Religious Diversity and Innovation: Historical Evidence from Patenting Activity

Transcription

1 Religious Diversity and Innovation: Historical Evidence from Patenting Activity Francesco Cinnirella Jochen Streb August 20, 2015 Abstract Religious norms related to specific denominations have been associated to economic outcomes. We argue that the religious composition of a population has also an impact on economic outcomes. In particular we argue that the presence of different cultures, skills, and abilities originating from different religions fosters innovation. Matching data on long-lasting patents with information on the religious composition of county population we find that religious diversity has an inverted-u shape relationship with patenting activity. The nonlinear effect is consistent with the notion of a tradeoff between costs and benefits of diversity. Separate instrumental variable estimates based on past religious diversity and on eastward territorial annexations of Prussia support a causal interpretation of our results. We find that the effect of religious diversity on innovation is relevant for large firms patents and in counties with a comparatively higher level of development. Keywords: Diversity, Fractionalization, Innovation, Patenting Activity JEL classification: O15, O31, R11, Z12, N33 DRAFT PREPARED FOR THE 2015 ANNUAL MEETING OF THE ECONOMIC HISTORY ASSOCIATION DO NOT CITE OR CIRCULATE WITHOUT PERMISSION We thank Philipp Ager, Davide Cantoni, Matteo Cervellati, Oliver Falck, Raphaël Franck, Oded Galor, Johannes Koenen, Lars Lønstrup, Stelios Michalopoulos, Alessandro Nuvolari, Ömer Özak, Giovanni Prarolo, Paul Sharp, Uwe Sunde, Fabian Waldinger, Dinand Webbink, David Weil, Ludger Woessmann, and seminar participants at the First German Economic History Congress, European Economic Association (EEA), Tinbergen Institute in Rotterdam, University of Southern Denmark, University of Munich, University of Bologna, Sant Anna School of Advanced Studies in Pisa, University of Mannheim, and the Ifo Institute in Munich for helpful comments. Corresponding author: Ifo Institute, CESifo, CEPR, and CAGE. cinnirella@ifo.de University of Mannheim. streb@uni-mannheim.de 1

2 1 Introduction There is a large and expanding literature which investigates the impact of religious norms on economic outcomes. Already at the beginning of the twentieth century Max Weber (1905) argued that hard work and thrift facilitated the rise of capitalism in Western Europe. Since then, numerous studies have attempted to identify religious traits which are directly or indirectly conducive to economic growth (Andersen et al., 2013; Becker and Woessmann, 2009, 2008; Botticini and Eckstein, 2005, 2007; Akçomak et al., 2015; Cantoni, 2014; Barro and McCleary, 2006, 2003; Hornung, 2014; Michalopoulos et al., 2014). In this paper we argue that the religious composition of the population has also an impact on economic outcomes. We argue that diversity in religious denominations brings about variety in cultures, skills, and abilities which positively affect the generation and adoption of new ideas (Alesina and La Ferrara, 2005). We test this hypothesis studying the impact of religious diversity on innovation during the second industrial revolution in nineteenth-century Prussia. Prussia provides an almost unique context of religious heterogeneity in a period of intense innovation such as the second industrial revolution. We study the effect of religious diversity on innovation merging county-level data on the religious composition of the population in 1871 with individual level data on longlasting patents for the period In this period Prussia was at the technological frontier in the chemical, electrical, and machine building sector (Pierenkemper and Tilly, 2004). Advancements in these sectors were based on the availability and use of scientific knowledge. Progress in scientific analysis lowered the effectiveness of secrecy and thus increased inventors dependency on patents (Moser, 2013). Furthermore, exploring the relationship between religious diversity and innovation for a single country significantly mitigates the problem of heterogeneous institutions which characterizes cross-country analysis. The availability of unique data on religious denomination and long-lasting patents at the micro-regional level allows us also to shed some light on the mechanisms that operates behind the relationship between religious diversity and innovation. Economic theory and empirical evidence suggest that there are also economic costs associated with cultural diversity. Coordination failure, mistrust, and heterogeneous preferences over the provision of public goods may lead to adverse economic outcomes (Alesina and La Ferrara, 2005; Alesina et al., 1999; Easterly and Levine, 1997). Therefore the existence of a trade-off between costs and benefits of diversity suggests that the relationship between religious diversity and innovation might be nonlinear. Indeed, our baseline estimates show that religious diversity has an inverted-u shape relationship with long-lasting patents. This relationship is robust to the introduction of a rich set of confounding factors such as the literacy rate, population size, urbanization, 2

