and Institutional Change
In this section we present evidence on the print media and the laws of the Protestant
Reformation and document the variation in laws and print media induced by distance
to Wittenberg. Figure 16 shows the number of cities passing their first Reformation law
and the total number of Reformation laws passed each year.
Figure 17 exploits data from the pre-1517 period to estimate the cross-sectional pre-
characteristic determinants of the adoption of the legal institutions of the reformation.
Consistent with Becker and Woessman (2009), distance to Luther is a significant correlate
of adoption – even with polity fixed-effects. Other pre-characteristics have little if any
significant relationship to whether or not cities adopted the laws of the Reformation.
Finally, in Figure 18 we also show the relative intensity of Protestant-type language in
cities that did and did not adopt city-level Reformation laws. Particularly notable is the
fact that until the mid-1520s, and thus including the first years of the Reformation, the
cities that did and did not adopt the legal institutions of the Reformation were producing
28
0
2
4
6
8
Law
s P
asse
d P
er Y
ear
1500 1520 1540 1560 1580 1600 1620 1640
Cities Passing First Reformation LawTotal Reformation Laws (5 Year Moving Average)
Figure 16: The city-level laws of the Reformation.Protestant 1517-1600 on Covariates as of 1517 (Cross Section)
Legal Legal Legal Legal Legal Legal
Variable Reform Reform Reform Reform Reform Reform
(1) (2) (3) (4) (5) (6) (7)
Books pre-1517 0.03** 0.00 0.02 0.01 0.02* 0.00
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Books in German pre-1517 -0.00 0.04** 0.01 0.04 0.01 0.04*
(0.03) (0.02) (0.03) (0.03) (0.03) (0.02)
University in 1517 -0.10 0.05 -0.05 0.02 -0.17 -0.04
(0.16) (0.13) (0.13) (0.13) (0.16) (0.12)
Distance to Wittenberg -0.15*** -0.16**
(0.03) (0.07)
Ln Distance to Wittenberg -0.27*** -0.17
(0.07) (0.11)
Principality Fixed Effects Yes Yes Yes
Observations 137 137 137 137 137 137
Figure 17: City-level adoption of Reformation laws. Linear probability model with bi-nary dependent variable capturing whether or not a given city adopted an evangelischekirchenordnungen. Books are measured in hundreds. Specifications with principalityfixed effects cluster standard errors at the principality level. All specifications controlfor imperial free city and Hanseatic League status and the number of monasteries andconvents at the city level in each of the Augustinian, Benedictine, Carmelite, Cistercian,Dominican, and the Brethren of the Common Life religious orders.
29
relatively similar print media as measured by our religion index. From the 1520s, a gap
opens and stabilizes. Impressionistically, this is suggestive of a story in which relatively
modest differences in exposure were associated with sorting into different equilibria.
0
.2
.4
.6
.8
1R
elig
ion
Inde
x (5
Yea
r M
A)
1475 1500 1525 1550 1575 1600
Cities with Laws Cities without Laws
Figure 18: The relative intensity of Protestant-type language in religious media.
6.2 Protestant and Non-Religious Media
Within cities, exposure to Protestant media was associated with increased production
of the specific literatures historians and economists have identified as crucial to the ac-
quisition of human capital and city growth in early modern Europe. In particular, we
show that increases in Protestant media were associated with increases in the production
of educational books designed to impart basic literacy and of the merchant’s manuals
that were the key texts in Renaissance business education and which Dittmar (2013)
were uniquely associated with city growth in early modern Europe. We also show that
increases in Protestant media were associated with increases in non-religious titles pub-
lished in German vernacular. Historical evidence indicates that the increase in books in
the vernacular was critical to the emergence of a practical, literate, bourgeois lay culture
that supported relatively open access to ideas and innovation in business and industry
(Chrisman 1981).
30
Our findings both support and contrast with the central findings in the economics
literature relating religion, and in particular Protestantism, to economic outcomes. The
economics literature has documented that Protestantism is associated with economic
attitudes (Guiso et al. 2003), literacy (Becker and Woessman 2009), and political insti-
tutions (Woodberry 2010). This literature typically takes religion as effectively fixed or
sufficiently slow moving to identify variations in economic outcomes.44 In contrast, our
results characterize a setting in which the signal in religious media and religious beliefs
were changing dramatically. We show that during this period of relatively rapid religious
change, variations in the exposure to religious media within and across cities were as-
sociated with variations in the diffusion of the content historians have tied to economic
growth and development in early modern Europe.
