Linguistic Distance in Applied Economics 7 2.1 Introduction

Language and Cultural Barriers in Labor Markets and International Factor Mobility

INAUGURALDISSERTATION zur Erlangung der Würde eines Doktors der Wirtschaftswissenschaft der Fakultät für Wirtschaftswissenschaft der Ruhr-Universität Bochum

Kumulative Dissertation, bestehend aus fünf Beiträgen

vorgelegt von Diplom-Ökonom Sebastian Otten aus Rhede 2013 Dekan: Prof. Dr. Helmut Karl Referent: Prof. Dr. Thomas K. Bauer Koreferent: Prof. Dr. Christoph M. Schmidt Tag der mündlichen Prüfung: 11.02.2014 i

Contents

List of Tables ...... vi List of Figures ...... vii

1 Introduction and Overview 1

2 The Costs of Babylon – Linguistic Distance in Applied Economics 7 2.1 Introduction ...... 7 2.2 Measuring Linguistic Distance ...... 9 2.2.1 Previous Literature ...... 9 2.2.2 The Levenshtein Distance ...... 10 2.3 Language Fluency of Immigrants ...... 11 2.3.1 Data and Method ...... 13 2.3.2 Results ...... 15 2.4 International Trade ...... 16 2.4.1 Data and Method ...... 17 2.4.2 Results ...... 19 2.5 Conclusion ...... 22 2.A Appendix ...... 28

3 Linguistic Barriers in the Destination Language Acquisition of Immi- grants 35 3.1 Introduction ...... 35 3.2 Measuring Linguistic Distance ...... 38 3.3 Data ...... 45 3.4 Method ...... 48 3.5 Results ...... 50 3.6 Conclusion ...... 55 3.A Appendix ...... 68 3.B Supplementary Appendix ...... 72 CONTENTS ii

4 Language and Cultural Barriers in International Trade and Investment 80 4.1 Introduction ...... 80 4.2 Empirical Model ...... 86 4.2.1 Theoretical Background: The Structural Gravity Model ...... 86 4.2.2 Empirical Strategy ...... 88 4.3 Data ...... 94 4.3.1 Measuring Linguistic and Genetic Distance ...... 94 4.3.2 Data on Trade Flows, Portfolio Investment, and Banking Claims . . 97 4.4 Empirical Findings ...... 98 4.4.1 The Eﬀect of Linguistic and Genetic Distance on Trade ...... 98 4.4.2 The Eﬀect of Linguistic and Genetic Distance on Investment . . . . 101 4.5 Conclusion ...... 102

5 The Role of Language Skills in the German Labor Market 114 5.1 Introduction ...... 114 5.2 Data ...... 116 5.3 Empirical Strategy ...... 119 5.4 Results ...... 121 5.5 Conclusions ...... 123

6 The Role of Source- and Host-Country Characteristics in Female Immi- grant Labor Supply 130 6.1 Introduction ...... 130 6.2 Background ...... 132 6.3 Method, Data, and Descriptive Statistics ...... 137 6.3.1 Method ...... 137 6.3.2 The European Social Survey ...... 138 6.3.3 Aggregated Data ...... 143 6.4 Basic Results ...... 149 6.4.1 Source- and Host-Country Fixed Eﬀects ...... 149 6.4.2 Source-Country FLFPR ...... 151 6.4.3 Host-Country FLFPR ...... 155 6.5 Sensitivity Analyses ...... 158 6.5.1 Control for Partner Characteristics ...... 158 6.5.2 Control for Parents’ Human Capital and Employment ...... 159 6.5.3 Ratio of FLFPR to MLFPR ...... 160 6.5.4 Source-Country Characteristics at Year of Migration ...... 162 6.5.5 Bias-Reduced Linearization of Standard Errors ...... 162 CONTENTS iii

6.6 Conclusion ...... 164 6.A Appendix ...... 177

Bibliography 182

Acknowledgments 196

Curriculum Vitae 198 iv

List of Tables

2.1 Closest and Furthest Language Pairs with Respect to the Levenshtein Distance 24 2.2 Descriptive Statistics of Dependent and Explanatory Variables – Immigra- tion Sample ...... 24 2.3 Immigrant’s Language Skills – Probit Results ...... 25 2.4 Descriptive Sample Statistics and Variable Definitions – International Trade Sample ...... 26 2.5 Effect of Language on Bilateral Trade – OLS Results ...... 27 2.A1 40-Items Swadesh Word List ...... 28 2.A2 Summary Statistics for the Language Variables – International Trade Sample 29 2.A3 Immigrant’s Language Skills – Probit Results, Including Native Speakers . 32 2.A4 Effect of Language on Bilateral Trade – OLS Results, Subsample Language Barrier > 0 ...... 33 2.A5 Effect of Language on Bilateral Trade – OLS Results, Subsample Linguistic Features Index ...... 34

3.1 Average Test Scores of US Language Students ...... 58 3.2 40-Items Swadesh Word List with Computational Examples ...... 59 3.3 Closest and Furthest Languages to English and German ...... 60 3.4 Rank Correlations among Linguistic, Geographic, and Genetic Distance Measures ...... 61 3.5 Distribution of Language Skills across Samples ...... 62 3.6 Language Ability and Linguistic Distance – Aggregated Results ...... 63 3.7 OLS Results of Linguistic Distance – ACS Sample ...... 64 3.8 Ordered Logit Marginal Eﬀects of Linguistic Distance – Model 1 ACS & SOEP Sample ...... 64 3.9 OLS Results of Linguistic Distance – SOEP Sample ...... 65 3.A1 Descriptive Statistics – ACS & SOEP Sample ...... 68 3.A2 Variables Description – ACS & SOEP Sample ...... 69 LIST OF TABLES v

3.A3 Robustness Checks: OLS Results of Linguistic Distance – Model 2 ACS Sample ...... 70 3.A4 Robustness Checks: OLS Results of Linguistic Distance – Model 2 SOEP Sample ...... 71 3.B1 OLS Results – Model 1 & 2 ACS Sample ...... 72 3.B2 Ordered Logit Results – Model 1 & 2 ACS Sample ...... 73 3.B3 Ordered Logit Marginal Effects – Model 1 ACS Sample ...... 74 3.B4 Ordered Logit Marginal Effects – Model 2 ACS Sample ...... 75 3.B5 OLS Results – Model 1 & 2 SOEP Sample ...... 76 3.B6 Ordered Logit Results – Model 1 & 2 SOEP Sample ...... 77 3.B7 Ordered Logit Marginal Effects – Model 1 SOEP Sample ...... 78 3.B8 Ordered Logit Marginal Effects – Model 2 SOEP Sample ...... 79