3 the share of Protestants, and the share of people employed in sectors related to natural endowments and characterized by a high propensity to innovate and patent. Counties with a high economic potential are generally more open to different cultures and might therefore be destination of migratory flows generating a problem of reverse causality. There could also be omitted variables affecting both the religious composition of the population and innovation. To address the issue of causality we propose two independent instrumental variable strategies. In the first case we exploit information on the religious composition of the population in 1816, a pre-industrial period. The logic behind this instrumentation is that the religious fragmentation observed in 1816 is the result of a complex set of historical processes, started with the Protestant Reformation in 1517, which are in large part exogenous to the innovative activity of the second industrial revolution. By including a rich set of contemporary controls in 1816 and accounting for some of the determinants of the adoption of Protestantism in 1517, we provide significant support for the exclusion restriction. Instrumental variable estimates support the finding of a nonlinear effect of religious diversity on long-lasting patents. The second identification strategy exploits the eastward annexations by the Kingdom of Prussia since its foundation in 1701 which successively included a large number of people of slavic ethnic origin of Catholic and Jewish religion. In particular we predict religious diversity in 1871 with distance to the eastern border, allowing the relationship to vary across annexations. The logic behind this instrumentation is that distance to the eastern border by annexation has an effect on innovation only through religious diversity. Because of the nature of the instrument and in order to have enough variation in religious diversity, this identification strategy focuses only on the counties east of the river Elbe. This constitutes also an advantage as we discard from the analysis the western regions of Prussia with superior natural endowments, larger cities, and more liberal economic institutions which make difficult the identification of the causal effect. Instrumental variable estimates for the counties east of the river Elbe strongly support the finding of a nonlinear effect of religious diversity on innovation. We show that income differences and linguistic fractionalization do not explain our main finding. Geographic controls such as latitude, longitude, precipitation, and temperature ensure that our instrument does not capture geographic features that can directly impact innovation. Previous literature has suggested that diversity is more beneficial at higher levels of development (Alesina and La Ferrara, 2005). The idea is that the benefits of skill complementarities are realized when the production process is diversified, therefore at higher levels of development. Using the information on the name of the patent holder and the legal form of the company (e.g. AG, GmbH ) we can distinguish between large firms patents and independent patents. Consistent with the hypothesis that diversity is more beneficial at higher levels of development, we find that the effect of religious 3

4 diversity on innovation is significant only for large-firms patents. Similarly, the effect is significant for counties with a literacy rate and industrialization rate above the median. The paper is structured as follows: Section 2 discusses the related literature; Section 3 describes the data and introduces our proxy for innovation; Section 4 presents our baseline estimates; Section 5 addresses the issue of causality and introduces our two instrumentations; Section 6 discusses potential channels through which religious diversity might affect innovation; Section 7 concludes. 2 Related literature [Incomplete] The literature on the economic benefits of diversity has mainly focused on the effect of fractionalization on productivity and income. Alesina et al. (2013) use an index of diversity based on people s birthplace and find that it is positively related to economic development. It is important to note that their index of birthplace diversity is almost orthogonal to standard measures of ethnic and linguistic diversity. Peri (2012) finds positive effects of diversity on productivity of US states. He argues that the positive effect of diversity is due to the immigration of unskilled workers who promote specialization. Similarly, Ottaviano and Peri (2006) find that increase in the cultural fractionalization of cities has a positive effect on productivity. Ager and Brückner (2011) study the relationship between diversity and economic growth in historical context, namely in US in the age of mass migration ( ). Exploiting within-county variation they find a positive effect of fractionalization on economic growth. There is also an extensive literature on the economic costs associated with diversity. 1 Easterly and Levine (1997) shows that ethnic diversity adversely affects public policies associated with economic growth such as black market, low provision of infrastructure, and low levels of education. Alesina et al. (1999) finds that the provision of public goods such as education, roads, and sewers is inversely related to ethnic fragmentation in US cities. They argue that this can happen because different ethnic groups have different preferences over the typology of public good to produce and because each ethnic group s utility level for a given public good is reduced if other groups also use it. Our study differs from the existing literature in several aspects. Previous studies have mainly focused on the relationship between birth, ethnic, and linguistic fractionalization and economic outcomes. To our knowledge this is the first work that explores the direct effect of religious diversity on innovation. We deviate from previous studies as the observed diversity is not driven by immigration but is the result of a process started after the Protestant Reformation in 1517 when the rulers of the various principalities and 1 See Alesina and La Ferrara (2005) and Alesina et al. (2003) for a comprehensive review of the literature. 4