In this section we present our findings on the incidence of Protestant media production
and the production of non-religious media. Motivated, by the historical literature on
economic development in early modern Europe, we present evidence for four key outcome
variables: the production of education titles, the production of legal texts on the theory
and study of jurisprudence, the production merchant manuals, and the production of all
non-religious titles published in German vernacular.45 In the data, our outcome variables
are all positive integer counts. We describe the estimator chosen for this setting and then
present the results of the estimation.
6.2.1 Estimation
The counts of types of media output across cities i over time t are modeled as Poisson
distributed random variables.46 The following exponential model specifies the conditional
mean of the Poisson model:
E(yit|xit) = exp(x′itβ)
Unobserved heterogeneity is modelled by specifying multiplicative individual effects:
E(yit|xit, αi) = exp(x′itβ + δi)
= λitαi
44Gentzkow and Shapiro (2010) notably argue that religiosity is effectively exogenous with respect tocontemporary media markets in the contemporary USA.
45In our data, ordinances themselves are classified separately from law books. The law books classi-fication we employ comprises jurisprudence including legal texts books, handbooks, and commentaries,but excludes edicts. In our data, printed copies of laws are classified as ordinances, together with edictsand proclamations.
46It is not uncommon to use a continuous approximation in place of the Poisson for applications inwhich the counts take on large values. However, when the counts are close to and include zero, thePoisson is better able to capture the qualities of the data.
31
Where we let:
λit = exp(x′itβ)
and
αi = exp(δi)
Because unobserved individual effects αi may be correlated with observed city-level data
xit such that E(xitαi) 6= 0, we choose to employ the standard fixed-effect Poisson model
of Hausman et al. (1984) (“HHG”) with an exponential conditional mean function and
multiplicative individual specific effects as described above.
The conditional probability of observing a count yit for a city i in a given period t
may be written
Pr(yit|xit, αi) =e−αiλit(αiλit)
yit
yit!
This leads to the following log likelihood function:
lnL = −αi∑t
λit + lnαi∑t
yit +∑t
yit lnλit −∑t
ln Γ(yit + 1)
where ln Γ is the log gamma function.
Following HHG, start from the assumptions that yit is conditionally Poisson dis-
tributed and that yit and yis are independent conditional on xit and αi. This implies
strict exogeneity:
E[yit|xi1, . . . , xiT , αi] = αi exp(x′itβ)
HHG develop a conditional maximum likelihood estimator using the approach of An-
dersen (1970) to eliminate αi from the likelihood, conditioning on its sufficient statistic∑t = yit in order to obtain a consistent estimator for β.
It has since been shown that the Poisson estimator does not suffer from the incidental
parameters problem so that the HHG estimator is equivalent to a maximum likelihood
estimator with individual specific effects. This estimator can also be obtained under
a single first moment condition by replacing the fixed effects with a ratio of within
group means (Blundell, Griffith, and Windmeijer, 2002; Cameron and Trivedi, 1998;
Wooldridge, 1999) such that ∑i
∑t
xit(yit − λityiλi
) = 0
These results imply that the only requirement for the consistency of our estimator is that
of strict exogeneity. This alone is enough to ensure a robust, consistent estimator of β
32
(Wooldridge 1999). We do not need any assumptions about the distribution or about
temporal dependence of yit. We only require that the conditional mean be correctly
specified. This means that our estimator is robust to departures from the Poisson distri-
bution, including its requirement that there is no underdispersion or overdispersion. If
the distributional and temporal independence assumptions do not hold, it is necessary
to use a robust estimator for the variance of β. All results contained in the next section
use the robust covariance estimate of Wooldridge (1999).
In addition to the conditional likelihood model for panel data, we use as a baseline
model a pooled Poisson quasi-maximum likelihood estimator (QMLE). We do this for the
following reasons. First, the QMLE estimator is also consistent under the assumption
that the conditional first moment is correctly specified. This follows from Gourieroux et
al.’s (1984) result on robustness of QMLE estimators in the linear exponential family.