4.1 Closest and Furthest Language Pairs with Respect to the Levenshtein Distance104 4.2 Effect of Linguistic and Genetic Distance on Bilateral Exports ...... 105 4.3 Effect of the Bilateral LDE Indicator on Bilateral Exports ...... 106 4.4 Effect of the Linguistic Distance toward English on Bilateral Exports . . . 107 4.5 Effect of Linguistic and Genetic Distance on Cross-Border Asset Stocks . . 108 4.6 Effect of Linguistic and Genetic Distance on Bilateral Banking Claims . . . 109 4.7 Effect of the Bilateral LDE Indicator on Cross-Border Asset Stocks . . . . 110 4.8 Effect of the Bilateral LDE Indicator on Bilateral Banking Claims . . . . . 111 4.9 Effect of the Linguistic Distance toward English on Cross-Border Asset Stocks112 4.10 Effect of the Linguistic Distance toward English on Bilateral Banking Claims113

5.1 Swadesh 40-Item List with Computational Examples ...... 124 5.2 Distance from German – Closest and Furthest Languages ...... 124 5.3 Descriptive Statistics by Oral German Ability ...... 125 5.4 Results of Employment Regressions ...... 127 5.5 Results of Wage Regressions – Oral Ability ...... 128 5.6 Results of Wage Regressions – Written Ability ...... 129

6.1 Descriptive Statistics – Individual Variables ...... 168 6.2 Descriptive Statistics – Aggregated Variables ...... 169 6.3 Model 1 – Source- and Host-Country Fixed Eﬀects ...... 170 6.4 Model 2 – Source-Country Characteristics ...... 171 6.5 Model 3 – Host-Country Characteristics ...... 172 6.6 Models 2 & 3 – Controlling for Partner Characteristics ...... 173 6.7 Models 2 & 3 – Controlling for Parents Characteristics ...... 174 LIST OF TABLES vi

6.8 Models 2 & 3 – Ratio of FLFPR to MLFPR ...... 175 6.9 Model 2 – Source-Country Characteristics at Year of Migration ...... 175 6.10 Models 2 & 3 – Bias-Reduced Linearization of Standard Errors ...... 176 6.A1 Explanatory Power of Source- & Host-Country Fixed Eﬀects ...... 177 6.A2 Explanatory Power of Source- & Host-Country Characteristics ...... 177 6.A3 List of Source Countries ...... 178 6.A4 Macroeconomic Data – Sources and Descriptions ...... 179 vii

List of Figures

2.A1 Comparisons of Linguistic Distance Using the Test-Score-Based Measure and the Levenshtein Distance – 2000 U.S. Census ...... 28 2.A2 Bivariate Kernel Density Estimation of Log Bilateral Trade and Levenshtein Distance – International Trade Sample ...... 30

3.1 Language Relations in the TREE Approach ...... 58 3.2 Predicted Language Assimilation Proﬁles for the ACS Sample ...... 66 3.3 Predicted Language Assimilation Proﬁles for the SOEP Sample ...... 67

5.1 German Ability, Employment Rate, and Hourly Wages by Years since Migration (5-Year Moving Average) ...... 126

6.1 Female Labor Force Participation Rate (Age 15-64) – Year 2011 ...... 166 6.2 Eﬀect of Source-Country FLFPR by Years since Migration ...... 167 6.3 Eﬀect of Host-Country FLFPR by Years since Migration ...... 167 The limits of my language mean the limits of my world. —Ludwig Wittgenstein, Tractatus Logico-Philosophicus (1922), p. 149.

Chapter 1

Introduction and Overview

Extreme poverty is a problem observed in all parts of the world. More than 1.2 billion people around the world live in extreme poverty, being defined as having to live from less than US$ 1.25 per day (United Nations, 2013). In addition, more than 870 million people world-wide suffer from extreme hunger and are chronically undernourished (IFPRI, 2013). In order to address these problems, the United Nations has set the fight against extreme poverty and hunger on top of the agenda of its millennium goals (United Nations, 2000). Both theoretical and empirical economic research suggest that international trade and migration can contribute to a reduction of differences in income between countries, initiate a progress of convergence in development, and that way help to overcome poverty (Freeman, 2006; OECD, 2011; Winters et al., 2004).1 Despite the positive impact of international migration, only around 3% of the world’s population live outside their country of birth (IOM, 2013). The relatively low share of migrants might partly be a result of the strict immigration legislation in the typical immigration countries (Freeman, 2006). While migration flows are hampered by the immigration legislation of many countries, the markets for goods and capital underwent a comprehensive liberalization of trade policies along with a reduction of trade barriers as a consequence of the collapse of the Bretton Woods system in 1973 (Wacziarg and Welch, 2007). This lead to a steady increase in the international flows of goods and capital. Nevertheless, not all countries have profited from the ongoing globalization of the international markets for goods and capital to the same extent. Even for countries with similar endowments and institutional frameworks, significant differences in trade volumes can be observed. This poses the question of what causes the differences in trade volumes. Theoretical and empirical economic studies provide clear answers and stress the importance of a country’s

1As John K. Galbraith states: “Migration is the oldest action against poverty. It selects those who most want help. It is good for the country to which they go; it helps break the equilibrium of poverty in the country from which they come. What is the perversity in the human soul that causes people to resist so obvious a good?” (Galbraith, 1979).