5 city-states within the Holy Roman Empire decided about the religious denomination to impose on their subjects (Cantoni, 2012). 3 Data 3.1 Patents We merge individual data on long-lasting patents granted in Prussia from 1877 to 1890 with county level data from 1871 and 1882 (Becker et al., 2014). The prime source for the patent database is the annual Verzeichnis der im Vorjahre erteilten Patente published by the German patent office since 1877, which lists all patents granted in the preceding year. From these periodical listings we keep all patents that were held for at least ten years, which therefore can be interpreted as valuable (Streb et al., 2006). The German patent law first introduced in 1877 was a system with periodical renewal fees (Seckelmann, 2006). The system allowed both private persons and firms to apply for patent protection that could last up to fifteen years. In order to keep a patent granted in force the patentee had to pay an annual renewal fee which grew from 50 Marks in the first and second year up to 700 Marks in the fifteenth year. Facing the rising expenditures for holding a particular patent a patentee was supposed to decide to renew his patent only when the costs of doing this were lower than the expected future returns of the patent. The assumption that this mechanism worked as intended is supported by the finding that, in the late nineteenth-century, 70 per cent of all German patents granted had already been given up after just five years. Only about five per cent of all patents reached the maximum age of fifteen years (Streb et al., 2006, 2007). We therefore assume that only those long-lived patents that survived at least ten years can be interpreted as valuable innovations. In this way we identify 1740 long-lasting Prussian patents which will be used as a proxy for the degree of innovation. Each patent document reports the name and the location of the patent holder, and the technological class of the invention. Whereas the name and the legal form of the of the patentee is used to distinguish between independent and large firms patents, the location of the patentee is used to assign a county of origin to the patent. Figure 2 shows the geographical distribution across counties of the stock of valuable patents granted in Prussia between 1877 and Religious diversity Religious diversity is measured with the fractionalization index and is based on the religious composition at the county level of five religious denominations: (i) Catholics, (ii) Protestants, (iii) other Christian, (iv) Jewish, (v) other religion. The Herfindahl index is constructed as follows 5

6 Div = 1 i s 2 i (1) where s i is the share of the religious group i over the total of the (county) population. The fractionalization index represents the probability that two randomly drawn individuals from the population belong to different religious groups. In Figure 3 we show the geographical distribution of religious fractionalization across Prussian counties in Since the ethnic group of Poles and Jews were defined also on the basis of their religion (Polish people were predominantly Catholic), our measure of fractionalization reflects to some extent also ethnic diversity. This is particularly true for the eastern regions of Prussia where the distinction between Protestant-German and Catholic-Poles was sharper. The control variables used in the regression analysis are mainly drawn from the ifo Prussian Economic History Database (ipehd, Becker et al. (2014)). The database contains variables such as literacy rate, landownership concentration, total population, and urbanization. In a further attempt to test the robustness of our results we will also control for the share of labor force in the industrial sectors with a high propensity to innovate such as mining, metallurgy, and chemistry though they are clearly endogenous to innovative activity. The cities of Berlin and Frankfurt show an overly large number of patents, 464 and 136 respectively. To ensure that these two data-points are not influential observations, we will present specifications excluding these two cities. The descriptive statistics of the variables used in the regression analysis are reported in Table 1. 4 Baseline Estimates The variable for long-lasting patents is a count variable characterized by over-dispersion (mean 3.8, s.d. 24) and excessive zeros (66%). Therefore we estimate our baseline model through a zero-inflated negative binomial model (ZINB): No. patents = β 0 + β 1 RelDiv X 1882γ + ε (2) where RelDiv is the fractionalization index in 1871 and X is a vector of control variables including the literacy rate, log of total population, urbanization rate, landownership concentration, the share of Protestants, and a binary variable for counties west of the river Elbe. Standard errors are clustered at the district level (Regierungsbezirk). The ZINB model assumes that the excess zeros are generated by a separate process from the count values and therefore should be modelled separately. In our case, to model the excess zero we use the share of people employed in manufacturing assuming that 6

7 counties with a larger industrial sector, independently on the specific sector of specialization, have a comparatively higher propensity to innovate and thus to patent. 2 results of the baseline estimates are presented in Table 2. Economic theory postulates a trade-off between costs and benefits of diversity (Alesina and La Ferrara, 2005; Ashraf and Galor, 2013). An excessive level of diversity might lead to mistrust and coordination failures which, in turn, can lead to adverse economic outcomes. The In a case with few religious groups, as in nineteenth century Prussia, an excessive level of diversity translates into a high level of polarization which might lead to conflicting situations. Therefore we test for a nonlinear relationship between diversity and long-lasting patents introducing a quadratic term for diversity. At the bottom of the regression tables we always report the level of religious diversity that maximizes the number of patents. As can be seen at the bottom of Table 2, the percentage of labor force employed in manufacturing is a strong predictor of zero patents in a county. The odds of an excessive zero decrease significantly with the percentage of people employed in manufacturing. Put differently, the larger the industrial sector, the larger the probability to innovate. Conditional on that result, the specifications in Table 2 show that religious diversity has a significant relationship with the number of valuable patents. In column 1 we estimate a simple bivariate correlation. We find a strong positive correlation between the fractionalization index and the absolute number of valuable patents. In column 2 we include a quadratic term: the significance of the quadratic specification suggests an inverted U-shape relationship. The level of diversity which maximizes the number of patents is about 0.33 and it remains fairly stable across different specifications. In column 3 we test for higher order polynomials. Yet, the third order polynomial does not enter the model significantly, therefore we keep the quadratic functional form as our preferred specification. In column 4 we introduce an important determinant of innovative activity, namely the stock of human capital proxied by the literacy rate in As expected the coefficient is positive and highly significant (Cinnirella and Streb, 2013). However the association between religious diversity and patents remains highly significant. In column 5 we include our baseline set of control variables to account for possible confounders. We include the log of total population, the urbanization rate, the concentration of large landownership, the share of Protestants, and we control for the counties on the west of the river Elbe. As expected, the size of the population and the level of urbanization are positively related with valuable patents. Although these variables are strongly correlated with religious 2 A test of the zero-inflated model versus the standard negative binomial model (Vuong test) indicates that the zeroinflated model is preferable. However by using a negative binomial regression without correcting for excessive zeros we obtain qualitatively the same results. See also the estimates in column 5 of Table 2 for a specification conditional on strictly positive patenting activity. 7

