doi:10.1111/j.1435-5957.2009.00271.

Migration and innovation: Does cultural diversity matter for regional R&D activity?*
pirs_271 563..586

Annekatrin Niebuhr1
1

IAB Nord, Regional Research Network, Institute for Employment Research, Projensdorfer Strae 82, D-24106 Kiel, Germany (e-mail: annekatrin.niebuhr@iab.de)

Received: 27 April 2009 / Accepted: 20 September 2009

Abstract. Recent theoretical research deals with economic costs and benets of cultural diversity related to immigration. However, empirical evidence regarding the impact of cultural diversity on economic performance is still scarce. We analyse the effect of cultural diversity of the labour force on patent applications for a cross-section of German regions. The results suggest that differences in knowledge and capabilities of workers from diverse cultural backgrounds enhance performance of regional R&D sectors. As regards innovation, the benets of diversity seem to outweigh the costs caused, for example, by communication barriers. JEL classication: C21, J61, O31 Key words: Cultural diversity, regional innovation, knowledge production function, Germany

1 Introduction The signicance of the immigration of qualied workers will rapidly increase in the ageing European economies since demographic change will cause a decline of the labour force and a sharp increase of the average age of workers. Foreign workers are already an important factor of the German economy. In 2004, almost 7% of all employees in Germany have foreign nationality. More than 100,000 highly skilled foreigners with a university degree work in Germany. Zimmermann (2005) notes that, in spite of the rising importance of migration, the issue is still controversial, and the understanding of the effects of international labour mobility is rather limited. Research on the economic consequences of migration has mainly focused on labour market effects and, more precisely, on the question whether immigrants depress wages and increase unemployment of native workers. Many analyses stress substitution effects among
* I would like to thank Christoph Grenzmann (Stifterverband fr die Deutsche Wissenschaft) for the generous provision of regional R&D data and Andrea Stckmann for excellent research assistance. Financial support from the Volkswagen Foundation is gratefully acknowledged as part of the Study Group on Migration and Integration Diversity, Integration and the Economy. I thank Herbert Brcker, Eckhardt Bode, Stefan Fuchs, Rdiger Wapler and anonymous referees for helpful comments and suggestions on an earlier version of this paper. The usual disclaimer applies.
2009 the author(s). Journal compilation 2009 RSAI. Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden MA 02148, USA. Papers in Regional Science, Volume 89 Number 3 August 2010.

564

A. Niebuhr

native and foreign workers. However, taking into account that labour is not homogenous, the impact of immigration depends on whether migrants are skilled or unskilled and on labour market conditions in the host country. The objective of this paper is to provide evidence on the impact of migration on innovation, a subject that has not received much attention in the migration literature up to now (see Hunt and Gauthier-Loiselle 2008). This papers analysis differs from many previous studies that focus on labour market effects of immigration.1 Moreover, there are only a small number of papers that investigate the signicance of labour mobility as regards knowledge spillovers and innovation (e.g., Almeida and Kogut 1999; Kim and Marchke 2005; Simonen and McCann 2008). The ndings so far suggest that the mobility of R&D staff is associated with the transfer of knowledge and positively affects patent applications. However, these studies do not consider the heterogeneity of researchers with respect to their cultural background. This is another aspect that differentiates this analysis from other studies. We do not restrict heterogeneity of labour to the level of education only. Due to their different cultural backgrounds, it is likely that migrants and native workers have fairly diverse abilities and knowledge. Thus, there might be skill complementarities between foreign workers and native in addition to those among workers of different qualication levels. Presumably foreign and native workers of the same educational level are imperfectly substitutable groups because of cultural differences. Fujita and Weber (2004) argue that cultural diversity of the labour force might be of special importance for R&D activity since the generation of new products and ideas heavily relies on individual talents and skills from diverse educational and cultural environments. The possibility that diversity can enhance productivity, innovation, and growth has already been considered in the economic literature. Diversity might refer to economic diversity, i.e. heterogeneity of rms and industries, or to the diversity of people. Most studies have concentrated on the impact of economic diversity and the effects of a diverse urban environment rather than on cultural or ethnic diversity of people.2 According to Jacobs (1969), diversity of geographically proximate industries promotes innovation and growth in cities. Glaeser et al. (1992) as well as Feldman and Audretsch (1999) provide corresponding empirical evidence for U.S. cities. Duranton and Puga (2001) and Henderson et al. (1995) investigate the role that a diversied urban environment plays in fostering innovation and attracting innovative industries. Romer (1990) highlights the signicance of a variety of intermediate inputs for productivity in his seminal endogenous growth model. As regards the diversity of people, Keely (2003) argues that interaction between heterogeneous skilled workers gives rise to knowledge spillovers and produces new research ideas. Hunt and Gauthier-Loiselle (2008) as well as Kerr and Lincoln (2008) investigate the effect of skilled immigration on innovation for a cross-section of U.S. States. According to the former study, immigrants patent at much higher rates than natives. This patenting advantage is due to the above average share of degrees in science and engineering among skilled immigrants. Kerr and Lincoln (2008) show that invention increases with higher admission levels through the direct contribution of immigrants to patenting. Empirical evidence provided by Anderson et al. (2005) suggests that creativity is greater in regions marked by more diverse employment bases. Finally, Audretsch et al. (2009) consider the impact of cultural diversity of the labour force on entrepreneurship. While there is an emerging theoretical literature dealing with the economic effects of cultural diversity (e.g., Lazear 1999b, 2000; Fujita and Weber 2004), there are surprizingly few empirical studies within the eld of economics. Theoretical models consider different costs and
1 See Longhi et al. (2008) for a meta-analysis of the labour market effects of immigration. Ozgen et al. (2009) consider the impact of migration on income convergence. 2 See Duranton and Puga (2000) for a survey of literature on diversity in cities and its economic effects.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

565

benets of diversity and specify various linkages between diversity and economic performance. However, corresponding empirical work, that can help determine whether positive or negative effects of cultural diversity prevail, remains scarce. Until now, there has been mainly crosscountry evidence and studies focusing on growth and productivity effects in U.S. regions (Easterly and Levine 1997; Ottaviano and Peri 2005, 2006). Up to now, evidence on the impact of skilled immigration on innovation, provided, for example, in Chellaraj et al. (2008), Hunt and Gauthier-Loiselle (2008), and Kerr and Lincoln (2008) is restricted to the U.S. Most investigations that analyse the relationship between innovation input and output fail to take cultural diversity into account (e.g., Anselin et al. 1997; Bottazzi and Peri 2003; Bode 2004,). The aim of this paper is to investigate the impact of cultural diversity on regional innovation in Germany. Therefore, we extend the knowledge production framework to analyse whether a more diverse labour force, from a cultural point of view, fosters innovation due to production complementarities, or whether negative effects of diversity, caused, for example, by language barriers, outweigh the benets. Due to data restrictions, we dene cultural diversity as diversity of workers nationality rather than ethnicity or cultural background. Regionally differentiated information on country of origin of inhabitants and employees is not available in German ofcial statistics. The rest of the paper is organized as follows. In Section 2, the theoretical framework of the analysis is outlined. Production complementarities and costs associated with cultural diversity are discussed. The data set applied in the empirical analysis is described in Section 3. An important issue of the investigation concerns the measurement of cultural diversity. In Section 4, we introduce the applied diversity indicators and provide some empirical evidence of cultural diversity in German regions. We employ the knowledge production function approach to investigate the impact of cultural diversity on regional innovation capacity. The corresponding regression model and some robustness issues are discussed in Section 5. The regression results are presented in Section 6. Conclusions follow.

2 Theoretical framework Ottaviano and Peri (2006) argue that skills of foreign workers might complement those of the native labour force. In their model of multicultural production, different cultural groups provide different services. Diversity has a positive impact on regional productivity. However, heterogeneity also hampers the exchange between different cultural groups: there are adverse productivity effects because of cultural distance. Other authors also recognize that there is a trade-off with respect to heterogeneity. Lazear (1999a, 2000) considers the positive productivity effects of ethnic diversity, but there are costs of diversity arising from barriers to communication caused by different languages and cultures.3 Thus, according to the literature, there appears to be an optimal degree of diversity which is inuenced by the nature of production. Some of the literature on this theme also examines the signicance of institutions in this context. An important result of this research is that the implementation of growth enhancing effects of diversity may require a specic set of rules, or regulatory framework. Ottaviano and Peri (2006) emphasize the role of a core of shared norms (integration) that might constitute a prerequisite for realizing the potential benets of diversity. There appears to be a link between the costs and benets of diversity on the one hand and the concept of ethnic identity described in Constant et al. (2006) on the other hand. According to the authors, migrants start out from their ethnicity, i.e. permanent characteristics associated
3 Costs of diversity might also be due to an inability to agree on common public goods and public policies. See Alesina and La Ferrara (2005).