Second, it does not require an assumption of strict exogeneity. Third, there is no restric-
tion on the time dependence of the observations. Finally, we can include time-invariant
regressors and regressors with a small variance relative to that of yit in xit – which allows
us additional insight into what drove the locational effects characterizing the diffusion of
ideas in print in the 1500s.
6.2.2 Results
Here we present the results of our specifications for the pooled Poisson without city fixed
effects and the fixed effects Poisson estimators. Two models are fit for each dependent
variable within each specification—one without time dummies and one with time fixed
effects, where time is measured in five year periods starting in 1517.47 In each case we
prefer the one with time dummies on the basis a Wald test on the joint hypothesis that
the dummy variable coefficients are zero. Since the conditional mean of the Poisson
is an exponential, the coefficient on the log transformed variables can be interpreted
as elasticities. Standard errors are reported in parantheses. All standard errors are
corrected for clustering at the city-level and are robust to serial correlation following
Wooldridge (1999). The results documenting the relationship between Protestant media
and the production of educational and non-religious vernacular media are robust over
the specifications.
The results in Table 6 are for the baseline pooled Poisson model without city fixed
effects. Both Catholic and Protestant media are associated with a positive increase in
the production of education works. The coefficients on both are not statistically different
47We collapse our annual-level data to 5 year periods to reduce sparseness.
33
Table 6: Baseline Poisson Model
educ law merchant vernaclnprot 0.58** 0.63** 0.45** 0.50** 0.92** 0.95** 0.76** 0.78**
(0.07) (0.08) (0.06) (0.07) (0.09) (0.09) (0.11) (0.11)
lncath 0.61** 0.63** 0.84** 0.81** 0.13 0.19 0.14 0.16*(0.08) (0.07) (0.04) (0.04) (0.12) (0.11) (0.10) (0.09)
time dummies no yes no yes no yes no yesN 2176 2176 2176 2176 2176 2176 2176 2176χ2lnprot=lncath 0.06 0.00 17.05 10.37 17.89 19.30 9.51 10.27ρ 0.21 0.20 0.24 0.22 0.19 0.18 0.55 0.58
Robust standard errors in parentheses
* p < 0.10, ** p < 0.05, ** p < 0.01
from one another as indicated by the χ2 statistic resulting from Wald test that coefficients
on lnprot and lncath are equal. This test statistic is distributed χ21 with a critical value
of 3.84 at the 95% confidence level. In the case of law books, Catholic media is associated
with a bigger effect than Protestant though both are statistically significant. The presence
of Protestant works is strongly associated with an increase in the number of merchant
manuals produced. A doubling of Protestant works leads to nearly a doubling of merchant
manuals. Catholic works do not have a statistically significant effect here. The same can
be said for the production of all non-religious works written in the vernacular German.
The production of Protestant works is overwhelming associated with a higher count of
these vernacular works.
Table 7 presents results from a pooled Poisson specification that includes covariates
that account for city-level institutional factors that could plausibly affect the production
of other media and might otherwise be conflated with the effects of Catholicism and
Protestant media. The institutions we control for are the presence of a university in the
city and the enactment of the first of Protestant reform laws, which marked the first
official approval of the Reformation and thus inaugurated the “decisive, final step in the
Protestant movement” (Ozment 1975, p. 125). These are two absorbing state variables
that become and stay equal to one for a given city when that city obtains the relevant
institutions.