1 CHAPTER 1. INTRODUCTION AND OVERVIEW 2 geographical and climatic features, the endowment with resources, the technological level, and the institutional framework. However, while some important drivers of international trade have been identified, the phenomenon is not completely understood. What remains unexplained is how linguistic and cultural differences between two countries affect their bilateral trade and capital flows. A similar thought experiment can be applied to the relatively small number of international migrants. Again, economic models and empirical studies provide numerous explanations. However, no satisfactory answer as to how linguistic and cultural differences affect international migration flows is given. Next to the question of what influences the selection of destination countries, the analysis of the factors that determine the immigrants’ integration process is of particular importance. Immigrants differ largely with respect to their success of integrating into their destination country’s society. Integration is mainly achieved by learning the destination country’s language and by the adaption to its culture. If the social and economic integration fails, the positive impact of immigration is undermined. Failed integration leads to a lack of perspectives and increases the immigrants’ risk of becoming subject to poverty again, though this time not in their country of origin, but in an unfamiliar society. Furthermore, the low level of income reduces the ability of immigrants to send remittances to their families in the country of origin. Anecdotal evidence points to the particular importance of linguistic and cultural differences for international transactions and migration.2 Nevertheless, these factors are mostly neglected in the standard economic models. This dissertation aims at partly filling this research gap. For this purpose, the following five chapters deal with empirical analyses of the influence of linguistic and cultural differences on different aspects of economic behavior.3 To analyze the influence of linguistic and cultural differences within an empirical framework, a suitable measure is required that permits such an analysis and that can be implemented into econometric models. Since the economic literature provides only a method for the measurement of cultural differences – the genetic distance – Chapter 2 introduces the Levenshtein distance as a new measure of linguistic distance. Subsequently, this method is tested in both a micro- and macroeconomic application and is compared to the most frequently used measures in both areas of research. Chapters 3 and 4 extend the empirical applications presented in Chapter 2. In Chapter 3, the linguistic distance measure is compared to three other measures of language differences which

2A prime example is the failed fusion of the automobile companies Daimler and Chrysler, which is to large parts attributed to differences in their management culture (Michler, 2011). 3The chapters of this thesis are available as independent articles and working papers. The chapters 2 and 3 are published and the chapters 4–6 are currently under revision. Electronic preprints are available from the author. CHAPTER 1. INTRODUCTION AND OVERVIEW 3 have been used in the economic literature. In doing so, the measures are analyzed within a microeconometric framework with respect to their explanation of the variation in the language proficiency of immigrants. In Chapter 4, the influence of linguistic and genetic distance on bilateral flows of goods and capital is analyzed in the context of a macroeconometric model. The second part of the dissertation is concerned with the labor market. Chapter 5 analyzes how differences in language skills affect the employment probabilities and wages of immigrants. Chapter 6 then focuses on identifying the role of source- and host-country culture in immigrant women’s labor supply. In the following, the contributions of this thesis to the economic literature are clarified and the main findings and implications of the succeeding chapters are summarized. Chapter 2 introduces the normalized and divided Levenshtein distance as the method of choice to quantify differences between languages and discusses its advantages over previous measures of linguistic distance. The measure is then used in two applications to explain the costs of linguistic differences on the micro- and macro-level: (i) the analysis of the language acquisition of immigrants and (ii) the analysis of linguistic barriers in international trade flows. On the micro-level, we use multiple datasets to estimate the initial disadvantages in the language acquisition of immigrants resulting from differences in linguistic origin. Estimations using the U.S. Census allow for a direct comparison of our approach to the approach by Chiswick and Miller (1999), who measure linguistic distance toward English using average test scores of language students. Both approaches lead to qualitatively comparable results. We further use the general applicability of our measure to broaden previous evidence to non-Anglophone countries. In doing so, we use data from the National Immigrant Survey of Spain and the German Socio-Economic Panel. The results reveal that immigrants who come from a more distant linguistic origin face significantly higher costs of language acquisition, which is reflected in their lower probability of reporting good language skills. On the macro-level, we apply the Levenshtein distance in the context of international trade, where language barriers have previously been addressed by controlling for common languages in bilateral trade flows. Using a comprehensive dataset of bilateral trade flows by Rose (2004), we estimate a standard gravity model using the Levenshtein distance as an additional explanatory variable and compare this approach to a previous approach based on shared linguistic features by Lohmann (2011). The results provide new and strong evidence indicating that language barriers affect trade above and beyond the simple effect of sharing a common language. Chapter 3 extends the analysis of the language acquisition of immigrants in the preceding chapter by comparing the Levenshtein distance to three other approaches CHAPTER 1. INTRODUCTION AND OVERVIEW 4 previously used in further applications in the economic literature to measure linguistic distance: (i) the WALS measure, which uses differences in language characteristics, (ii) the TREE measure, which is based on a priori knowledge on language families, and (iii) a measure based on average test scores of native U.S. foreign language students (SCORE). The information on language differences is applied to German and U.S. micro data – the German Socio-Economic Panel and the American Community Survey – in order to provide a comprehensive analysis of the influence of the linguistic origin on the acquisition of the destination language proficiency. The results suggest that the linguistic barriers raised by language differences play a crucial role in the determination of the destination-country language proficiency of immigrants. Regardless of the method employed, we estimate large initial disadvantages by linguistic distance for immigrants both in the U.S. and in Germany. In Germany, these initial differences in language skills decrease with a moderate convergence over time. Contrarily, in the U.S., the initial disadvantages increase over time. We interpret this difference in assimilation patterns as a potential outcome of stronger enclave effects in the U.S. This crucial difference highlights the importance of extending the analysis beyond the case of Anglophone countries. Chapter 4 extends the macro-economic analysis of linguistic barriers by evaluating the extent to which language and cultural barriers affect different types of international factor movements, i.e., international trade flows, cross-holdings of assets, and consolidated international banking claims. In addition to disentangling the impact of language and cultural dissimilarities, which might vary substantially over the aforementioned factor movements, I analyze the effect of English proficiency on international factor movements, an impact factor so far neglected in the literature. In the empirical analysis, I apply a gravity model, which was first proposed by Tinbergen (1962) and has since then been applied in numerous empirical studies on factor mobility. The results show that linguistic and genetic distance have varying effects on the examined factor movements. While controlling for a host of other possible determinants, I find strong evidence that a higher linguistic distance between two countries reduces cross-border trade and investment holdings between these countries. The results for genetic distance, however, are mixed. While cultural differences significantly reduce bilateral trade, they have no effect on international investments. When including the country’s linguistic distance toward English in the model, the estimates indicate a significant negative impact of a higher linguistic distance toward English on international factor mobility. These findings are in line with the theoretical expectations and provide supportive evidence that language differences contribute to higher informational frictions across countries, thereby reducing bilateral trade and cross-border investment flows, respectively. CHAPTER 1. INTRODUCTION AND OVERVIEW 5