Papers in Regional Science, Volume 89 Number 3 August 2010.

566

A. Niebuhr

with the country of origin, and then develop their ethnic identity as they are exposed to the culture and values of the host country. Ethnic identity is dened as the balance between commitments with the host country and commitments with the country of origin. Constant et al. (2006) distinguish four states of ethnic identity: assimilation, integration, marginalization and separation. Assimilation seems to imply a strong decline of both costs and benets of cultural diversity since it is characterized by a strong identication with the host country and conformity to the corresponding norms and codes. With respect to the economic effects of diversity, integration might be interpreted as the best state because it involves commitment to the host society, but also a strong dedication to the culture of origin, thus, still ensuring high benets but relatively low costs of diversity. In contrast, in case migrants are primarily identied as marginalized or separated, cultural diversity may mainly entail high costs. The benets of diversity might be of particular importance in the R&D sector, whereas in industries specialized on more standardized forms of production it might be more likely that costs of a diverse labour force outweigh the positive effects. Alesina and La Ferrara (2005) argue that cultural diversity may lead to innovation and creativity since it involves variety in abilities and knowledge. Fujita and Weber (2004) note that knowledge production relies heavily on the talents and skills of employees coming from a wide range of cultural backgrounds. The nature of R&D activity calls for interaction between different workers and a pooling of different ideas and abilities. Berliant and Fujita (2008) also refer to the signicance of cultural diversity for knowledge creation and transfer. The heterogeneity of people is important for the creation of new ideas. As outlined by Alesina and La Ferrara (2005), ethnic diversity can affect economic performance in different ways. Diversity might have a direct impact on economic outcomes via different preferences or by inuencing individual strategies. Moreover, diversity might have an inuence on the production process. Our analysis focuses on the latter approach. As in Romer (1990), we assume that the production of new knowledge in region i requires people engaged in R&D. Following Ottaviano and Peri (2005), we augment the production function to allow for effects of cultural diversity of the workforce. R&D staff Lin differs with respect to nationality n. In order to develop new designs of products Pi, they can make use of the existing technological knowledge Ai:

Pi = [1 ( divi )] Ai1 ( Lin )

n =1

(1)

where t(divi) (0,1) is transaction cost that is an increasing function of cultural diversity. Thus, only a fraction [1 - t(divi)] of inputs is available for production due to communication barriers and other differences between cultural groups which hamper interaction among them. The additively separable function with respect to labour input implies that, given the total number of researchers, output increases as jobs are distributed across more groups of workers. With a given number of nationalities N, it is more productive to distribute employees evenly across the N groups than to concentrate on a single nationality. As regards the qualication level, we assume labour to be homogeneous. Heterogeneity only refers to the cultural background of the workers.4 Researchers with the same educational attainment but different nationality are not perfect substitutes for each other. With this functional form, productivity in R&D is higher, the larger the number of nationalities and the more balanced their supply. The magnitude of the positive impact of diversity depends on the elasticity of substitution among nationalities s = 1/(1 - a).

4 In the regression analysis, we control for qualication of the workers since we focus on a specic skill level, namely, highly qualied R&D employees.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

567

In order to differentiate the impact of a simple increase of R&D staff from the effect of diversity, we multiply Equation (1) by (Li/Li)a
1 Pi = [1 ( divi )] Ai1 L Li divi i ( Lin Li ) = [1 ( divi )] Ai

(2)

n =1

Cultural diversity of R&D workers is given by divi = ( Lin Li ). The diversity index is inuenced by the number of nationalities that are present in the regional workforce and by the distribution of R&D employment across nationalities. The marginal effect of diversity on knowledge output is given by:
n =1

Pi 1 = Ai1 L divi i [1 ( divi )] [1 ( divi )] divi divi

(3)

Thus, the size and direction of the diversity effect is inuenced by two factors: the elasticity of substitution between nationalities, i.e. the strength of the complementarity, and the transaction costs associated with diversity. With respect to the trade-off associated with cultural diversity, some issues emphasized by Lazear (2000) are noteworthy. Lazear argues that the gains from diversity should be greatest if different groups have information sets that are disjoint because then individuals can learn a great deal from each other. In contrast, if knowledge and capabilities completely overlap, there are no benets from diversity. Moreover, the information and knowledge must be relevant for other groups. With respect to innovation it is most likely that primarily the know-how of high skilled foreigners is relevant. The educational attainment of workers might also affect the costs of diversity. Costs that arise due to language barriers might be lower among employees with a university degree. This is in line with Keely (2003) who argues that interaction cannot take place if the technology gap between workers is too large. Thus, the skill level of workers probably inuences both the relevance of information and the barriers of information exchange.

3 Data Point of departure of our empirical analysis is the knowledge production function outlined in Section 2 that links R&D input to R&D output, namely, new products, processes and ideas. Thus, we rst of all need adequate proxies for regional innovation and R&D input to investigate the impact of cultural diversity on knowledge production. Regional data on patent applications, used as a measure for knowledge output,5 and on R&D inputs in Germany are available on the county level (NUTS 3) and for planning regions (so-called Raumordnungsregionen) which consist of several NUTS 3 regions linked by intense commuting. We have to restrict the regression analysis to planning regions due to some data restrictions for NUTS 3 regions. Overall, our cross section contains 95 regions. Data for these regions is generated by aggregation of NUTS 3 level information if data is not available for planning regions. Furthermore, the analysis takes into account the region type. Starting from a classication based on a typology of

5 Bottazzi and Peri (2003) argue that patent applications can be considered as a good approximation of innovation. For an in-depth discussion of the caveats of applying patents as a measure of innovation see also Griliches (1990).

Papers in Regional Science, Volume 89 Number 3 August 2010.

568

A. Niebuhr

settlement structure according to the criteria population density and size of the regional centre, we differentiate between agglomerated, urbanized and rural regions.6 Patent applications, applied as an indicator for innovative output of the region, comprise patents published by the German and the European patent ofce that have been assigned to the innovators region of residence. As Bode (2004) notes, this approach avoids potential mismeasurement due to centralized patenting of multi-site companies. Annual patent data is available for the period 1995 to 2000.7 Information on R&D input was provided by the German Stifterverband. R&D data include R&D staff as well as R&D expenditure of commercial rms. The data come from a census and are available for 1995, 1997 and 1999. However, we can only use data for 1997 and 1999 in our analysis. Data for 1995 is not compatible due to some changes in the delineation of regions. Thus, the investigation is restricted to a panel data set with only two observations in the time dimension. Finally, we include several explanatory variables in the regression model based on employment data provided by the German Federal Employment Agency for the period 1993 to 2000. The employment statistic covers all employment subject to social security contributions.8 The information is given on the NUTS 3 level and refers to workplace location. We use employment data differentiated by nationality, educational level, branch, occupation, and rm size in order to generate our diversity measure and several control variables that enter into the regression model.