The point estimates with the addition of these two variables are broadly similar to
our baseline results. However, there are some notable differences. For education works,
we still cannot statistically discern the effects due to Protestant and Catholic media. It
is evident, however, that the enactment of the Protestant reform law is associated with
an environment that leads to more education books being published. In fact, this effect
34
Table 7: Pooled Poisson Model
educ law merchant vernaclnprot 0.51** 0.55** 0.39** 0.45** 0.73** 0.74** 0.66** 0.70**
(0.06) (0.06) (0.07) (0.06) (0.10) (0.10) (0.10) (0.09)
lncath 0.67** 0.67** 0.73** 0.73** 0.26** 0.30** 0.32** 0.34**(0.06) (0.06) (0.07) (0.06) (0.08) (0.09) (0.08) (0.08)
reform law 0.48** 0.49** 0.37* 0.15 1.61** 1.56** 0.66** 0.56**(0.13) (0.12) (0.19) (0.15) (0.34) (0.35) (0.21) (0.22)
university 0.01 0.05 0.92** 0.61** 0.16 0.12 -0.51 -0.59(0.17) (0.19) (0.24) (0.22) (0.38) (0.40) (0.34) (0.37)
time dummies no yes no yes no yes no yesN 2176 2176 2176 2176 2176 2176 2176 2176χ2lnprot=lncath 2.16 1.29 7.04 7.73 7.84 6.21 4.35 5.10ρ 0.18 0.17 0.22 0.20 0.14 0.14 0.48 0.49
Robust standard errors in parentheses
* p < 0.10, ** p < 0.05, ** p < 0.01
is present across all outcome variables. The only exception being the specification for
law books with time fixed effects included. The production of legal texts is strongly
associated with the presence of a university in a city, consistent with the fact that uni-
versities housed law faculties and produced jurists. Legal output is strongly correlated
with Catholic media production as most universities in 16th century Germany were and
remained predominantly Catholic.48
The row labeled ρ in Tables 6 and 7 reports the mean of the unique off-diagonal
estimates of serial correlation developed in HHG. If the correct specification is one in
which city-level effects are indeed important, then the residuals will exhibit serial corre-
lation. To compute ρ, we take the generalized (Pearson) residuals of the pooled model,
εit = (yit − λit)/√
(λit) and obtain the correlation coefficient between εit and εis, esti-
mated as∑
i εitεis√∑i ε
2it
√∑ε2is
. Under the null of no individual effects, we expect the correlation
coefficient to be zero for t 6= s. These estimates are all positive, indicating that some
small degree of serial correlation is likely present. However, if the serial correlation is due
to unchanging unobserved city-level effects αi, we would expect that ρ is fairly constant
(t−s). This is a slighty tougher case to make. The standard deviation of the correlations
is about 0.2 for each dependent variable.
Table 8 contains the results for the specification with city fixed effects. The point
48While the universities at Wittenberg and Basel became centers of reformist thought, the universitiesat Koln and Louvain developed as centers of Catholic counter-reformation thinking.
35
Table 8: Fixed effects Poisson model
educ law merchant vernaclnprot 0.32** 0.44** 0.40** 0.42** 0.58** 0.62** 0.41** 0.49**
(0.07) (0.08) (0.07) (0.09) (0.11) (0.10) (0.06) (0.06)
lncath 0.24** 0.26** 0.74** 0.60** -0.00 0.07 0.08** 0.11**(0.06) (0.06) (0.09) (0.07) (0.09) (0.12) (0.04) (0.03)
reform law -0.00 0.08 1.40** 0.53 1.09** 1.13** 0.26* -0.19(0.17) (0.21) (0.37) (0.36) (0.43) (0.49) (0.14) (0.15)
time dummies no yes no yes no yes no yesN 1232 1232 1200 1200 512 512 1712 1712χ2lnprot=lncath 0.63 2.41 6.69 1.85 11.67 11.24 14.36 29.39
q 158.93 49.90 84.55 67.38 118.66 23.24 92.76 105.68
Robust standard errors in parentheses
* p < 0.10, ** p < 0.05, ** p < 0.01
estimates are stable with the exception of the education specification. Here, the interpre-
tation is reversed with higher Protestant media output associated with the publication
of more educational books. However, there is still not enough evidence to say that the
elasticities of educational print media with respect to Protestant and Catholic media
are statistically different from one another. It should be pointed out that we now have
varying sample sizes for each model. This is due to the fact that the fixed effects Poisson
drops cases for which∑
i yit = 0 as there is no sufficient statistic for αi in this case.
To help determine the preferred model specification, we check for serial correlation.
Since the conditional MLE does not provide consistent individual effects αi, we cannot
easily compute residuals yit − αiλit. HHG (1984) develop a test of serial correlation for
the fixed effects Poisson model. Based on the fact that yi1, . . . , yiT |∑
t yit is multinimial-
distributed with probability pit = λit∑s λis
, they introduce a conditional moments test based
on the usual moment conditions of a multinomial-distributed random variable. Here the
test statistic is q ∼ χ2105 with critical value 129.92 at a confidence-level of 95%. The only
specification that exhibits significant serial correlation is that of the educational books in
the case without time dummies. However, with the addition of time dummies this ceases
to be the case.