The second part of the thesis, Chapters 5 and 6, analyzes the role of language and culture in the labor market. Chapter 5 investigates the effects of language skills on labor market outcomes of immigrants in Germany using data from the German Socio-Economic Panel. Knowledge about the extent to which language skills affect employment probabilities and wages may improve our understanding of the underlying mechanisms of a successful integration of immigrants and illustrate the need and scope for government intervention such as the provision of student loans or free language courses. To address the problem of unobserved heterogeneity4 and a potential measurement error, we employ an instrumental variable approach to identify the causal effect of language skills on labor market outcomes. The instrument is based on the relationship between immigrants’ duration of residence in their host country and their language skills, taking into account heterogeneity in the linguistic distance among immigrants from different countries of origin. We find that the effect of language skills on employment probabilities is insignificant, which is in line with the economic literature on residential segregation of immigrants. In particular, it seems likely that geographic clustering allows immigrants to find jobs even without knowledge of the host-country language. In contrast, we observe a significantly positive effect of language skills on wages. However, this effect diminishes when we control for occupation, indicating that the returns to language skills are a result of the sorting of immigrants across occupations. We further demonstrate that simple OLS regressions systematically underestimate the positive effects of language skills on wages. Chapter 6 investigates the labor force participation of female immigrants in Europe. A central aim of this chapter is to provide evidence on the role of culture in women’s labor market behavior. Specifically, we are interested in whether immigrant women’s labor supply in their host country is affected by the female labor force participation rate in their source country, which serves as a proxy for the country’s preferences and beliefs regarding women’s roles. The effect of source-country culture on immigrants’ behavior, however, might weaken as immigrants assimilate to the culture of their host country and adapt to the labor supply behavior of natives. A second aim of this paper is to shed light on such assimilation patterns by investigating the role of host-country female labor force participation in immigrant women’s labor supply decisions. In the empirical analysis, we employ data from five rounds of the European Social Survey (ESS), covering immigrants in 26 European countries surveyed between 2002 and 2011. These data are augmented with an extensive set of aggregated source- and host-country variables as well as bilateral data describing the relationship between both countries, such as the geographic, linguistic, and genetic distance between the immigrants’ source and host country.

4In particular, the identiﬁcation of a causal eﬀect of language skills on labor market outcomes is challenging because language skills and labor market outcomes are both determined by unobserved individual ability. CHAPTER 1. INTRODUCTION AND OVERVIEW 6

We find that the labor supply of both first- and second-generation immigrants is positively associated with the FLFPR in their (parents’) source country. This result supports previous evidence for immigrants in the U.S. and suggests that immigrant women’s labor supply is affected by preferences and beliefs regarding women’s roles in society in her source country. The effect of this cultural proxy on the labor supply of immigrant women is robust to controlling for spousal characteristics, parental characteristics, and a variety of source-country characteristics. Moreover, we find evidence for a strong positive correlation between the FLFPR in the immigrant’s host country and immigrant women’s decision to participate in the labor market. This result suggests that immigrant women adapt to the culture, institutions, and economic conditions in their host country and that way assimilate to the work behavior of natives. Again, this result is robust to various sensitivity analyses. Taken together, the results of this dissertation stress the importance of linguistic and cultural differences for many aspects of economic behavior. Both in theoretical and empirical economic models, the standard variables should be complemented by factors that capture language and cultural differences. The proper modelling and implementation of these factors allows to gain new insights and can help to address unsolved puzzles. An example in this context is the irrational preference of investors to invest disproportionally high shares of their investment portfolio in their home country, which contradicts economic theory. The analyses in this dissertation have shed light on the fact that language and cultural differences impose burdens on the integration of immigrants. As a consequence, groups of immigrants that face higher linguistic and cultural hurdles are disadvantaged relative to other groups of immigrants. Due to this disadvantage, they have a lower employment probability and lower wages, which in turn increases their risk of poverty. Here, it is up to political decision makers to consider this disadvantage and address it with specifically designed measures. Public debates on immigrants’ abilities to integrate into their host country, as conducted in Germany and other industrialized countries, are often polemical. The fact that these debates do not consider the factors mentioned above stresses the necessity of particular target-oriented political actions. Language and cultural differences, in particular in relation to the leading industrial nations, negatively affect bilateral trade and capital flows. They act as an additional export duty and thus lead to a disadvantage on the globalized markets for goods and capital. This disadvantage affects developing countries in particular. It is up to the World Trade Organization to counteract this disadvantage and ensure fair competition on the international goods and capital markets. 7

Chapter 2

The Costs of Babylon – Linguistic Distance in Applied Economics∗

2.1 Introduction

According to biblical accounts, the Babylonian Confusion once stopped quite effectively the construction of the tower of Babel and scattered the previously monolingual humanity across the world, speaking countless different languages. In economic research, linguistic diversity is believed to be a crucial determinant of real economic outcomes, due to its impact on communication and language skills (see, e.g., Chiswick and Miller, 1999), and as accumulated costs affecting international trade flows (see, e.g., Lohmann, 2011). The operationalization of differences between languages is not straightforward and only few, but problematic, approaches have been undertaken so far. This study proposes to use a measure of linguistic distance developed by linguistic researchers. Linguistic distance is defined as the dissimilarity of languages, including, but not restricted to, vocabulary, grammar, pronunciation, scripture, and phonetic inventories. The Automatic Similarity Judgment Program (ASJP) by the German Max Planck Institute for Evolutionary Anthropology offers a descriptive measure of phonetic similarity: the normalized and divided Levenshtein distance. This distance is based on the automatic comparison of the pronunciation of words from different languages having the same meaning. We use this measure in two applications to explain the costs of linguistic differences on the micro-