4 Regional disparities in cultural diversity of the labour force We apply different methods to quantify cultural diversity in order to check the robustness of results with respect to a variation in measurement. A rst indicator of cultural diversity is rooted in the literature on growth effects of ethnic fragmentation (e.g., Easterly and Levine 1997). In these studies, the probability that two randomly drawn individuals belong to two different groups is frequently used as a measure of fragmentation. The corresponding diversity measure is calculated as 1 minus the Herndahl index of concentration across groups:
2 DIV _ Hit = 1 sint n =1 Ni

(4)

where sint is the share of employees with nationality n among all employees of region i in year t. Ni is the number of different nationalities actually present in region i. Ottaviano and Peri (2006) note that this indicator accounts for both richness of the distribution, i.e. number of nationalities, and a relatively even distribution across nationalities. Thus, according to this measure, cultural diversity will increase if the number of nationalities rises or if the shares of different nationalities in employment converge. A disadvantage of this diversity measure is, however, that the index assigns disproportionately high weights to the largest nationalities. The result is largely driven by the share of the dominant population group, i.e. the natives. The Herndahl measure will therefore be highly correlated with the employment share of Germans (foreigners, respectively). An entropy index, such as the Theil index, offers a more adequate way of measuring cultural diversity than the Herndahl index. The corresponding diversity measure is dened as follows:
6 Four planning regions had to be merged due to restricted data availability. The classication has been developed by the Federal Ofce for Building and Regional Planning. For details see URL: http://www.bbr.bund.de/raumordnung/ europa/download/spesp_indicator_description_may2000.pdf 7 See Greif and Schmiedl (2002) for more detailed information on the patent data base. 8 Hence, civil servants and self-employed are not recorded in the employment statistic.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

Ni

569

DIV _ Tit = sint ln ( sint )

n =1

(5)

The maximum of the Theil index is ln(Ni), indicating that employment is evenly distributed across the Ni different nationalities in region i, i.e. sint = 1/Ni, "n. If the labour force consists of just one ethnic group, the index takes the minimum value ln(1) = 0. The marginal contribution of an additional worker to cultural diversity is ceteris paribus the higher, the smaller the ethnic group to which she belongs. The Theil index is a measure of cultural diversity that reects both the share and the variety of the foreign population in the region under consideration (Audretsch et al. 2009). And nally, we apply the so-called Krugman index to measure cultural diversity of employment:

DIV _ K it = sint
n =1

Ni

1 Ni

(6)

The index is calculated as the sum of absolute differences between the employment shares of nationalities and a reference that is given by 1/Ni. Thus the reference corresponds with an even distribution across nationalities. This implies that a value of DIV_Kit = 0.5 indicates that at least one quarter (0.5 DIV_Kit) of the total workforce has to change nationality for the employment distribution to correspond exactly to the reference distribution (see Bickenbach and Bode 2008). So in contrast to the diversity measures introduced above, a low Krugman index points to a relatively diverse workforce. Whereas most studies on the effects of cultural diversity are based on population data, we use employment data instead. The advantage of our measure is a closer connection to the production process. Moreover, nationality denes cultural identity of employees in the present analysis. Country of birth is the most widely used indicator in this context. However, information on country of birth is not available in German statistics. Applying nationality to determine cultural identity has advantages and drawbacks. Referring to nationality implies that naturalized citizens do not enter into the diversity measure as foreign persons. However, using country of origin as a denition of the foreign workforce implies that we do not consider people with a migration background born in Germany, unless we have information on the country of birth of the parents. Naturalized employees probably tend to be more successful with respect to educational attainment and labour market integration due to the terms of naturalization in Germany (minimum duration of stay, required language skills). Therefore, our diversity measure might be imprecise with respect to the highly qualied labour force and affected by a downward bias. The indicators of cultural diversity in this study are based on regional employment data differentiated by occupation, educational attainment, and nationality. We can differentiate between 3 levels of education (no formal vocational qualication, completed apprenticeship, university degree) and 213 nationalities. We use total employment and R&D employment to calculate diversity measures. R&D employment is dened on the basis of occupations. We consider engineers, chemists, physicists, mathematicians, technicians, other specialized technical staff, and natural scientists. Different diversity indices are calculated for total employment and R&D employees: aggregate measures as well as qualication-specic indices corresponding to the levels of education mentioned above. By considering the cultural diversity of the labour force at different qualication levels, we can check whether education matters, i.e. taking into account that it might be cultural diversity of highly qualied workers only that affects the process of innovation.
Papers in Regional Science, Volume 89 Number 3 August 2010.

570 Table 1. Cultural diversity of the labour force (Theil index), 2000 Total employment All skill groups Mean Standard deviation Maximum Minimum Agglomerated regions Urbanized regions Rural regions East Germany West Germany Germany 0.35 0.21 0.86 0.03 0.57 0.35 0.26 0.15 0.54 0.47 High skilled 0.22 0.11 0.47 0.03 0.33 0.22 0.16 0.13 0.34 0.29

A. Niebuhr

R&D employment All skill groups 0.15 0.10 0.47 0.01 0.28 0.16 0.15 0.09 0.26 0.23 High skilled 0.18 0.11 0.55 0.00 0.31 0.20 0.16 0.11 0.30 0.27

The share of foreign employees in Germany amounts to 7.1% in 2000.9 This corresponds with a value of the overall diversity measure (Theil index) of 0.47. Table 1 shows summary statistics and results of different diversity measures for regions types. Agglomerated regions achieve highest levels of cultural diversity, whereas rural areas are marked by a relatively low diversity of the workforce. Moreover, there are distinct differences between East and West Germany. In East German regions, cultural diversity is signicantly lower than in the western part of the country. The Herndahl and the Krugman index show similar spatial patterns. In particular, the regional distribution of the Herndahl and the Theil index correspond very closely. The Krugman index differs somewhat as density at the upper tail of the distribution tends to be higher than for the other measures. Furthermore, density increases more quickly at the lower tail of the distribution. There are also distinct differences between the diversity measures for high skilled, R&D, and total employment. Cultural diversity is higher among all workers than among high skilled and R&D employees. This might reect to some extent a downward bias in measurement of diversity for high skilled workers mentioned above. Due to the terms of naturalization, such as skill requirements, one might expect that naturalized workers are on average characterized by a higher level of educational attainment than foreign employees. This might result in a bias of a diversity measure that is solely based on foreign workers. But recent empirical evidence suggests that there are no signicant differences in the share of the high skilled among foreigners and the naturalized population in Germany (see Statistisches Bundesamt 2009). Moreover, the overall number of nationalities is likely to be lower in high skilled and R&D employment than in the total work force. However, for R&D employment, cultural diversity is higher among high skilled as compared to total R&D staff. The most diversied regions are Stuttgart, Munich and Frankfurt, highly agglomerated regions in the South-West of Germany (see Figure 1). There are no cities from the northern part of the country among the leading regions. Lowest diversity measures arise in East Germany, the only exceptions being Berlin and one neighbouring region. On average, East German regions do poorly as regards diversity of their R&D staff, most notably the rural peripheral areas. 5 Econometric issues We apply the knowledge production function introduced by Griliches (1979) to investigate the impact of cultural diversity of R&D employment on innovation. This function is widely used to
9

This refers to the labour force subject to social security contributions only.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

571

Fig. 1. Regional disparities in cultural diversity of R&D employment (Theil index), 2000

investigate the effects of R&D inputs on regional levels of innovation (e.g., Acs et al. 1994, Anselin et al. 1997). Referring to the theoretical outline in Section 2, if (Pi/divi) > 0, diversity fosters regional innovation because the positive impact associated with the production complementarity outweighs negative effects linked to a labour force marked by more diverse cultural
Papers in Regional Science, Volume 89 Number 3 August 2010.

572

A. Niebuhr

backgrounds. In contrast, (Pi/divi) < 0 implies that production complementarities are too weak to compensate for the negative effects associated with cultural diversity. We check whether positive or negative effects dominate in the regression analysis.

5.1 Basic specication The regression analysis is applied to the knowledge production function that links R&D inputs to innovation output, namely, patent applications. Since the number of patents is affected by the size of the regional economy, we investigate the relationship between patents and R&D input in per capita terms. R&D staff and R&D expenditure per inhabitant are used as proxies for R&D activity. The basic regression model is given by:

ln Pit = 0 + 1 ln RDit 1 + 2 ln HCit + 3 ln DIVit + k Ckit + uit

k =1

(7)