Finally, we consider the elasticity of media production with respect to distance to
Wittenberg over the period before and after the Reformation. We do this by estimat-
ing our baseline Poisson model with interactions that capture the effect of distance to
Wittenberg on media production in each five year time period. Figure 19 documents
that distance to Wittenberg was not associated with any significant difference in book
36
production pre-1517 and was systematically post-1517. It also shows that places farther
from Wittenberg produced systematically less educational books from the mid-1500s on-
wards. These facts are documented in Figure 19 by plotting the estimated elasticity of
print media production with respect to distance to Wittenberg as it evolves over time.
-1.5
-1
-.5
0
.5
1
1.5
2
2.5
Dis
tanc
e to
Witt
enbe
rg E
last
icity
1500 1520 1540 1560 1580 1600
Total Books
-1.5
-1
-.5
0
.5
1
1.5
2
2.5
Dis
tanc
e to
Witt
enbe
rg E
last
icity
1500 1520 1540 1560 1580 1600
Educational Books
-1.5
-1
-.5
0
.5
1
1.5
2
2.5
Dis
tanc
e to
Witt
enbe
rg E
last
icity
1500 1520 1540 1560 1580 1600
Vernacular Books
Figure 19: Elasticity of print media with respect to distance to Wittenberg.
6.3 Institutions, Media, and Long-Run Outcomes
This section documents the relationship between educational outcomes in the 1800s and
exposure to the legal institutions of the Protestant Reformation. The key new fact we
present is that populations in counties where Reformation ordinances were passed in
the 1500s were more literate in the 1800s. This holds conditional on variations in the
religious composition of the population, which is not a significant correlate of literacy
once we control for legal institutions. We also show that while variations in both religious
composition and literacy in the 1800s are explained by exposure to the institutions of
the Reformation, they are not explained by our measure exposure to the message of
the Reformation in historical print media. These findings provide new evidence on the
channels through which the Reformation may have impacted development, and suggest
an institutional underpinning for the human capital impacts of religious change.49
49An extensive literature ties religion to economic behavior and outcomes (e.g. Barro and McCleary2005). A number of recent economic analyses examine Protestantism and economic outcomes in German-speaking Europe. Davide Cantoni (2009) finds that Protestant cities enjoyed no advantage in populationgrowth in the Holy Roman Empire. Becker and Woessman (2009) find that Protestantism was associatedwith higher levels of literacy in 19th century Prussia and argue that Protestantism operated as a demandshift for human capital that subsequently facilitated the adoption of new technology and economicgrowth. Basten and Betz (2013) find exposure to Calvinist Protestantism reduces preferences for leisure,redistribution, and government intervention along the border between Calvinist Protestant and Catholicregions of Switzerland.
37
Our findings suggest an institutional interpretation of the key fact that Protestantism
was associated with higher literacy in the 1800s across Prussian (German) counties.
Becker and Woessman (2009) propose a causal interpretation of the observed relationship
between Protestantism and literacy, using distance to Wittenberg as a source exogenous
variations in Protestantism. Our evidence qualifies the relationship between religioun
and education observed in the 1800s. We document that specific historical institutions
established by the Reformation explain variations in literacy observed in the 1800s, while
exposure to Reformation ideas in the 1500s and religious denomination in the 1800s itself
do not.
We measure exposure to religious laws as follows. For all city or municipal level
Reformation ordinances in Sehling (1910) we identify the subsequent Prussian county
locations of the city or municipality.50 This identifies a set of counties with urban or
city-level Reformation ordinances. For all territorial Reformation ordinances in Sehling
(1910) we proceed in two steps. We identify and geocode the extent of each polity
(principality) with a territorial law. We then overlay the map of Prussian counties from
the 1800s on the historic map of territorial ordinances to determine which counties were
treated by which ordinances. In cases, where the territory of an ordinance covers part
but not all of a particular county, we assign treatment on a proportional basis. For
example, we observe a territorial ordinance in Electoral Saxony. The 1849 boundaries of
the county (kreis) of Sagan show that 72% of the county is inside Electoral Saxony, so we
assign a treatment value of 0.72 to Sagan for any ordinance passed in Electoral Saxony.