8 diversity, the association of diversity with valuable patents remains virtually unaffected. In this case the level of religious diversity that maximizes the number of patents is Previous research has shown that large landowners opposed the spread of education (Galor et al., 2009; Cinnirella and Hornung, 2013). Such a conservative environment might have also negatively affected the innovative spirit of a county. Our baseline estimates do not show a significant relationship between large landownership and valuable patents. Instead, as expected, we find that counties west of the river Elbe have a significantly higher propensity to patent. This is consistent with the notion that western counties are richer in natural endowments and more industrialized. We control also for distance to Berlin since the Imperial patent office was located in Berlin and, in order to patent an innovation, one had to hand in the proposal to the central office. Yet we do not find any significant relationship with distance to Berlin. Similarly, the share of Protestants in the county seems to have no relationship with the number of valuable patents. In column 6 we estimate our baseline model for the subset of counties with positive patenting activity. Although the sample is reduced by about two thirds, the relationship between religious diversity and valuable patents remains highly significant. The coefficients of the control variables are also largely unaffected. This is an important result as it suggests that our relationship of interest is not simply driven by the large number of counties with zero patenting activity. Similarly in column 7 we drop from the regression sample the city-counties, i.e. counties with a 100% urbanization rate. The coefficients for religious diversity are less precisely estimated and this is due to the fact that a large part of religious diversity is to be found in large cities or in urban environments. The distribution of valuable patents is not independent from the geographical location of innovative industries. Crafts and Wolf (2014) show that the location of the UK cotton industry in 1838 is positively related to the availability of water power, ruggedness of terrain, and to proximity to ports. Gutberlet (2014) analyses the extent to which the dependence on natural resources such as water and coal affected the location of manufacturing in Germany at the end of the nineteenth century. In our case, large chemical firms settled at the banks of the rivers Rhine and Main, which were not only important navigable waterways, but were also used as a water source and to get rid of effluents. Firms engaged in the field of mechanical engineering were particularly concentrated in the neighborhood of iron and steel producers, namely in the Greater Ruhr area, and near textile manufacturers as in the Province of Saxony. To account for these factors, in column 8 we include the shares of people employed in the high-patenting sectors such as mining, metallurgy, and chemistry. Clearly these variables are endogenous in our model as they might be the result of innovative activity, therefore constituting bad controls. Yet, in combination with the indicator for the western counties, these controls allow to account for the location of the highly innovative sectors. Indeed we find that 8

9 these variables, in particular that for chemistry, are strongly correlated with the number of valuable patents. The concave relationship between religious diversity and patenting activity remains, however, highly significant. We also have a variable for the share of people employed in the service sector which includes banks and insurances. This is potentially an important correlate of patenting activity given the relatively expensive nature of the German patent system compared, for example, to the US system (Kahn and Sokoloff, 2004). A developed financial sector might have enhanced the patenting process. Indeed we find a significant positive association between the share of people in the service sector and the number of valuable patents (column 8). In sum, the estimates in Table 2, which account for the excessive number of counties with zero patents, strongly suggest an inverted-u shape relationship between religious diversity and the number of long-lasting patents. 5 Identifying Exogenous Variation The issue of endogeneity in our research question is of particular concern. Religious diversity could be the result of migration and concentration of people of different cultures and abilities in regions characterized by high economic potential. Counties with pronounced diversity might also be characterized by a more liberal socio-economic environment, open to new cultures and ideas. These unobserved characteristics are a serious threat as they might bias, or in the worst case invalidate, the hypothesized relationship between religious diversity and innovation. Thus, in order to get as close as possible to causal estimates of the effect of religious diversity on innovation, we propose two independent instrumental variable approaches equipped with a large number of robustness checks. In our first approach we use information on religious fractionalization in the preindustrial period, namely in In fact, an advantage of our variable of interest religious diversity is its strong geographical persistence. The religious composition of the population in Prussian counties is the result of a complex set of historical processes started with the Protestant Reformation in 1517, successively affected by the Peace of Augsburg (1555) which established the principle Cuius regio, eius religio, and by the Peace of Westphalia (1648) which ratified the end of the Thirty Years War. Along this time period the rulers of the various principalities and city-states decided about the religious denomination to impose on their subjects (Cantoni, 2012). We argue that the religious composition of the population in 1816 is to a considerable extent the result of a random process possibly orthogonal to innovative activity at the end of the nineteenth century. In econometric terms, our identification strategy is based on the assumption that religious diversity in 1816 has an impact on the stock of long-lasting patents in only through religious diversity in