where Pit is the number of patents per capita in region i and year t. RDit-1 is R&D personnel or R&D expenditure per capita in year t-1 and uit is the error term. In order to appropriately model the relationship between R&D input and output, the input variable enters into the model with a time lag of one year. Patents as well as R&D input refer to rms only. We assume that knowledge is embodied in human capital (HCit) and use the share of high skilled employees (university degree) in total employment as a proxy for the regional knowledge stock. However, human capital might also foster the innovation process via facilitating knowledge spillovers. Finally, the inclusion of a human capital variable enables us to check whether diversity among highly qualied workers just works as an approximation of the human capital endowment of the region. With respect to the objective of the investigation, the coefcient of interest is a3 that captures the impact of cultural diversity of R&D workers on patent applications. The diversity index DIVit is calculated according to Equations (4) to (6), depending on the considered measure. Separate models are estimated for diversity measures based on total R&D employment and high skilled R&D employment.10 Furthermore, we expand the original knowledge production function by some control variables Ckit in order to avoid misspecication due to omitted variables. Controls comprise an indicator for the sectoral composition of regional economies, more precisely the ratio of manufacturing to service employment in the region STRUCit. The industry structure is considered because the propensity to patent is higher in manufacturing than in the service sector. According to Bode (2004), the propensity to patent might also be affected by the size of rms. In order to capture corresponding effects, two additional variables are considered: the employment shares of small (less than 20 employees) and large (500 or more employees) rms (SMALLit, LARGEit). As the innovation process in highly agglomerated areas may signicantly differ from the process in rural peripheral regions, we take into account the region type as well by means of a categorical variable REGTYPEi that differentiates between agglomerated, urbanized and rural (see Appendix for details). Finally, we allow for systematic differences in R&D activity between East and West Germany (Dummy EAST) that might for example, be due to disadvantages caused by some kind of heritage of the centrally planned socialist economy still at work.
10 We also considered local university research as a potential determinant of R&D output and included R&D staff at universities and polytechnics per inhabitant as an explanatory variable. However, results are rather disappointing since the coefcient of university research is insignicant in most specications. The poor performance of university research might be caused by the fact our data set does not allow to focus on applied research at universities and research institutes. Corresponding results are available upon request.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

573

5.2 Robustness checks To investigate the robustness of our empirical results, a number of additional regression models are applied. First, we have to consider potential effects of measurement errors and simultaneity. Due to endogeneity of explanatory variables, the relationship estimated by ordinary least squares (OLS) cannot be interpreted as causal. The estimated effect of diversity on R&D output might be biased due to potential endogeneity of cultural diversity. Innovation activity may inuence labour migration and thus, cultural diversity of the labour force since regions characterized by high innovation output, growth, and favourable labour market conditions are preferred destinations of immigrants. We use diversity measures calculated with low skilled employment lagged both in space and time (by 5 years) and the latitude of the region centre as instruments (Zit) for contemporaneous diversity indices. As an alternative instrument, we use the lagged share of foreigners in low skilled employment. The variables are valid instruments if they are relevant [corr(Zit,DIVit) 0] and uncorrelated with the errors term [corr(Zit,uit) = 0].11 Lagged cultural diversity in neighbouring regions is supposed to inuence current diversity in the region under consideration due to migration ows. As migration occurs primarily across shorter distances and is inuenced by herd behaviour (see Bauer et al. 2002; Burkert et al. 2008), cultural diversity in a region is likely to affect future diversity in neighbouring areas. Moreover, diversity among low skilled workers should reect the existence of amenities that will also attract high skilled workers. However, employees without a formal vocational qualication should not play an important role in innovation. The latitude is used as an approximation of the main destinations of immigrants in Germany in the 1950s and 1960s that are located in the Southwest of the country. Relevance of both instruments rests on the idea that migrants limit their destination choices to places with signicant prior immigration from the same origin. The instrument variables (IV) are highly correlated with the diversity measures and unlikely to be affected by reverse causation. This applies especially to the latitude variable as a pure geographic variable. Second, xed and random effects panel data models are applied so as to control for unobserved time-invariant effects:

ln Pit = 0 + 1 ln RDit 1 + 2 ln HCit + 3 ln DIVit + k Ckit + i + t + it

k =1

(8)

where hi denotes a region-specic effect, controlling for unobserved regional characteristics, lt captures time effects, and nit is a white noise error term. Moreover, evidence provided by Bode (2004) and Anselin et al. (1997) suggest that geographically bounded spillovers and spatial dependence are important for regional innovation activity. Therefore, we check for misspecication due to omitted spatial effects indicated by spatial autocorrelation in the regression residuals. Depending on the results of corresponding LM-tests, we might estimate spatial lag models or spatial error models. The spatial lag model is given by:

ln Pit = 0 + wij ln Pjt + 1 ln RDit 1 + 2 ln HCit + 3 ln DIVit + k Ckit + uit

j =1 k =1

(9)

Thus, we extend the non-spatial model by a spatial lag of the dependent variable

w
j =1

ij

ln Pjt where wij is an element of the R R spatial weights matrix W. We consider two

11 More precisely, relevance requires a partial correlation of the instrument with the endogenous regressor, namely, the coefcient of the instrument variable should be signicant in the rst stage regression.

Papers in Regional Science, Volume 89 Number 3 August 2010.

574

A. Niebuhr

alternative raw adjacency matrices W in this analysis. A rst and frequently applied specication is a binary spatial weights matrix such that wij = 1 if the regions i and j share a border and wij = 0 otherwise. Second, wij is set to the inverse of travel time between the capitals of regions i and j with a cut-off point at 150 km, i.e. wij = 0 if distance between regions centres exceeds this cut-off point. From the raw matrices W spatial weights matrices W are derived by normalizing such that j wij = 1, i . Taking into account the weighted sum of patent applications in neighbouring regions implies that spatial autocorrelation of the error term is due to omission of some substantive form of spatial dependence caused by interaction among regions. We assume that there might be important knowledge spillovers that result in interdependent innovation processes of adjacent R&D departments leading to spatial dependence in patent applications. Localized knowledge spillovers are likely to be caused by interregional mobility of R&D workers which in turn is affected by frictional effects of distance. Moreover, evidence in Scherngell and Barber (2009) suggests that geographical distance and co-localization of organizations in neighbouring regions are important determinants of R&D collaborations. Following Bode (2004), we apply the patents granted to researchers in neighbouring regions as a proxy for the knowledge stock in adjacent locations. In contrast, the spatial error model will be the appropriate specication if the misspecication is due to nuisance dependence. Spatial autocorrelation in measurement errors or in variables that are otherwise not crucial to the model might entail spatial error dependence. The spatial error model may be expressed as:

ln Pit = 0 + 1 ln RDit 1 + 2 ln HCit + 3 ln DIVit + k Ckit + uit

k =1

(10)

with uit = wij u jt + it as the spatial autoregressive error term.

j =1

Finally, we take into account that outlying observations might have a marked effect on the regression results. To address this issue, we apply quantile regressions as introduced by Koenker and Basset (1978). The median regression corresponds to the least absolute deviation estimator and is a robust alternative to OLS. Quantile regressions minimize an objective function which is a weighted sum of absolute deviations:

min yi xi + (1 ) yi xi i : yi < xi i: yi xi

(11)

Here yi is the dependent variable and xi is the vector of explanatory variables which is multiplied by the coefcient g. The objective function can be interpreted as an asymmetric linear penalty function of deviations from predicted to actual patents per capita. An important special case is the median regression (q = 0.5). Since this regression puts less weight on outliers than OLS, it is a robust alternative.

6 Regression results and discussion A point of departure of the regression analysis is a basic pooled model that is estimated with different versions of the pivotal explanatory variable, cultural diversity of R&D employment based on different diversity indicators. We apply the three measures introduced in Section
Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