∗Co-authored with Ingo E. Isphording (IZA). This chapter is published in the Review of International Economics, 21(2), 2013. A preliminary version of this chapter is available as Ruhr Economic Paper #337. The authors are grateful to Thomas K. Bauer, John P. Haisken-DeNew, Ira N. Gang, Julia Bredtmann, Jan Kleibrink, Maren Michaelsen, the participants of the RGS Doctoral Conference 2012, the RES 2012, the SOLE 2012, and the ESPE 2012 for helpful comments and suggestions. We are also very thankful to Andrew K. Rose for providing parts of the trade dataset and Johannes Lohmann for the data of his language barrier index. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 8 and macro-level. On the micro-level, we use multiple datasets, the 2000 U.S. Census, the National Immigrant Survey of Spain and the German Socio-Economic Panel, to estimate the initial disadvantages due to differences in linguistic origin in the language acquisition of immigrants. On the macro-level, trade-flow gravity models are estimated using the bilateral trade-flow data by Rose (2004) to analyze accumulated costs of linguistic barriers in international trade. Epstein and Gang (2010) point out that differences in culture, though crucially affecting economic outcomes, are typically treated as a black box in empirical investigations. One main channel of this effect of cultural distance on economic outcomes are differences arising from different linguistic backgrounds. Differences in language are arguably the most visible manifestation of such cultural differences. Previous studies relied on approaches that measure linguistic distance using average test scores of language students (Chiswick and Miller, 1999) or classifications by language families (Guiso et al., 2009). Test-score-based approaches assume the difficulty of learning a foreign language for students to be determined by the distance between the native and a foreign language. Unfortunately, due to data limitations test-score-based measures are only available for the distances toward the English language and are therefore strongly restricted in its use. Approaches using language family trees to derive measures of linguistic distance rely on strong assumptions of cardinality and have to deal with arbitrarily chosen parameters. Against this background, we contribute to the existing literature in several respects. First, we introduce the normalized and divided Levenshtein distance as an easy and transparently computed, cardinal measure of linguistic distance. We use the general applicability of this measure to broaden the evidence on disadvantages in the language acquisition of immigrants to non-Anglophone countries. Second, we apply the measure in the context of international trade, where language barriers have previously been addressed by controlling for common languages in bilateral trade flows. The Levenshtein distance allows to overcome this very narrow definition of linguistic barriers. Our results confirm the existence of significant costs of language barriers on the micro- and macro-level. Immigrants coming from a more distant linguistic origin face significantly higher costs of language acquisition. A higher linguistic distance strongly decreases the probability of reporting good language skills. To illustrate the results for immigrants into the U.S., a Vietnamese immigrant coming from a very distant linguistic origin faces an initial disadvantage compared to a German immigrant from a close linguistic origin which is worth of 6 additional years of residence. In the case of accumulated costs on bilateral trade flows, our results indicate that not only a shared common language but also a related but not identical language accelerates trade by lowering transaction costs. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 9

The paper is organized as follows. Section 2.2 provides a short overview of previous attempts to measure linguistic distance and introduces the Levenshtein distance, discussing its advantages and potential shortcomings. We present our results concerning the explanation of immigrants’ language skills in Section 2.3. The second application, the explanation of international trade ﬂows, is discussed in Section 2.4. Section 2.5 summarizes the results and concludes.

2.2 Measuring Linguistic Distance

2.2.1 Previous Literature

Linguistic distance is the dissimilarity of languages in a multitude of dimensions, such as vocabulary, grammar, pronunciation, scripture, and phonetic inventories. This multidimensionality of linguistic distance makes it difficult to find an appropriate empirical operationalization to be used in applied economic studies. A very straightforward approach is the evaluation of linguistic distances between languages by counting shared branches in language-family-trees (see, e.g., Guiso et al., 2009). This language-tree approach has to deal with strong cardinality assumptions and arbitrarily chosen parameters. Additionally, the approach offers only low variability between different language pairs and is difficult to implement for isolated languages such as Korean. A widely used approach to measure linguistic distance has been introduced by Chiswick and Miller (1999), who use data on the average test score of U.S. language students after a given time of instruction in a certain foreign language. They assume that the lower the average score, the higher is the linguistic distance between English and the foreign language. Similar measures have been used to analyze the effect of language barriers on international trade (Hutchinson, 2005; Ku and Zussman, 2010). This test-score-based measurement of linguistic distance relies on strong assumptions. It has to be assumed that the difficulty of U.S. citizens to learn a particular foreign language is symmetric to the difficulty of foreigners to learn English. Further, it has to be assumed that the average test score is not influenced by other language-specific sources. Dörnyei and Schmidt (2001) give an overview of the potential role of intrinsic and extrinsic motivation in learning a second language. Intrinsic motivation, the inherent pleasure of learning a language, and extrinsic motivation, the utility derived from being able to communicate in the foreign language, are likely to differ across languages, but are not distinguishable from the actual linguistic distance in the test-score-based approach. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 10

2.2.2 The Levenshtein Distance

Drawing from linguistic research, it is possible to derive an operationalization of linguistic distance without strong identification assumptions that underlie previous approaches. The so-called Automatic Similarity Judgment Program (ASJP) developed by the German Max Planck Institute for Evolutionary Anthropology aims at automatically evaluating the phonetic similarity between all of the world’s languages. The basic idea is to compare pairs of words having the same meaning in two different languages according to their pronunciation. The average similarity across a specific set of words is then taken as a measure for the linguistic distance between the languages (Bakker et al., 2009). This distance can be interpreted as an approximation of the number of cognates between languages. The linguistic term cognates denotes common ancestries of words. A higher number of cognates indicates closer common ancestries. Although restricting its computation on differences in pronunciation, a lower Levenshtein distance therefore also indicates a higher probability of sharing other language characteristics such as grammar (see Serva, 2011). The acquisition of a second language is crucially affected by such differences in pronunciation and phonetic inventories, as they determine the difficulty in discriminating between different words and sounds. For a recent overview of the linguistic literature on language background and language acquisition see Llach (2010). The algorithm calculating the distance between words relies on a specific phonetic alphabet, the ASJPcode. The ASJPcode uses the characters within the standard ASCII1 alphabet to represent common sounds of human communication. The ASJPcode consists of 41 different symbols representing 7 vowels and 34 consonants. Words are then analyzed as to how many sounds have to be substituted, added, or removed to transfer the one word in one language into the same word in a different language (Holman et al., 2011). The words used in this approach are taken from the so-called 40-item Swadesh list, a list including 40 words that are common in almost all the world’s languages, including parts of the human body or expressions for common things of the environment. The Swadesh list is deductively derived by Swadesh (1952), its items are believed to be universally and culture independently included in all world’s languages.2 The ASJP program judges each word pair across languages according their similarity in pronunciation. For example, to transfer the phonetic transcription of the English word you, transcribed as yu, into the transcription of the respective German word du, one simply has to substitute the first consonant. But to transfer maunt3n, which is the transcription of mountain, into bErk, which is the transcription of the German Berg, one has to remove or substitute each 7 consonants and vowels, respectively.