10 Clearly, we do not claim that the whole variation in religious diversity in 1816 is the result of a random process orthogonal to innovative activity. Cantoni (2012) studies the determinants of the adoption of Protestantism for a cross-section of 119 territories in central Europe and finds that the nature of a territory s rule, its size, and its geographic location (i.e. distance to Wittenberg) were important determinants. For example, Free Imperial Cities were important economic centers during the reformation period with independent institutions which allowed them to control their own trade and to accumulate a large amount of wealth. This freedom and prosperity might have attracted people of different religious groups giving origin to the religious diversity we observe in the nineteenth century. A similar story would apply to territories belonging to the Hanseatic League during the late Middle Age. To make our instrumentation credible we will account for such historical characteristics. In addition, if we claim that religious diversity has an effect on innovation and is related to population size and urbanization at the end of the nineteenth century, we have to ensure that such effects did not already take place in In a series of validity checks we will show that our proposed instrumentation holds also when accounting for population size, urbanization, population density, and enrollment rates in Instrumental variable estimates using past religious diversity The population census in 1816 reports information on the religious composition of the population at the county level. In particular the religious groups reported in the 1816 census are only four compared to the five groups reported in The four groups are Protestant (distinguished between Lutheran and Reformed Protestant), Catholic, Jewish, and Mennonite. To render the grouping in 1816 more consistent with the religious denominations in 1871 we consider Mennonites as other Catholic religion. 3 Yet, one disadvantage of using religious diversity in 1816 to predict religious diversity in 1871 is that we need to aggregate the data at the county borders in force in This explains why in the following instrumental variables estimates the number of observations is reduced to 280. Also differently from the baseline estimates, in the instrumental variables approach we resort to a standard two stage least squares using the logarithm of 1 + patents per 10,000 individuals as dependent variable. This log transformation has the nice feature of being zero for counties with zero patents and is also useful to account for counties with extremely high number of patents such as Berlin which reports 464 long-lasting patents for the period In Figure 4 we plot the relationship between religious diversity in 1871 and in As already mentioned, we find a very strong persistence of the religious composition of 3 The Mennonites are an Anabaptist Christian group named after Menno Simons ( ), a religious leader from Friesland. 4 An explanation on how to merge Prussian data from different years is explained in Becker et al. (2014). 5 It is important to note that our results do not depend on the logarithmic transformation. Qualitatively identical results are obtained without the logarithmic transformation. 10

11 the population. This strong correlation is mirrored by the large value of the F-statistic of the first stage reported in Table 3. In column 1 of Table 3 we report the OLS estimates with the 280 observations of the 1816 borders. Also in this case the estimates suggest an inverted-u shape relationship. In column 2 we estimate a two stage least squares model without control variables. The coefficients for religious diversity are highly significant. In column 3 we include the proxy for human capital, the literacy rate, whereas in column 4 we include our standard set of control variables. 6 The estimates are largely consistent with the OLS estimates, namely we find a strong positive impact on valuable patents of urbanization and for the counties west of the river Elbe. The nonlinear effect of religious diversity on patenting activity is also highly significant. In column 5 we include the share of people employed in industrial sectors with a comparatively higher propensity to patent (mining, metal, chemistry, plus the service sector) and which rely on the presence of natural endowments. In this way we can account for the location of industries which depends on natural resources which could affect the cultural and skill diversity of the population. The estimates show that, for a given level of employment in such sectors, higher levels of religious diversity positively affect innovation, up to a certain point. Interestingly, we find that once accounting for the employment in the mentioned industrial sectors the level of diversity which maximizes the number of valuable patents is smaller (0.28). We obtain a similar result when constraining the sample to counties with positive patenting activity (column 6): the coefficients for religious diversity are highly significant and the maximizing level of diversity is about 0.3. The large F-statistic of the first stage for the subsample with positive patenting activity suggests that the effect of diversity on innovation is not identified through the zero-patenting counties. In column 7 and 8 we test whether the results are affected by the functional form of the dependent variable or the functional form of the variables on the right-hand side. In particular in column 7 we use as dependent variable the number of patents per 10,000 individuals. As one can see the significance and magnitude of the coefficients of interest are not affected. In column 8, instead, we include the squared term of literacy, urbanization, population density, and landownership concentration. In fact one could argue that the robustness of our results for religious diversity is due to the fact that the control variables enter the model linearly whereas religious diversity enters the model in a quadratic fashion. The estimates in column 8 do not support this hypothesis: the coefficients for diversity still suggest a significant inverted-u shape effect and the level of diversity maximizing the number of patents remains around The estimated effect of religious diversity on innovation is also economically significant. Referring to the estimates in column 7, moving along the distribution of religious diversity 6 Since the dependent variable is already computed per 10,000 individuals, in the regression we control for population density instead of population size. 11