575

4 to total R&D employment (DIVit) and high skilled R&D staff (DIVit high). Table 2 shows the results of the basic regression model. The specications in columns 1 to 6 only differ with respect to the diversity measure included. In line with previous evidence on the knowledge production function, we get a highly signicant impact of R&D expenditure on innovation output.12 A positive effect is also associated with the regions human capital endowment. However, the coefcient is only marginally signicant at the 10% level in the models 2 and 4, those specications including diversity among high skilled R&D staff only. Furthermore, some control variables appear with signicant coefcients, indicating that structural characteristics of the regions matter for innovation activity. The relative size of the industrial sector is associated with a signicant effect on innovation output. According to the estimates, a specialization of regions on manufacturing as compared to services tends to raise patents per capita. Furthermore, areas characterized by a relatively large share of small rms on average seem to perform better than other regions. The negative coefcient of the region type variable implies that there might be systematic differences between the innovation processes of metropolitan areas, urbanized and rural regions. More precisely, less densely populated regions, especially rural areas, tend to be marked ceteris paribus by lower productivity of R&D activity. This might point to some kind of positive agglomeration effect at work. However, the region type variable is marginally signicant in some specications only. Moreover, the signicant coefcient of the dummy variable suggests that there are still systematic differences in R&D activity between East and West Germany. Even after controlling for R&D input and other factors that inuence regional patent applications East German regions achieve a lower R&D output per capita than locations in the western part of the country. Finally, turning to the most important variable, the results point to an innovation-enhancing effect of cultural diversity of the workforce. The coefcient of cultural diversity is positive and highly signicant, irrespective of applied diversity measure. Only in model 6, is diversity not signicant at the 5% level. Thus, the impact of diversity is rather robust with respect to a variation of measurement. Differences in size of the coefcients, that show up especially for the Krugman index on the one hand and the Theil and Herndahl measures on the other hand, can be traced back to different denitions of the indices. Further, the impact of diversity among highly educated R&D employees is smaller than the effect that is detected for total R&D staff. Thus, the regression results indicate that cultural diversity is a factor which positively inuences the process of knowledge creation. The qualication level of labour might not matter in this context because the educational attainment of R&D workers and their communication skills tend to be on average relatively high. However, as indicated by the tests for spatial autocorrelation (Morans I, Lagrange multiplier error and Lagrange multiplier lag tests), regional R&D activity is marked by some spatial interaction not captured by the regression model so far. The differences between the test statistics suggest that problems are caused by omission of some kind of substantive form of spatial dependence that might rest upon knowledge spillovers between neighbouring regions.13 In order to check whether the identied impact of cultural diversity is affected by the omission of spatial dependence, we include a spatial lag of patent applications per capita in some specications. Moreover, the results of the Hausmann-Wu test for endogeneity of the diversity measures point to a potential problem that might adversely affect the estimates. However, for diversity of total R&D employment the test is only signicant at the 5% level. Nevertheless, we apply instrument variable (IV) regression in order to check the robustness of results with respect to potential endogeneity of our diversity variables.
Substituting R&D expenditure per capita by R&D personnel per capita does not signicantly change the results. Higher signicance of LM lag tests indicates that the spatial lag model is the appropriate specication. The corresponding decision rule is proposed by Anselin and Florax (1995).
13 12

Papers in Regional Science, Volume 89 Number 3 August 2010.

576

Table 2. Regression results for alternative diversity measures (Pooled OLS) ln(patents per capita) Herndahl Index (2) 2.66** (2.81) 0.36** (7.93) 0.31 (1.92) 2.83** (2.93) 0.36** (7.73) 0.26 (1.61) 0.19** (2.93) 0.69** (8.66) 0.80** (2.90) 0.13 (1.00) -0.07 (1.62) -0.41* (2.51) 0.87 190 6.69* 1.73 0.80 0.02 5.21* 4.43* 0.87 189 7.78** 2.08* 1.50 0.01 6.57** 5.08* 0.16** (3.06) 0.72** (9.08) 0.79** (2.76) 0.09 (0.68) -0.08 (1.79) -0.48** (3.35) 2.85** (2.98) 0.36** (7.91) 0.32 (1.96) (3) (4) (5) 8.08** (2.79) 0.38** (8.28) 0.17 (0.91) 9.22* (2.13) 0.68** (8.42) 0.89** (3.04) 0.16 (1.22) -0.09* (2.10) -0.51** (3.05) 0.87 190 4.99* 2.04* 1.35 0.00 6.02* 4.67* Krugman Index (6) 5.36* (2.48) 0.39** (8.45) 0.25 (1.47)

Dependent variable

Theil Index

(1)

Papers in Regional Science, Volume 89 Number 3 August 2010.

Constant ln(RDit-1) ln(HCit) ln(DIVit) ln(DIVit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) REGTYPEi Dummy EAST 0.19** (3.18) 0.71** (9.06) 0.80** (2.80) 0.09 (0.66) -0.07 (1.61) -0.47** (3.21) 0.87 189 7.95** 2.05* 1.45 0.01 6.72** 5.28*

2.69** (2.83) 0.35** (7.66) 0.25 (1.50) 0.24** (3.40)

0.69** (8.69) 0.80** (2.95) 0.13 (1.05) -0.06 (1.34) -0.35* (2.07)

4.84 (1.54) 0.69** (8.57) 0.91** (3.00) 0.15 (1.14) -0.10* (2.20) -0.61** (4.06) 0.87 190 5.76* 2.18* 1.70 0.00 8.15** 6.45*

Adj. R2 Observations Wu-Hausman F-Test Morans I LM error Robust LM error LM lag Robust LM lag

0.87 190 6.27* 1.59 0.61 0.05 4.86* 4.30*

Notes: t-statistics in parentheses are based upon Huber-White standard errors. ** signicant at the 0.01 level, * signicant at the 0.05 level. All models include time xed effects. Test on spatial autocorrelation were conducted with different weight matrices in order to check robustness. The results presented in the table are based on a weights matrix using inverse distance with a cut-off point of 150 km. A. Niebuhr

Migration and innovation Table 3. IV regression (2SLS) Dependent variable (1) Constant ln(RDit-1) ln(HCit) ln(DIV_Tit) ln(DIV_Tit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) REGTYPEi Dummy EAST Hansen J statistica Shea Partial R2 F-Test of excluded instruments 3.11** (3.00) 0.28** (4.74) 0.12 (0.63) 0.63** (3.37) 0.69** (8.02) 0.51* (2.35) 0.07 (0.55) 0.001 (0.02) 0.24 (0.73) 1.12 (0.2893) 0.20 21.71** ln(patents per capita) (2) 3.25** (2.87) 0.25** (3.83) 0.25 (1.22) 0.71** (3.27) 0.77** (7.50) 0.30 (1.13) -0.14 (0.97) -0.002 (0.02) 0.28 (0.76) 0.46 (0.4983) 0.11 10.62** (3) 3.11** (3.01) 0.28** (4.70) 0.12 (0.62) 0.63** (3.35) 0.69** (7.99) 0.50* (2.30) 0.06 (0.53) 0.002 (0.04) 0.25 (0.75) 0.17 (0.6813) 0.20 22.63** (4)

577

3.26** (2.86) 0.25** (3.75) 0.25 (1.20) 0.72** (3.31) 0.78** (7.44) 0.29 (1.05) -0.14 (1.00) 0.001 (0.01) 0.30 (0.82) 0.15 (0.7017) 0.10 10.46**

Notes: t-statistics in parentheses are based upon Huber-White standard errors. ** signicant at the 0.01 level, * signicant at the 0.05 level. All models include time xed effects. Latitude of region centre, diversity measures (Krugman index) lagged in space and time are used as instruments in models 1 and 2. Latitude and the lagged share of low skilled foreigners in total employment are applied as instruments in the models 3 and 4. a p-value in parentheses.

The results of the IV regressions indicate that endogeneity of cultural diversity is unlikely to be a major problem (see Table 3). The diversity measures are instrumented by diversity measures calculated with low skilled employment lagged both in space and time, lagged shares of foreigners in low skilled employment, and the latitude of the region. In the following, we focus on results for Theil diversity measures only. The impact of cultural diversity on innovation output is even reinforced in the IV regressions. All IV estimates are positive, signicant, and larger than their OLS counterparts. This is surprising since simultaneity should result in upward biased OLS estimates of the impact of cultural diversity. This suggests that the simultaneity bias in the OLS estimates is relatively small. The gap between OLS and IV estimates might reect a downward bias in the OLS estimates caused by measurement errors. This may indicate that measurement errors bias towards zero is more important than the upward bias due to the endogenous location choice of immigrants. Hunt and Gauthier-Loiselle (2008) provide similar evidence for US States. Another explanation is that there is heterogeneity in the impact of diversity on innovation output, and that the IV estimates tend to recover effects for a subset of regions with relatively strong impact of diversity on patent intensity.14 The quality of the IV results critically hinges on the quality of the instruments, i.e. whether they are relevant and exogenous. We apply the test of overidentifying restrictions to check instrument exogeneity. The results of the Hansen J-statistic suggest that we can not reject the hypothesis that the instruments are exogenous. The F-statistics of excluded instruments as well as Sheas Partial R2 indicate that the partial correlation between the instruments and the endogenous explanatory variables is sufcient to ensure unbiased estimates and relatively small standard errors.15
See Card (2001) for a corresponding reasoning with respect to returns to schooling. The F-statistic refers to the test of the hypothesis that the coefcients of the instruments equal zero in the rst stage of 2SLS. Staiger and Stock (1997) propose a F-statistic of at least 10. A statistic less than 10 indicates that the instrument variables are weak, in which case the 2SLS estimator is biased and inference is unreliable.
15 14

Papers in Regional Science, Volume 89 Number 3 August 2010.