1American Standard Code for Information Interchange, keyboard-character-encoding scheme. 2A list of the 40 words is available upon request. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 11

The following formalization of the computation follows Petroni and Serva (2010). To normalize the distance according to the word length, the resulting number of changes is divided by the word length of the longer word. Denoting this normalized distance between

item i of language α and β as Di(α, β), the calculation of the normalized linguistic distance (LDN) is computed as the average across all i = 1, ..., M distances between synonyms of the same item:

1 X LDN(α, β) = D(αi, βi). (2.1) M i To additionally account for potential similarities in phonetic inventories which might lead to a similarity by chance, a global distance between languages is deﬁned as the average Levenshtein distance of words with diﬀerent meanings:

1 X Γ(α, β) = D(αi, βj). (2.2) M(M − 1) i6=j The ﬁnal measure of linguistic distance is then the normalized and divided Levenshtein distance (LDND), which is deﬁned as:

LDN(α, β) LDND(α, β) = . (2.3) Γ(α, β) The resulting measure expresses a percentage measure of similarity between languages, although, by construction, it might take on values higher than 100% in cases in which languages do not even possess those similarities which are expected to exist by chance. Table 2.1 lists the closest and furthest languages toward English, German, and Spanish. The measurement via the normalized and divided Levenshtein distance is in line with an intuitive guessing about language dissimilarities. Although there is clearly a strong positive correlation between the Levenshtein distance and the test-score-based approach by Chiswick and Miller (1999), the Levenshtein distance oﬀers a higher variability in its measurement and we believe it to be more exact.3 Some languages are found to be distant according to the Levenshtein distance, but have a comparably low distance using the test-score-based measure, indicating that the test-score-based measure might also entail incentives to learn a foreign language instead of solely measuring linguistic distance.

2.3 Language Fluency of Immigrants

Language skills of immigrants are known to be a crucial determinant of the economic success of immigrants in the host country labor market. The economic literature concerning

3A figure which shows the relationship between both measure is available upon request. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 12 the determinants of language fluency of immigrants was initiated by the influential work by Chiswick (1991). Based on this seminal paper, Chiswick and Miller (1995) developed a theoretical human capital framework of host country language skill acquisition. In this framework, linguistic distance is a crucial determinant of language skills by lowering the efficiency of learning a language and inducing higher learning costs. This theoretical implication has been subsequently tested for various countries using the test-scores-based measure by Chiswick and Miller (1999). Due to its exclusive availability to the English language, these applications have been restricted to studies concerning the immigration to English-speaking countries such as the U.S. or Canada (Chiswick and Miller, 2005). This restriction does not hold for the Levenshtein distance as a measure of linguistic distance, which is not restricted to any home or host country, and may therefore be applied to a broader range of countries. This feature allows for providing evidence on the relationship between linguistic distance and language fluency in an international perspective. In doing so, we utilize data from three different sources. First, we use data from the 2000 U.S. Census to apply both the test-score-based measure by Chiswick and Miller (1999) and the Levenshtein distance to the same dataset. To compare the influence of linguistic distance across different countries, we additionally use data from the German Socio-Economic Panel (SOEP), and the National Immigrant Survey of Spain (NISS). The U.S., Germany, and Spain have very different migration histories that make an international comparison worthwhile. The United States have been an immigrant country since its foundation and currently a legal permanent residence status is granted to about 1 million immigrants per year. In 2000, this immigration flow consisted mainly of immigrants from other North-American countries (40%, including 21% from Mexico), followed by Asian (32%) and European immigrants (15%) (U.S. Department of Homeland Security, 2010). These inflows are also resembled in the stocks of the immigrant population. In the 2000 U.S. Census, 11% of the population of the U.S. were foreign-born. Neither does Germany have such a long-running immigration history as the U.S., nor can it look back on an extensive colonial history as Spain. Mass immigration started off only shortly after World War II with large waves of ethnic German expellees, followed by the so-called “guestworker”-programs aimed at attracting mainly unskilled workers from Mediterranean countries such as Turkey, Yugoslavia, Italy, or Spain. These two first waves of immigration were followed by a strong immigration phase by family re-unification during the 1970s and 1980s. The third large wave of immigration consists of immigrants and Ethnic Germans from former Soviet states during the 1990s (Bauer et al., 2005). Compared to the U.S. and Spain, Germany has a very old immigrant population with long individual migration histories. In 2009, 10.6 million (approx. 13%) of the German population have CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 13 immigrated after 1949, 3.3 million as Ethnic Germans. This third immigration wave was accompanied by large numbers of refugees and asylum seekers from Ex-Yugoslawia. The major part of the immigrant population is born in EU member states (32%), followed by 28% from Turkey and 27% from former members of the Soviet Union. Although Spain has a long-running colonial history, it is a comparably young immigration country. After large waves of emigration until the 1970s, net immigration began in the early nineties, and accelerated considerably during the last 20 years. Between 1997 and 2007, the number of migrants increased by around 700%, initially including mostly migrants from Africa and Western Europe. Nowadays, the majority of immigrants comes from Latin America and, since the EU enlargement, increasingly from Eastern Europe. Today, about 10% or 4.5 million of the population in Spain are foreign-born (see Fernández and Ortega, 2008).