12 from the 25th percentile (0.03) to the 75th percentile (0.3) implies an increase in the number of patents per 10,000 individuals from 0.26 to 0.60 (+130%). 5.2 The exogeneity of religious diversity in 1816 The proposed instrumentation is based on the assumption that religious diversity in 1816 has an impact on patenting activity in the period only through religious diversity in Although economic theory postulates that, in general, diversity has a positive effect on economic outcomes at higher level of development (Alesina and La Ferrara, 2005), 7 one could argue that religious diversity in 1816 had already a positive effect on proto-industrialization which, in turn, is related to innovation and patenting activity during the second industrial revolution. This persistent process would invalidate our instrumental variable approach. To directly address this issue we estimate our model substituting our standard vector of control variables for the period 1871/1882 with a set of variables which are contemporary to the instrument (1816) and thus account for the possible impact of religious diversity in 1816 on contemporary development. In Table 4 we successively include variables for total population, urbanization rate, population density, and primary school enrollment rate in We also keep the indicator variable for the counties west of the river Elbe. It is important to include this regional control also for this time period as these counties had already a set of liberal economic institutions more conducive to economic growth (Acemoglu et al., 2011). These regions enjoyed also a richer set of natural endowments (water-power, soil quality, and coal) suitable for the development of the pre-industrialization period and for the heavy industry of the late nineteenth century. The estimates in Table 4 indicate that many of the variables capturing the level of economic development in 1816 are strongly correlated with valuable patents during the second industrial revolution. In column 1 we include population size in This variable is strongly correlated with patents in , yet the second stage estimates for religious diversity are highly significant and precisely estimated. In column 2 we include the urbanization rate in 1816 which appears also to be strongly correlated with successive patenting activity. On the contrary, population density and enrollment rates in primary school are not correlated with patenting activity at the end of the century. In column 5 we include simultaneously all the controls for The coefficients for religious diversity are highly significant and the maximum of the concave function is always in the same range ( ). Finally, in column 6 we report the reduced form estimate in which we regress the log of patents per capita on religious diversity in 1816, plus the set of controls in Not surprisingly, given the strong persistence of religious fragmentation, the estimates in column 6 confirm the inverted-u shape relationship between religious diversity and patenting activity. 7 In section 6 we will provide some evidence in support to this hypothesis 12

13 Accounting for the level of development in 1816 is important to ensure the orthogonality of religious diversity in 1816 with respect to innovation at the end of the nineteenth century. Yet this could not be enough. The strong persistence of religious fragmentation suggests to account also for those factors that determined in the first place the decision to adopt a denomination. Cantoni (2012) studies the determinants of the adoption of Protestantism for a cross-section of 119 territories in central Europe. He finds that the nature of a territory s rule, its size, and its geographic location (distance to Wittenberg) were important factors for the adoption of Protestantism. For example Free Imperial Cities were important economic centers during the reformation period with independent institutions which allowed them to control their own trade and to accumulate a large amount of wealth. This freedom and prosperity might have attracted people of different religious groups giving origin to the religious diversity we observe in the nineteenth century. Examples of Free Imperial Cities in our sample are Cologne, Frankfurt am Main, Dortmund, and Danzig (Gdansk). Hanseatic cities were also relatively economically advanced in the sixteenth-century and might have also attracted people of different religious denominations. A similar argument could apply to areas that were advanced in high education (Cantoni and Yuchtman, 2014). These factors are related to religious fractionalization and at the same time might have a persistent impact on innovative activity. In Table 5 we present an extensive set of robustness checks to exclude the persistent effect of historical characteristics on innovative activity during the second industrial revolution. In column 1 we include an indicator variable for counties containing a Free Imperial City in 1517 (n = 18), in column 2 an indicator for counties containing cities that belonged to the Hanseatic League by 1517 (n = 22), and in column 3 we control for the presence of a University in 1517 (n = 8). It appears that these variables are not significantly correlated with valuable patents during the second industrial revolution. In a recent paper Hornung (2014) investigates the effects of Huguenots migration to Prussia on productivity. He finds a substantial long-term effect of Huguenots settlement on the productivity of textile manufactories. Huguenots were French Reformed Protestants who might have contributed to increase the religious diversity of a county. In column 4 we include the log of Huguenots in The point estimate suggests a positive relationship with patents in , yet the coefficient is not precisely estimated. This (non)result is not surprising for two reasons: first, the sector of specialization of the Huguenots (textile) was not one of the leading innovative sectors during the second industrial revolution in Germany; second, the presence of Huguenots was regionally clustered reducing thus variation across counties. In columns 5 and 6 we successively include distances to Mainz and Berlin. Dittmar (2011) shows that the diffusion of the printing press has a sizable effect on city growth in 8 We thank Erik Hornung for making the data available to us. 13