578

A. Niebuhr

Before turning to the signicance of spatial interaction, we check whether unobserved region-specic effects are important and adversely affect the estimates of the pooled model. In Table 4 the results of the random and xed effects models as well as the between estimator are summarized. The results of the random effects models (columns 1 and 2) are very similar to the estimates of the OLS regression of the pooled data. The coefcient of R&D expenditure slightly declines but is still highly signicant. The impact of cultural diversity turns out to be very stable as well. The effect of diversity of total R&D employment is even reinforced. However, the ndings change dramatically in the xed effects models (columns 3 and 4). The coefcients of the diversity measures are insignicant, although still of the same sign at least for total R&D employment.16 In contrast, applying an alternative specication with xed effects on the NUTS 1 level, namely for those German Lnder that consist of several planning regions, does not signicantly change the results with respect to cultural diversity. This additional robustness check is conducted for the specication including the Theil index. The coefcients of the diversity measures slightly decline from 0.24 to 0.21 for ln(DIV_Tit) and from 0.19 to 0.16 for ln(DIV_Tit) high. Furthermore, the effect of cultural diversity is still signicant at the 5% level in both specications. The problem of the xed effects model might be linked to the quality of the data on R&D input, survey data that may be affected by measurement errors. Johnston and DiNardo (1997) note that estimates may be biased towards zero due to mismeasurement of explanatory variables. The attenuation bias can be aggravated by xed effects estimation, in particular if the explanatory variables are highly correlated across time, as is frequently the case when the time period between the two cross sections is small (see also Griliches and Hausman 1986). With respect to the data set used in the regression analysis, this applies to R&D expenditure per capita as well as to the diversity indices. Although there is a considerable variation across regions, there is much less variation in time.17 This is conrmed by the results of the between estimator (columns 5 and 6) which only uses variation between cross sectional units by applying OLS to the time-averaged Equation (7). The random effects estimators is the weighted average of the xed effects and the between estimator. Differences between random effects and between estimator results are, however, fairly small. Because of the completely implausible implications and the outlined methodological problems of the xed effects specication, we rely on the random effects model that conrms the OLS and IV results. We also check whether spatial interaction affects the estimates by applying spatial regression models. The results in Table 5 suggest that there are signicant spillover effects among neighbouring regions. The coefcient of the spatially lagged dependent variable is signicantly different from zero in almost all models, the only exceptions being the two stage least square (2SLS) estimates in column 6. Thus, the neighbourhood of innovative regions promotes regional R&D activity. This conrms the evidence on localized knowledge spillovers provided in Bottazzi and Peri (2003) and Bode (2004). The result is robust with respect to a variation of the spatial weights matrix (binary contiguity, inverse distance). However, the inclusion of spatially lagged patents per capita does not seriously affect the coefcient of cultural diversity. The main differences regarding the impact of diversity emerge between the maximum likelihood (ML) estimates and the 2SLS results. The size of the coefcient increases once we instrument cultural diversity in the 2SLS regressions. This is in line with IV results summarized in Table 3. Altogether, the basic ndings regarding the impact of cultural diversity on innovation are not changed in the spatial lag models.
16 This is also in line with evidence surveyed in Griliches (1990). Several studies indicate that patents are more correlated with R&D expenditure within the cross section than in the time series dimension. 17 Corresponding ndings on the within and between dispersion of the variables are summarised in Table A3 in the Appendix.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

Table 4. Panel models ln(patents per capita) Fixed effects (2) 3.90** (3.20) 0.24** (3.94) 0.50* (2.30) 6.34* (2.37) -0.21 (1.87) -0.52 (0.70) 0.20 (1.16) 1.09 (1.69) 0.08 (1.34) 0.78* (2.06) -0.07 0.93 0.13 0.64 (0.58) (1.48) (1.91) (1.89) 5.96* (2.32) -0.22 (1.87) -0.36 (0.53) (3) (4) (5) 2.73 (1.84) 0.39** (6.05) 0.23 (0.97) 0.21* (2.05) 0.66** (5.64) 1.06* (2.51) 0.18 (1.09) -0.06 (1.04) -0.38 (1.68) 0.51 0.21 0.22 189 0.00 0.90 0.87 190 Between regression (6) 1.97 (1.29) 0.39** (6.19) 0.26 (1.11)

Dependent variable

Random effects

(1)

Constant ln(RDit-1) ln(HCit) ln(DIV_Tit) ln(DIV_Tit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) REGTYPEi Dummy EAST 0.15* (2.04) 0.87** (7.50) 0.40 (1.58) 0.07 (0.55) -0.08 (1.38) -0.62** (3.33) 0.29 0.88 0.86 189 0.53 0.45 0.45 190

4.02** (3.34) 0.22** (3.56) 0.40 (1.84) 0.30** (3.55)

0.85** (7.42) 0.35 (1.42) 0.11 (0.84) -0.05 (0.88) -0.34 (1.74)

0.17 (1.86) 0.67** (5.70) 1.07* (2.52) 0.16 (0.92) -0.07 (1.18) -0.45* (2.26) 0.39 0.90 0.63 189

R2 within R2 between R2 overall Observations

0.35 0.88 0.87 190

Notes: t-statistics in parentheses are based upon Huber-White standard errors. ** signicant at the 0.01 level, * signicant at the 0.05 level. All models include time xed effects.

579

Papers in Regional Science, Volume 89 Number 3 August 2010.

580

Table 5. Spatial lag models ln(patents per capita) ML (binary) (2) 1.31 (1.30) 0.36** (8.39) 0.21 (1.33) 1.14 (1.13) 0.35** (8.02) 0.17 (1.11) 0.19** (2.56) 0.49** (5.01) 0.51* (2.16) 0.03 (0.26) -0.08* (1.98) -0.43** (2.67) 0.20** (3.07) 0.15** (2.68) 0.50** (5.11) 0.50* (2.05) -0.01 (0.04) -0.09* (2.19) -0.51** (3.76) 0.21** (3.18) 1.06 (1.06) 0.36** (8.46) 0.21 (1.41) (3) (4) (5) 2.89** (3.15) 0.30** (6.01) 0.15 (0.85) 0.51** (2.93) 0.61** (7.33) 0.81** (2.67) 0.12 (0.99) 0.02 (0.38) 0.16 (0.51) 0.47* (2.02) 0.35 0.20 32.6** 14.7** 1.85 ML (inverse distance) 2SLS (inverse distance) (6) 3.23** (3.13) 0.26** (4.42) 0.29 (1.53)

Dependent variable

(1)

Papers in Regional Science, Volume 89 Number 3 August 2010.

1.41 (1.39) 0.35** (7.96) 0.16 (1.01) 0.21** (2.94) 0.17** (3.07) 0.55** (6.13) 0.53* (2.02) 0.02 (0.18) -0.08 (1.84) -0.43** (2.99) 0.16** (2.96)

Constant ln(RDit-1) ln(HCit) ln(DIV_Tit) ln(DIV_Tit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) REGTYPEi Dummy EAST W_ln(Pit) (r)

0.54** (5.99) 0.55* (2.15) 0.06 (0.52) -0.07 (1.61) -0.35* (2.05) 0.15** (2.81)

0.65** (3.29) 0.72** (7.12) 0.52 (1.66) -0.11 (0.71) 0.02 (0.28) 0.24 (0.74) 0.35 (1.36) 0.33 0.11 31.8** 8.3** 3.73*

Shea Partial R2 W_ln(Pit) ln(DIV_Tit) F-Test of excluded instruments W_ln(Pit) ln(DIV_Tit) Wu-Hausman F-Test

Notes: t-statistics in parentheses are based upon Huber-White standard errors; ** signicant at the 0.01 level, * signicant at the 0.05 level. All models include time xed effects. Latitude of region centre, diversity measures (Krugman index) lagged in space and time are used as instruments for diversity measures in the models 5 and 6. Patent applications lagged in space and time is used as an instrument for the spatial lag of the endogenous variable. The results presented in the table are based on a binary weights matrix and a matrix using inverse distance with a cut-off point of 150 km. A. Niebuhr