We map city-level print media exposure to Prussian counties in the 1800s in the same
manner.
Table 20 documents the positive relationship between the legal institutions of the
reformation and county-level literacy data from 1871 (Becker, Woessman, and Hornung
2006).51 Column (2) shows the baseline correlation between Protestantism and literacy
across 334 Prussian counties. Column (3) introduces our indicator for city-level laws
institutionalizing the Reformation. In column (3) we show that when we also include our
measure of territory-level exposure to the legal institutions of the Reformation, Protes-
tantism in the 1800s ceases to be a significant predictor of literacy in the 1800s. Column
(5) shows that exposure to Protestant print media in the 1500s does not positively pre-
dict literacy in the 1800s. Column (6) shows that non-religious, German language print
50For cities that produced print media 1450-1600, we similarly associated them with the counties theywere subsequently located in.
51We use the set of counties with consistent data available in all periods. Thus we exclude countiesthat appear in the data in the 1870s due to Prussian expansion. Following Becker and Woessman (2006),we use religion as of 1811 in order to take an observation on religion before the industrial revolution ofthe 1800s took off.
38
media of the 1500s does predict literacy in the 1800s. Finally, column (7) shows that the
relationship between historic exposure to German-language non-religious media and sub-
sequent literacy is much larger and is only statistically significant in the post-reformation
(1517-1600) era. Literacy, Religion, Institutions, Print Media -- Cross Section Prussian Counties
Literacy Literacy Literacy Literacy Literacy Literacy
Variable in 1871 in 1871 in 1871 in 1871 in 1871 in 1871
(1) (2) (3) (4) (5) (6) (7)
Share Protestant 1811 0.06*** 0.05*** 0.03 0.03 0.03 0.03
(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
City Reformation Law 0.07*** 0.06*** 0.06*** 0.05*** 0.05***
(0.02) (0.01) (0.02) (0.02) (0.02)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Territorial Reformation Law 0.05*** 0.05*** 0.05*** 0.05***
(0.02) (0.02) (0.02) (0.02)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Protestant Books 1517-1600 0.01 -0.01** -0.01*
(0.01) (0.00) (0.00)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1454-1600 0.17***
(0.05)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1454-1516 0.11
(0.53)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1517-1600 0.17**
(0.08)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Observations 334 334 334 334 334 334
Figure 20: Literacy across Prussian counties in 1871.Religion, Distance to Wittenberg, Institutions, Print Media -- Cross Section Prussian Counties
% Protestant % Protestant % Protestant % Protestant % Protestant % Protestant
Variable in 1811 in 1811 in 1811 in 1811 in 1811 in 1811
(1) (2) (3) (4) (5) (6) (7)
Distance to Wittenberg -0.79*** -0.75*** -0.12 -0.12 -0.13 -0.11
(0.15) (0.15) (0.19) (0.19) (0.19) (0.19)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
City Reformation Law 0.18*** 0.15*** 0.19*** 0.20*** 0.18***
(0.06) (0.06) (0.05) (0.05) (0.05)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Territorial Reformation Law 0.39*** 0.39*** 0.39*** 0.39***
(0.06) (0.06) (0.06) (0.06)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Protestant Books 1517-1600 -0.04 -0.02 -0.05***
(0.03) (0.03) (0.01)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1454-1600 -0.22
(0.31)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1454-1516 -10.28***
(1.29)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Non-Religious German Books 1517-1600 0.76***
(0.18)0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Observations 334 334 334 334 334 334
Figure 21: The share of Protestants across Prussian counties in 1811.
Our interpretation of these facts is that the religious denomination of citizens recorded
by census enumerators in the 1800s tells us much less about variations in literacy than the
long effects of laws that formalized the institutions of the reformation.52 Our finding is
52The census classes people as Protestant, Catholic, Jewish, or “other christian religion.” See iPEHD(2013).
39
thus consistent with the argument in Becker and Woessman (2009) that “Protestantism”
was associated with variations in educational attainment.