14 Germany between 1500 and Dittmar identifies exogenous adoption of the printing press using distance to Mainz, the location where Johannes Gutenberg established the first printing Press around The printing press contributed to the spread of the Protestants Reformation and could therefore be related to religious diversity (Rubin, 2014). At the same time, through its effect on growth, the printing press might also have a persistent effect on innovative activity thus biasing our identification strategy. The estimates in column 5 suggests that this concern is unfounded. The same applies to distance to the capital Berlin which, as already seen in Table 2, is not correlated with valuable patents. 9 Finally in column 7 we include all the control variables simultaneously. The specifications presented in Table 5 show that both the strength of the instrument and the nonlinear effect of religious diversity on innovation are unaffected when accounting for the historical characteristics that contributed to the spread of Protestantism, thus potentially fostering religious diversity. 5.3 Religious diversity and innovation in East Prussia An alternative instrumental variable approach In this section we propose an alternative instrumental variable based on the eastward territorial annexations of the Kingdom of Prussia. The advantage of this alternative instrumentation is twofold: (i) we can exploit a different source of variation in religious diversity which permits us to corroborate our claim of causality; (ii) this alternative instrumental variable approach allows to identify variation in religious diversity in the eastern regions of Prussia. The latter is an advantage as we can drop from the analysis those regions west of the river Elbe which are comparatively more developed because of better natural endowments and with a different set of economic institutions. In fact these conditions make the identification of the causal effect of religious diversity on innovation more difficult. By focusing on the eastern regions we can identify variation in religious fractionalization which is to a lesser degree related to growth in urbanization or industrialization. The disadvantage is that we have less innovation. We identify exogenous variation in religious diversity in the counties east of the river Elbe exploiting the Prussian eastward annexations that occurred from the sixteenth century until the Congress of Vienna in The logic behind this instrumentation is that each territorial expansion towards east generated an exogenous variation in the religious composition of the population. For example the Partitions of Poland at the end of the eighteenth century contributed substantially to the religious diversity of the Prussian population because of the Catholic and Jewish religion of ethnic Poles. 10 This can be 9 In this context another relevant distance would be distance to Wittenberg, the place from where Martin Luther spread the Protestant ideas. However, since we are already controlling for the share of Protestants and the literacy rate, it would be redundant to further account for distance to Wittenberg. 10 The Partitions of Poland were a series of three partitions executed by the Kingdom of Prussia, the Russian Empire, and Habsburg Austria that ended the existence of the Polish-Lithuanian Commonwealth. The first partition of Poland took place in 1772 and Prussia gained 36,000 km 2 and about 600,000 people. The Polish-Lithuanian Commonwealth lost about 30 percent of its territory and half of its population. The territories annexed by Prussia became a new province in 14

15 seen from the figures in Table 6 which report the average religious denominations in 1871 by annexation. In the last column we report the average religious diversity. It can be seen that the acquisition of Silesia (starting in 1742) substantially increases religious diversity with its 49% of Catholics and 1% of Jews. The contribution to religious diversity of the First and Second Partitions of Poland are even more noteworthy with an absolute majority of Catholics and a significant share of Jews. On the contrary the territories acquired after the Congress of Vienna, mainly part of the Kingdom of Saxony, are very homogeneous in terms of denominations. In Figure 5 we show the different annexations for the counties east of the river Elbe. We predict county-level religious diversity in 1871 for the regions east of the river Elbe estimating the following equation: RelDiv i = j β log DistBorder i Annex j + j γ Annex j + X δ + ε (3) where log DistBorder is the log distance in km to the eastern border in 1871, Annex are 8 dummy variables for the different annexations as reported in Table 6. The vector X contains the control variables of the second stage estimate, namely literacy, population density, urbanization rate, the share of Protestants, and landownership concentration. To ensure that the time of the annexation has not a direct impact on our outcome, we also include a variable which computes the number of years since annexation to Prussia with respect to Furthermore we include latitude and longitude of the county centroid to ensure that we are not capturing other factors related to the geographic position of the county. We assume that the further a county is from the eastern border, the larger the share of Protestants in the population and thus the lower is religious diversity. On the contrary, counties closer to the eastern border will be more heterogenous with a comparatively larger number of Catholics, Jewish and other small religious groups. With the interaction terms we allow the relationship between distance to the eastern border and diversity to differ across annexations. We identify variation in religious diversity also through the indicators for the annexations called West Prussia. According to Clark (2007, p. 233), about 54 percent of the population of the new territories were German-speaking Protestants. In the second partition of Poland, which took place in 1793, Prussia got an additional 58,000 km 2 and the Commonwealth was reduced to 215,000 km 2. Among others, Prussia received the cities of Danzig and Thorn. Prussia organized its newly acquired territories into South Prussia, which became the province of Posen following Napoleon s victory over Prussia in 1807 and the reorganization after the Congress of Vienna in Finally the third partition of Poland in 1795 effectively ended the Polish-Lithuanian Commonwealth. Prussia received Podlachia, the remainder of Masovia, and Warsaw, with 1 million people. Yet, these territories became part of the Kingdom of Poland and went under the influence of Russia after the Congress of Vienna. Prussia in return received 40 percent of the Kingdom of Saxony, later known as the province of Saxony. 11 This variable varies from 53 to 346 years 12 In order to get predictions of religious diversity between 0 and 1, in equation 3 we use a logit transformation of the dependent variable. 15