Migration and innovation Table 6. Quantile regression Dependent variable (1) Constant ln(RDit-1) ln(HCit) ln(DIV_Tit) ln(DIV_Tit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) REGTYPEi Dummy EAST W_ln(Pit) (r) 2.55* (2.10) 0.34** (6.60) 0.26 (1.12) 0.24 (1.92) 0.64** (5.58) 0.63* (2.08) 0.05 (0.31) -0.10 (1.41) -0.47 (1.46) (2) 2.32 (1.67) 0.35** (5.49) 0.29 (1.30) 0.20* (2.24) 0.65** (6.05) 0.54 (1.27) -0.002 (0.01) -0.10 (1.36) -0.52* (2.29) ln(patents per capita) (3) 2.61* (2.05) 0.34** (6.28) 0.22 (0.99) 0.28* (2.47) 0.60** (5.39) 0.82 (1.91) 0.10 (0.55) -0.08 (1.05) -0.32 (1.16) 0.32 (1.71) (4)

581

2.12 (1.74) 0.36** (6.50) 0.36 (1.74) 0.13 (1.42) 0.57** (5.22) 0.69* (2.32) 0.01 (0.08) -0.07 (0.79) -0.64** (2.86) 0.39 (1.94)

Notes: t-statistics in parentheses are based upon bootstrap standard errors with 100 replications. ** signicant at the 0.01 level, * signicant at the 0.05 level. All models include time xed effects.

Furthermore, we investigate whether outlying observations affect the estimates by applying quantile regressions. Table 6 shows results for the basic specication and spatial lag models. The coefcients of the median regression are rather similar to the previous estimates for the corresponding diversity measures in Tables 2 and 5, indicating that the effect of cultural diversity is not subject to serious bias caused by outliers. Signicance levels of the coefcients decline, but the impact of cultural diversity of total R&D employment is still signicant at the 10% and 5% level (columns 1 and 3). However, as regards diversity among high skilled R&D staff, we detect an important impact in the non-spatial model only. Finally, we turn to the size of the effect of cultural diversity focusing on the evidence for total R&D employment since these results seem to be more robust than those for high skilled R&D workers. Total R&D employment in Germany amounted to roughly 1,971,000 employees in the year 2000. The share of foreign workers in R&D employment is 2.9% (58,000 foreign employees) and corresponds to a Theil index of 0.23. If the number of foreign R&D workers grows by 100%, and the increase of all nationalities is proportional to their size, the Theil index will rise to 0.41. Applying the lowest coefcient estimates (ML spatial lag model: 0.19), this change in cultural diversity of R&D employment will cause an increase of patents per capita by almost 12%. In contrast, the IV estimate of the spatial lag model (0.51) gives rise to an increase of patents per capita by 34% (direct effect).18 Thus, there is a signicant impact of diversity on innovation. However, one has to keep in mind that these effects result from a doubling of foreign R&D staff and rest on strong assumptions regarding the distribution of these workers across nationalities. Moreover, compared with evidence provided in Hunt and Gauthier-Loiselle (2008), at least the effect associated with the ML estimate seems to be rather small. According to their results, a one percentage point increase in the skilled immigrant share is associated with an increase of the mean patents per capita by 17% in the U.S. However, their analysis does not refer to the impact of cultural diversity on innovation. The benecial effects of immigration rests upon the above average contribution of college graduate immigrants to patenting that is caused by the relatively large share of foreign born graduates with science and engineering education. Our diversity measures are based on employment shares of more than 200 nationalities. Depending on the size of the group, immigration of specic nationalities might result in quite
18 We do not discuss the indirect and induced effects that are due to spillover effects among neighbouring regions in the spatial lag model.

Papers in Regional Science, Volume 89 Number 3 August 2010.

582

A. Niebuhr

different changes of diversity. The marginal contribution of an additional individual to cultural diversity is the higher, the smaller the ethnic group to which that individual belongs. Thus, the same overall increase of foreign R&D workers will involve a stronger increase of diversity if small nationalities realize disproportionately high growth. Therefore, the same growth of foreign workers might give rise to different effects on innovation because the impact of immigration on cultural diversity might vary.

7 Conclusions We have analysed the impact of cultural diversity of R&D employment on innovation for a cross section of German regions. In principle, the effect of diversity on innovation might be positive or negative since there is likely a trade-off between productivity effects and transaction costs associated with diversity. The regression results indicate that cultural diversity might indeed matter for innovation activity. The evidence points to differences in knowledge and capabilities of workers from diverse cultural backgrounds that may enhance performance of regional R&D sectors. The positive impact of diversity on innovation seems to outweigh negative effects. IV estimates suggest that we may interpret this correlation as a causal effect from diversity to innovation. This corresponds with evidence on the economic effects of cultural diversity provided by Ottaviano and Peri (2005, 2006) for the U.S. The ndings are also in line with studies that emphasize the importance of labour mobility for knowledge transfer and innovation (e.g., Almeida and Kogut 1999; Simonen and McCann 2008). Moreover, our results indicate that the benets of diversity are not conned to R&D employees with a university degree. However, this is a plausible result as the educational attainment of R&D workers and their communication skills might on average be relatively high. Thus, we cannot generalize our ndings since only cultural diversity among skilled R&D staff is considered, and the educational attainment of workers is likely to matter for the economic effects of cultural diversity. Theoretical literature on economic effects of cultural diversity stresses the signicance of institutions in this context. The implementation of growth-enhancing effects of diversity may require a specic set of rules, or regulatory framework. Signicant benets of diversity among skilled workers and the fact that our focus is on employed migrants suggest that institutions and regulatory framework concerned with education and labour-market integration of immigrants play a particular role in realizing the economic benets. Moreover, our ndings have implications for immigration policy that needs to be more in favour of high skilled foreigners if they foster innovation and growth. Growing international competition for high skilled labour has already led some governments to revise their immigration policies (Bauer and Kunze 2004). European countries, especially, developed new arrangements directed towards highly qualied migrants in order to catch up with countries that offer more favourable conditions for high skilled immigrants. However, despite recent changes, labour migration regulation is still rather restrictive in Germany, even for highly educated foreign nationals.

References
Acs Z, Audretsch DB, Feldman M (1994) R&D spillovers and recipient rm size. Review of Economics and Statistics 76: 336367 Alesina A, La Ferrara E (2005) Ethnic diversity and economic performance. Journal of Economic Literature 43: 762800 Almeida P, Kogut B (1999) Localization of knowledge and the mobility of engineers in regional networks. Management Science 45: 907917
Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation

583

Anderson R, Quigley JM, Wilhelmsson M (2005) Agglomeration and the spatial distribution of creativity. Papers in Regional Science 83: 445464 Anselin L, Florax JGM (1995) Small sample properties of tests for spatial dependence in regression models: Some further results. In: Anselin L, Florax JGM (eds) New directions in spatial econometrics. Springer, Berlin Anselin L, Varga A, Acs Z (1997) Local geographic spillovers between university research and high technology innovations. Journal of Urban Economics 42: 422448 Audretsch DB, Dohse D, Niebuhr A (2009) Cultural diversity and entrepreneurship. A regional analysis for Germany. Annals of Regional Science (forthcoming) DOI 10.1007/s00168-009-0291-x Bauer T, Epstein G, Gang IN (2002) Herd effects or migration networks? The location choice of Mexican immigrants in the U.S. IZA Discussion Paper No 551 Bauer T, Kunze A (2004) The demand for high-skilled workers and immigration policy. Brussels Economic Review 47: 5775 Berliant M, Fujita M (2008) Knowledge creation as a square dance on the Hilbert cube. International Economic Review 49: 12511295 Bickenbach F, Bode E (2008) Disproportionality measures of concentration, specialization, and localization. International Regional Science Review 31: 359388 Bode E (2004) The spatial pattern of localized R&D spillovers: An empirical investigation for Germany. Journal of Economic Geography 4: 4364 Bottazzi L, Peri G (2003) Innovation and spillovers in regions: Evidence from European patent data. European Economic Review 47: 687710 Burkert C, Niebuhr A, Wapler R (2008) Regional disparities in employment of high skilled foreigners. Determinants and options for labour-migration policy in Germany. Journal of International Migration and Integration 9: 383 400 Card D (2001) Estimating the returns to schooling: Progress on some persistent econometric problems. Econometrica 69: 11271160 Chellaraj G, Maskus KE, Mattoo A (2008) The contribution of skilled immigration and international graduate students to U.S. innovation. Review of International Economics 16: 444462 Constant A, Gataullina L, Zimmermann KF (2006) Ethnosizing immigrants. IZA Discussion Paper No 2040 Duranton G, Puga D (2000) Diversity and specialisation in cities: Why, where and when does it matter? Urban Studies 37: 533555 Duranton G, Puga D (2001) Nursery cities: Urban diversity, process innovation, and the life cycle of products. American Economic Review 91: 14541477 Easterly W, Levine R (1997) Africas growth tragedy: Policies and ethnic divisions. Quarterly Journal of Economics 111: 12031250 Feldman M, Audretsch DB (1999) Innovation in cities: Science-based diversity, specialization and localized competition. European Economic Review 43: 409429 Fujita M, Weber S (2004) Strategic immigration policies and welfare in heterogeneous countries. FEEM Working Paper No 2 Glaeser EL, Kallal HD, Scheinkman JA, Shleifer A (1992) Growth in cities. Journal of Political Economy 100: 11261152 Greif S, Schmiedl D (2002) Patentatlas Deutschland Ausgabe 2002. Dynamik und Strukturen der Erndungsttigkeit, Mnchen Griliches Z (1979) Issues in assessing the contribution of R&D to productivity growth. Bell Journal of Economics 10: 92116 Griliches Z (1990) Patent statistics as economic indicators: A survey. Journal of Economic Literature 28: 16611707 Griliches Z, Hausman J (1986) Errors in variables in panel data. Journal of Econometrics 31: 93118 Henderson V, Kuncoro A, Turner M (1995) Industrial development in cities. Journal of Political Economy 103: 10671090 Hunt J, Gauthier-Loiselle M (2008) How much does immigration boost innovation? NBER Working Paper No 14312 Jacobs J (1969) The economy of cities. Random House, New York Johnston J, DiNardo J (1997) Econometric methods. McGraw-Hill, New York Keely LC (2003) Exchanging good ideas. Journal of Economic Theory 111: 192213 Kerr W, Lincoln WF (2008) The supply side of innovation: H-1B visa reforms and U.S. ethnic invention. Harvard Business School Working Paper No 09-005 Kim J, Marchke G (2005) Labor mobility of scientists, technological diffusion, and the rms patenting decision. The Rand Journal of Economics 36: 298317 Koenker R, Bassett G (1978) Regression quantiles. Econometrica 46: 3350 Lazear EP (1999a) Culture and language. Journal of Political Economy 107: 95126 Lazear EP (1999b) Globalisation and the market for team-mates. The Economic Journal 109: C15C40

Papers in Regional Science, Volume 89 Number 3 August 2010.

584

A. Niebuhr

Lazear EP (2000) Diversity and immigration. In: Borjas GJ (ed) Issues in the economics of immigration. University of Chicago Press, Chicago Longhi S, Nijkamp P, Poot J (2008) Meta-analysis of empirical evidence on the labour market impacts of immigration. IZA Discussion paper No 3418 Ottaviano GIP, Peri G (2005) Cities and cultures. Journal of Urban Economics 58: 304337 Ottaviano GIP, Peri G (2006) The economic value of cultural diversity: Evidence from US cities. Journal of Economic Geography 6: 944 Ozgen C, Nijkamp P, Poot J (2009) The effect of migration on income convergence: Meta-analytic evidence, Tinbergen Institute Discussion Paper No 2009-022/3 Romer PM (1990) Endogenous technological change. Journal of Political Economy 98: S71S102 Scherngell T, Barber MJ (2009) Spatial interaction modelling of cross-region R&D collaborations: Empirical evidence from the 5th EU framework programme. Papers in Regional Science 88: 531546 Simonen J, McCann P (2008) Firm innovation: The inuence of R&D cooperation and the geography of human capital inputs. Journal of Urban Economics 64: 146154 Staiger D, Stock JH (1997) Instrumental variables regression with weak instruments. Econometrica 65: 557586 Statistisches Bundesamt (2009) Bevlkerung und Erwerbsttigkeit. Bevlkerung mit Migrationshintergrund Ergebnisse des Mikrozensus 2007. Statistisches Bundesamt, Wiesbaden Zimmermann KF (2005) European labour mobility: Challenges and potentials. De Economist 153: 425450

Appendix: data R&D data from Stifterverband fr die Deutsche Wissenschaft. R&D personnel (full-time equivalents) and R&D expenditure (1,000 euros) 1997, 1999. Patent data from Patentatlas Deutschland edition 2002. Patent applications 19952000. Employment data from the German Federal Employment Agency. Employment by nationality, qualication level and occupation 1993, 1995, 1998, 2000. Region type classication from the German Federal Ofce for Building and Regional Planning. Conditional on population density and the size of regional centres three groups of regions are distinguished: agglomerated (1), urbanized (2) and rural regions (3). Distance and travel time. Interregional travel time bases on estimates for NUTS 3 regions. Travel time for planning regions was generated by calculating weighted averages of NUTS 3 data. Data is available for NUTS 3 regions or planning regions. In the analysis, the functionally dened planning regions are applied. Information for planning regions is generated by aggregation of NUTS 3 level data in case data is only available for NUTS 3 regions. Planning regions are dened on the basis of NUTS 3 regions, i.e. they consist of several NUTS 3 regions that are linked by intense commuting.

Papers in Regional Science, Volume 89 Number 3 August 2010.

Migration and innovation Table A1. Summary statistics Mean Standard deviation 27.8 35,683 300 0.03 Minimum

585

Maximum

Patents per 100,000 inhabitants R&D expenditure per 100,000 inhabitants R&D staff per 100,000 inhabitants Employment share of high skilled employees (university degree) Diversity measures Theil index Total R&D employment High skilled R&D employment Herndahl index Total R&D employment High skilled R&D employment Krugman index Total R&D employment High skilled R&D employment Ratio of manufacturing to service employment Employment share of small rms (<20 employees) Employment share of large rms (>500 employees)

40.7 27,608 273 0.07

4.1 476 8.5 0.03

140.2 184,344 1,669 0.15

0.16 0.18 0.04 0.05 1.95 1.94 0.48 0.30 0.18

0.10 0.11 0.03 0.03 0.02 0.03 0.23 0.04 0.07

0.004 0.009 0.001 0.002 1.88 1.85 0.13 0.11 0.06

0.47 0.55 0.14 0.16 1.99 1.99 1.19 0.43 0.35

Table A2. Between and within dispersion of variables Standard deviation Overall ln(Pit) ln(RDit-1) ln(HCit) ln(DIV_Tit) ln(DIV_Tit) high ln(DIV_Hit) ln(DIV_Hit) high ln(DIV_Kit) ln(DIV_Kit) high ln(STRUCit) ln(SMALLit) ln(LARGEit) 0.97 1.15 0.37 0.84 0.85 0.93 0.91 0.012 0.014 0.51 0.16 0.41 Between 0.96 1.15 0.36 0.84 0.87 0.93 0.92 0.012 0.014 0.51 0.15 0.41 Within 0.15 0.14 0.04 0.09 0.12 0.10 0.12 0.001 0.002 0.04 0.04 0.04

Papers in Regional Science, Volume 89 Number 3 August 2010.

doi:10.1111/j.1435-5957.2010.00271.x

Migracin e innovacin: inuye la diversidad cultural en la actividad regional de I+D?

Annekatrin Niebuhr

Resumen. La investigacin terica reciente trata los costos econmicos y los benecios de la diversidad cultural relacionados con la inmigracin. Sin embargo, aun son escasas las pruebas empricas relacionadas con el impacto de la diversidad cultural sobre el desempeo econmico. Analizamos el efecto de la diversidad cultural de la fuerza laboral sobre solicitudes de patentes para una muestra transversal de regiones alemanas. Los resultados sugieren que las diferencias en conocimiento y capacidad de los trabajadores de orgenes culturales diversos mejoran el rendimiento de los sectores regionales de I+D. en cuanto a la innovacin, parece ser que los benecios de la diversidad superan con creces los costos que implican, por ejemplo en barreras de comunicacin. JEL classication: C21, J61, O31 Palabras clave: Diversidad cultural, innovacin regional, funcin de produccin de conocimiento, Alemania

2010 the author(s). Papers in Regional Science 2010 RSAI. Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden MA 02148, USA. Papers in Regional Science, Volume 89 Number 3 August 2010.