Because distance from Wittenberg provides a source of plausibly exogenous varia-
tion in exposure to Protestantism, we examine the importance of distance in predicting
Protestantism in the early 1800s. Table 21 shows that distance to Wittenberg is a highly
significant predictor of variations in county-level Protestantism – before we control for
exposure to city and territorial laws. Columns (4) to (7) show that once we control for
exposure to Reformation ordinances, distance to Wittenberg ceases to be a significant
predictor of variations in religion at the county level. This finding does not mean that
distance to Wittenberg only impacted subsequent religion through the institutional chan-
nel. But it provides prima facie evidence consistent with narrative evidence documenting
the importance of Protestant educational institutions. A caveat is in order on these re-
sults: We are currently exploring the spatial structure of the data to determine whether
there is robust evidence of spatial spillovers. The current estimates abstract from spatial
questions except insofar as distance from Wittenberg is concerned.
7 Conclusion
The print media revolution dramatically lowered the cost of transmitting ideas. The
Protestant Reformation was the first mass movement to exploit the new media. Previous
research has not documented the diffusion of Protestantism in quantitative terms.
We use new data on the universe of known books and pamphlets printed in Germany
to document the diffusion of Protestant ideas across time and space. We find that the
diffusion of Protestant ideas in the 1500s was associated with increased access to edu-
cational and non-religious vernacular in the 1500s, institutional change, and ultimately
higher rates of literacy by the 1800s.
References
[1] Erling Bernhard Andersen. Asymptotic properties of conditional maximum-
likelihood estimators. Journal of the Royal Statistical Society. Series B (Method-
ological), pages 283–301, 1970.
[2] Christoph Basten and Frank Betz. Beyond work ethic: Religion, individual, and
political preferences. American Economic Journal: Economic Policy, 5(3):67–91,
40
2013.
[3] Sascha O. Becker, Francesco Cinnirella, Erik Hornung, and Ludger Woessmann.
ipehd - the ifo prussian economic history database, cesifo working paper no. 3904.
August 2012.
[4] S.O. Becker, E. Hornung, and L. Woessmann. Education and catch-up in the indus-
trial revolution. American Economic Journal: Macroeconomics, 3(3):92–126, 2011.
[5] S.O. Becker and L. Woessmann. Was weber wrong? a human capital theory of
protestant economic history. The Quarterly Journal of Economics, 124(2):531–596,
2009.
[6] S.O. Becker and L. Woessmann. The effect of protestantism on education before the
industrialization: Evidence from 1816 prussia. Economics Letters, 107(2):224–228,
2010.
[7] Albert Bifet and Eibe Frank. Sentiment knowledge discovery in twitter streaming
data. In Discovery Science, pages 1–15. Springer, 2010.
[8] Richard Blundell, Rachel Griffith, and Frank Windmeijer. Individual effects and
dynamics in count data models. Journal of Econometrics, 108(1):113–131, 2002.
[9] Johan Bollen, Huina Mao, and Alberto Pepe. Modeling public mood and emotion:
Twitter sentiment and socio-economic phenomena. In ICWSM, 2011.
[10] A Colin Cameron and Pravin Trivedi. Regression analysis of count data, volume 53.
Cambridge University Press, 2013.
[11] E. Cameron. The European Reformation. Clarendon Press, 1991.
[12] Davide Cantoni. The economic effects of the protestant reformation: Testing the
weber hypothesis in the german lands. 2010.
[13] Davide Cantoni. Adopting a new religion: The case of protestantism in 16th century
germany. Economic Journal, 122(560):502–531, 2012.
[14] P.W. Carey and J.T. Lienhard. Biographical Dictionary of Christian Theologians.
ABC-Clio ebook. Greenwood Press, 2000.
[15] R.A. Crofts. Printing, reform, and the catholic reformation in germany (1521-1545).
The Sixteenth Century Journal, 16(3):369–381, 1985.
[16] J. De Vries. European Urbanization: 1500-1800. Methuen, 1984.
41
[17] J. Dittmar. New media, firms, ideas, and growth: European cities after gutenberg.
[18] J.E. Dittmar. Information technology and economic change: the impact of the
printing press. The Quarterly Journal of Economics, 126(3):1133–1172, 2011.
[19] Jr. Edwards, Mark U. Lutheran pedagogy in reformation germany. History of
Education Quarterly, 21(4):pp. 471–477, 1981.
[20] M.U. Edwards Jr. Catholic controversial literature, 1518-1555: Some statistics.
Archiv fur Reformationsgeschichte, 79, 1988.