16 The estimates of equation 3 are reported in the appendix. The model explains more than 80 percent of the variation in religious diversity in The coefficients estimated are in line with our expectations: β is generally negative and highly significant, consistent with the notion that counties nearer to the border are denominationally more heterogeneous. Yet the interaction terms reveal some important differences: for example the negative impact of distance in Mark Brandenburg (annex 4) and in Western Pomerania (annex 7) is significantly larger than in the far eastern regions of the the Duchy of Prussia (annex 1). On the contrary, distance to the eastern border is not significantly related to religious diversity in the counties acquired through the Second Partition of Poland (annex 5). The latter result is not surprising as these are counties belonging to the province of Posen which is located near the border. In many counties of this region being near the border implies a low level of religious diversity as the majority is constituted by Catholic Poles. A similar result is obtained for the counties in Silesia (annex 6) for which the negative coefficient is small and not significant. In Figure 6 we plot actual religious diversity in 1871 against religious diversity predicted on the basis of equation 3. As one can see the model does a remarkably good job in predicting religious diversity in We thus use the predicted diversity and its squared term in an instrumental variable approach. 13 The second stage estimates are shown in Table 7. For comparison, in column 1 and 2 we provide estimates of an OLS and a zero inflated negative binomial model for the eastern subsample (n = 212). 14 Although the coefficients in the OLS model are not significant, the maximum of the concave function at 0.30 is statistically different from zero. In column 3 we show the instrumental variable estimates with the standard set of control variables. The alternative instrumentation based on the east-elbe sample supports our main finding of a nonlinear effect of religious diversity on innovation. The fact that most of the Catholic population in the eastern counties is of Polish ethnic origin might constitute an important confounding factor. We can separate the two factors by using information of languages. The education census in 1886 collected information on the language spoken by pupils at home. In particular, the census in 1886 reports the number of pupils speaking German, Polish, Lithuanian, Wendish, Slavic, Danish and other language. Thus we construct a variable for the share of non-german speaking people. This variable is included in column 4. Indeed, counties with a higher share of non-german speaking people are associated with lower innovation. Clearly, the inclusion of the share of non-german speaking people weakens our instrumentation as distance to the border is related to that variable. However our relationship of interests is virtually unaffected suggesting that religious difference is the salient factor relevant for innovation Standard errors in the two-stage least squares are bootstrapped with 100 repetitions. 14 Another advantage of this alternative instrumentation is that we can keep the borders structure as in place in Our results remain unaffected also if we include a variable for linguistic fractionalization. 16

17 We have previously argued that one advantage of focusing on the east-elbe sample is that we can abstract from the rapid industrializing areas of Rhineland and Westphalia, rich also in natural endowments. That setting proves more difficult in identifying exogenous variation in religious denominations which is not related to economic potential. In column 5 we show that the effect of religious diversity on innovation in east-elbian counties remains (though weaker) also when accounting for the share of people employed in manufacturing. Finally in column 6 we show that Imperial of Hanseatic cities in 1517 (n = 12) do not impact innovation at the end of the nineteenth century and do not affect our relationship of interest Possible mechanisms We have seen that also when using an alternative instrumentation and focusing on the less developed regions east of the river Elbe we find a nonlinear effect of religious diversity on valuable patents. For the rest of the paper we go back to the analysis based on the full sample, which uses religious diversity in 1816 as instrument, to investigate some possible channels. A potential channel through which religious diversity affects innovation is education. It is possible that the levels of education within each religious denomination affect the nonlinear effect of religious diversity. For instance, low levels of human capital of a particular denomination could explain the negative marginal effect associated with high levels of diversity. We can test this channel using information on the number of illiterate people by denomination reported in the 1871 census. In particular in column 1 of Table 8 we include the share of illiterate people (older than 10 years) among Protestants, Catholics, and Jews, respectively, and we drop the control for total literacy. 17 The controls for religious-specific illiteracy are not related to patents and do not affect the impact of religious diversity. Literacy, however, captures only basic skills which might not be relevant for innovation and patenting activity. Ideally we would like to have data on attendance rates of secondary schools and technical institutes. To our best knowledge such data are not available at the county level for Prussia in the period under consideration. A second best solution is to control for (i) the presence of a university in the county or the (ii) distance to the nearest university (columns 2 and 3). 18 In both cases we do not find a significant association with valuable patents. 16 Imperial cities and Hanseatic cities are perfectly collinear in the east-elbe sample. We did not include the control for universities in 1517 as it concerns only one case, namely Greifswald. 17 Summary statistics for the shares of illiterate individuals by religious denomination support the notion that Protestants and Jews had higher literacy than Catholics. The average illiteracy rate for Protestant is 6%, for Catholics is 10%, while for Jews is 5%. 18 The universities present within the 1816 Prussian borders are Königsberg, Strasburg, Berlin, Greifswald, Breslau, Halle, Munster and Bonn. 17