2.3.1 Data and Method

Our data are restricted to male immigrants who entered the respective country after the age of 16 and are younger than 65 and who do not speak the host country language as their first language. The sample drawn from the 1%-PUMS (Public Use Microdata Series) 2000 U.S. Census file consists of 59,889 individuals. Similar data is extracted from the German Socio-Economic Panel, a long-run longitudinal representative study. Using cross-sectional data from 2001, the sample consists of 675 male immigrants.4 The National Immigrant Survey of Spain, conducted in 2007, also offers comprehensive cross-sectional information on the socio-economic characteristics and migration history of immigrants.5 The sample includes 2,513 male immigrants. All datasets include self-reported assessments of language fluency taking four or five possible values, which we have converted into a dichotomous measure taking a value of 1 if language skills are “Good” or “Very Good” and 0 otherwise. This variable serves as a the dependent variable in Probit regressions. This dichotomization decreases the probability of misclassification, which would lead to biased estimates in the case of Probit models, as pointed out by Dustmann and van Soest (2001). Moreover, it avoids dealing with violated proportional odds assumptions in the case of Ordered Probit models, as discussed by Isphording and Otten (2011). Further, the recoding enhances the comparability of the estimations between the different datasets and to previous approaches, as e.g., Chiswick and Miller (1999). Denoting this dichotomized indicator variable of host country language skills as our

4For further information about the SOEP see Haisken-DeNew and Frick (2005). The SOEP data was extracted by using the Stata-add-on PanelWhiz (Haisken-DeNew and Hahn, 2006). 5For further information about the NISS see Reher and Requena (2009). CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 14 dependent variable yi, the estimated probability of reporting good language skills can be speciﬁed as:

P r [yi = 1| LDi,Xi] = Φ (β0 + β1 LDi + Xi γ), (2.4) where Φ(·) is the cumulative distribution function for the standard normal distribution. LD is the linguistic distance between the native language and the host country language, the parameter β1 is our main parameter of interest, the disadvantage by linguistic origin in the language acquisition process. The main variable of interest is the measure of linguistic distance introduced in Section 2.2.2. Both measures of linguistic distance, the test-score-based measure and the normalized and divided Levenshtein distance, are expressed as percentiles of their respective distribution. This allows for a direct comparison between effects. As additional control variables, all three datasets offer comparable information on the age at migration, years since migration, years of education, marital status, number of children and an indicator variable denoting a former colonial relationship between home- and host country. We additionally include the distance between capitals in kilometers to proxy migration costs. For the 2000 U.S. Census we include some additional regional information about living in a non-metropolitan area, living in the Southern states and the share of the minority speaking the language of the individual. For Germany, we include a dummy for coming from a neighboring country. We control for refugee status (U.S. and Germany) and political reasons for migration (Spain), respectively. The U.S. data further includes information about having been abroad 5 years ago, while the German data includes information on having family abroad. This information serves as a proxy for return migration probability. Finally, each specification includes 17 world-region dummies to account for potential cultural differences correlated with linguistic distance. Sample means of the included variables are reported in Table 2.2. They show significant differences across the datasets, related to the different migration histories summarized above. Immigrants in Germany display the highest number of years since migration, as the sample consists in large parts of former guestworkers who immigrated during the 1960s and early 1970s. The German immigrant population also has the lowest mean education, but a higher share of married couples and a higher number of children, which might partly be due to the higher average age. The low average distance to the home country indicates the high share of guestworkers and immigrants from Eastern and Southern Europe. In contrast, both immigrants to Spain and the U.S. have a high average distance to the home country, as many immigrants come from overseas. Spain has the youngest immigrant population, resembling its relatively short immigration history starting off in the 1990s. Each dataset has a comparable share of “Good” or “Very Good” host country language CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 15 skills of around half of the sample.

2.3.2 Results

Table 2.3 lists the results of the Probit regressions across datasets, reported as marginal effects evaluated at the mean of the covariates. Columns (1) and (2) show the results for the U.S. data, using the test-score-based measure and the Levenshtein distance, respectively. Column (3) shows the results for the German SOEP data, and column (4) for the Spanish NISS data. The results confirm a significantly negative effect of linguistic distance on the probability of reporting good or very good language abilities in the host country language throughout all estimations. For the U.S., the effects for the test-score-based measure and the Levenshtein distance are qualitatively comparable. The effect is lower, however, when applying the Levenshtein distance. To illustrate the effect of linguistic distance, we can look at the additional amount of years of residence that make up for an initial disadvantage by linguistic origin. This amount of years of residence can be calculated by equating the marginal effect of years since migration with the marginal effect of a certain difference in linguistic origins and solving for the years since migration. In the U.S., the initial disadvantage of an immigrant with a distant linguistic origin, e.g., a Vietnamese who is in the 97th percentile of the distribution of linguistic distance, compared to an immigrant with close linguistic origin, e.g., a German who is in the 1st percentile, is worth around 6 years of additional residence. For a Turk (79th percentile), the largest immigrant group in Germany, the disadvantage compared to a linguistically closer Dutch migrant (3rd percentile) is worth 8 years of residence. Switching the measure of linguistic distance in the U.S. data from the test-score-based to the normalized and divided Levenshtein distance does not qualitatively affect the coefficients of the control variables. The coefficients are in line with previous studies and theoretical predictions. We see a positive impact of education, at around 5 percentage points for the U.S. and Germany and around 3 percentage points for Spain. The initial negative effect of age at migration decreases over time. Being married and having children is associated with higher language skills. The signs of these relationships are stable across all datasets, with the exception of lower language skills for immigrants with children in Spain. Being born in a former colony has a strong positive effect for both immigrants in the U.S. and in Spain. In Germany those immigrants from a neighboring country report higher language skills. Refugees in the U.S. and in Germany report lower average language skills compared to immigrants without refugee status. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 16 2.4 International Trade