[21] Robert B Ekelund Jr, Robert F Hebert, and Robert D Tollison. An economic analysis
of the protestant reformation. Journal of Political Economy, 110(3):646–671, 2002.
[22] Richard Gawthrop and Gerald Strauss. Protestantism and literacy in early modern
germany. Past & Present, (104):pp. 31–55, 1984.
[23] Matthew Gentzkow and Jesse M. Shapiro. What drives media slant? evidence from
u.s. daily newspapers. Econometrica, 78(1):35–71, 2010.
[24] Alec Go, Richa Bhayani, and Lei Huang. Twitter sentiment classification using
distant supervision. CS224N Project Report, Stanford, pages 1–12, 2009.
[25] Christian Gourieroux, Alain Monfort, and Alain Trognon. Pseudo maximum like-
lihood methods: applications to poisson models. Econometrica: Journal of the
Econometric Society, pages 701–720, 1984.
[26] Christian Gourieroux, Alain Monfort, and Alain Trognon. Pseudo maximum likeli-
hood methods: Theory. Econometrica: Journal of the Econometric Society, pages
681–700, 1984.
[27] M. Greengrass. The Longman companion to the European Reformation, c. 1500-
1618. Longman companions to history. Longman, 1998.
[28] Luigi Guiso, Paola Sapienza, and Luigi Zingales. People’s opium? religion and
economic attitudes. Journal of monetary economics, 50(1):225–282, 2003.
[29] Jerry Hausman, Bronwyn H Hall, and Zvi Griliches. Econometric models for count
data with an application to the patents-r & d relationship. Econometrica, 52(4):909–
938, 1984.
[30] W. Klaiber. Katholische kontroverstheologen und reformer des 16. Jahrhunderts
(Reformationsgeschichtliche Studien und Texte 116), 136, 1978.
42
[31] H.J. KOhler. Flugschriften als Massenmedium der Reformationszeit: Beitrage zum
Tubinger Symposion 1980. Klett-Cotta, 1981.
[32] Jimmy Lin and Alek Kolcz. Large-scale machine learning at twitter. In Proceedings
of the 2012 ACM SIGMOD International Conference on Management of Data, pages
793–804. ACM, 2012.
[33] J. Mark U. Edwards. Printing, Propaganda And Martin Luther. Augsburg Fortress
Pub, 2004.
[34] Rachel M McCleary and Robert J Barro. Religion and economy. The Journal of
Economic Perspectives, 20(2):49–72, 2006.
[35] M. Mullett. Historical Dictionary of the Reformation and Counter-Reformation.
Historical Dictionaries of Religions, Philosophies, and Movements Series. Scarecrow
Press, 2010.
[36] Steven E Ozmint. The Reformation in the Cities: The Appael of Protestantism to
Sixteenth Century Germany and Switzerland, volume 357. Yale University Press,
1975.
[37] Alexander Pak and Patrick Paroubek. Twitter as a corpus for sentiment analysis
and opinion mining. In LREC, 2010.
[38] Marco Pennacchiotti and Ana-Maria Popescu. A machine learning approach to
twitter user classification. In ICWSM, 2011.
[39] A. Pettegree. Reformation and the Culture of Persuasion. Cambridge University
Press, 2005.
[40] A. Pettegree. The Book in the Renaissance. Yale University Press, 2010.
[41] Jared Rubin. Printing and protestants: reforming the economics of the reformation.
Available at SSRN 1742523, 2011.
[42] R.W. Scribner. Popular Culture and Popular Movements in Reformation Germany.
History series. Hambledon Press, 1987.
[43] M. Taddy. Multinomial inverse regression for text analysis. Journal of the American
Statistical Association, (just-accepted), 2012.
[44] M.A. Taddy. Inverse regression for analysis of sentiment in text. arXiv preprint
arXiv:1012.2098, 2010.
43
[45] J. Witte and M.E. Marty. Law and Protestantism: The Legal Teachings of the
Lutheran Reformation. Cambridge University Press, 2002.
[46] Robert D Woodberry. Religion and the spread of human capital and political insti-
tutions christian missions as a quasi-natural experiment, 2011.
[47] Jeffrey M Wooldridge. Distribution-free estimation of some nonlinear panel data
models. Journal of Econometrics, 90(1):77–97, 1999.
44