Costs imposed by linguistic barriers can also be found on the macro-level. The trade- increasing effect of a common language is an undisputed fact in international economics. It is intuitive that trade between countries with a common language is cheaper than between countries with different languages. In their survey article, Anderson and van Wincoop (2004) report an estimate of the tax equivalent of “representative” trade costs for industrialized countries of about 170%. Of these, language-related barriers account for 7 percentage points, which is similar in magnitude to policy barriers and information costs. The question is whether and to what extent the dissimilarity between two languages matters if trading partners do not share a common language. Certainly, a range of dominating languages (English in the Western countries, Russian in Eastern Europe, French in Africa, and Spanish in Latin America) plays a major role in international trade. Especially the role of English as a lingua franca has been addressed by Ku and Zussman (2010). However, in the development of longer-term business partnerships, the crucial variable of interest is the linguistic knowledge in the trade partner’s home country language (Hagen et al., 2006), captured by the direct linguistic distance between the dominant languages of the trade partners. The method of choice in examining determinants of international bilateral trade is the gravity model first proposed by Tinbergen (1962). The basic theoretical gravity model assumes that the size of bilateral trade between any two countries depends on a function of each country’s economic size measured by (log of) GDP. Trade costs in their simplest form are approximated by the distance between the trading countries (Anderson and van Wincoop, 2004). Extensions are proxies for trade frictions, such as the effect of trade agreements (McCallum, 1995), and cultural proximity (Felbermayr and Toubal, 2010). To incorporate language-related barriers into these gravity models, common empirical practice is to use an indicator variable that equals 1 if two countries share the same official language and 0 otherwise (see Anderson and van Wincoop, 2004). While most studies employ the former approach, Mélitz (2008) goes beyond official languages and develops two different measures. The first measure depends on the probability that two randomly chosen individuals from either country share a common language spoken by at least 4% of both populations. The second measure is an indicator variable that equals 1 if two countries have the same official language or the same language is spoken by at least 20% of the populations of both countries. These measures share the shortcoming that they only look at whether countries share the same language, but do not account for heterogeneity in the degrees of similarity between languages. The degree of similarity, however, is likely to affect trade costs, e.g., CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 17

by lower costs of learning the trade partner’s language or by lowering translation costs (Hagen et al., 2006). Our results of Section 2.3.2, showing that linguistic barriers crucially affects second language acquisition, lend further support to this hypothesis. Moreover, lower host country language skills diminish the ability of immigrants to promote trade and commerce between their host country and their country of origin (Hutchinson, 2005). The only two approaches we know of that take similarities and differences between a multitude of languages into account are the ones by Hutchinson (2005) and Lohmann (2011). By relying on the measure by Chiswick and Miller (1999), Hutchinson’s approach is restricted to distances toward English. Lohmann (2011) uses data from the World Atlas of Language Structures (WALS; see Dryer and Haspelmath, 2011) to construct an index of 139 potentially shared linguistic features between languages. Similar to our application, he applies this index to explain international trade flows using data from Rose (2004). This approach counts shared language features within language pairs and builds up a language features index normalized to the interval of [0; 1], where 0 means sharing all features.6

2.4.1 Data and Method

To ensure a high degree of comparability with the previous literature, we use a widely accepted empirical methodology and a standard dataset of bilateral trade flows. The dataset constructed by Rose (2004) has been widely used previously by Mélitz (2008), Ku and Zussman (2010), and Lohmann (2011).7 The sample covers bilateral trade between 178 countries over the years 1948 to 1999 leading to 234,597 country-pair-year observations.8 The variables of interest are Rose’s binary common language variable, two versions of linguistic distance between trading partners’ languages as measured by the Levenshtein distance, and finally the linguistic features index calculated by Lohmann (2011). The Levenshtein distance is computed for every country-pair in the dataset. In monolingual countries we assign the respective native language to the country. In multi-lingual countries, the most prevalent native language is assigned, which was identified using a multitude of sources, including CIA’s World Factbook, encyclopedias, and Internet

6The measure by Lohmann (2011) is assigned at the country-level using the most widely spoken official language of each country. In the Spanish and U.S. micro-data, we can rely on a more detailed assignment using information on the mother tongue of each individual. This makes it unfeasible to include this alternative measure in the micro-data regressions in Section 2.3. 7The data and their sources are explained in detail in Rose (2004) and posted on his website. Following Tomz et al. (2007), we have defined our WTO membership variable broadly to include both countries that are either formal members of the organization or have agreements that involve rights and obligations toward it. 8The annual value of bilateral trade between a pair of countries is created by averaging the imports and exports. Country-pairs with zero bilateral trade flows are not included in the sample. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 18 resources.9 To analyze the sensitivity of the results with respect to the measurement of linguistic distance, we calculate an alternative specification of the Levenshtein distance, replacing the most prevalent language with the prevailing lingua franca in a country. A lingua franca is defined as a language typically used to enable communication between individuals not sharing a mother tongue. These languages are often third languages, which are widely spoken in a particular regional area and are not necessarily an official language.10 Subsequently, we compare the effect of these two definitions of the Levenshtein distance with the approach by Lohmann (2011). Descriptive statistics of the variables used in the empirical analysis are shown in Table 2.4. The average Levenshtein distance decreases from 90.3 to 75.1 when we use lingua francas instead of the most prevalent native language to calculate the linguistic distance. This indicates that lingua francas may have come into existence to decrease costs imposed by language barriers in the first place. Following Rose’ definition, 22.3% of the country-pairs share a common language. This quite high share relies on a very broad definition of official languages by Rose. For example, even country pairs such as the U.S. and Denmark or France and Egypt are coded to have the same language. Using the Levenshtein distance, only 4.7% of the country-pairs show a distance of zero, which is equivalent to sharing a common language, increasing to 18.4% for the Levenshtein distance measure based on lingua francas. The linguistic features index by Lohmann (2011) is zero for 9.4% of the country-pairs, meaning that both languages share all linguistic features considered. We use the gravity model to estimate the impact of language barriers on trade between pairs of countries. The model has a long record of success in explaining bilateral trade flows and becomes the standard model for applied trade analysis. Following Rose (2004), we augment the basic gravity equation with a number of additional variables that affect trade in order to control for as many determinants of trade flows as possible. Our empirical strategy is to compare trade patterns for trading partners with different language barriers using variation across country-pairs. If a common language or a high similarity between languages has a positive effect on trade, we expect to observe significantly higher trade for these country-pairs than for others. We compare three different specifications. First, we adopt the original specification by Rose (2004) including an indicator variable for country-pairs sharing the same language. This basic approach is then augmented by the Levenshtein distance and the language features index by Lohmann (2011). The exact

9For example, we use English as the native language in the United Kingdom, because it is a monolingual country and English is the national language. In a multi-lingual country such as Canada we use English instead of French, because English is the most prevalent native language. A comprehensive index of assigned languages with further explanations is available upon request. 10For example, we use Russian as lingua franca for most countries of the former Soviet Union. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 19 speciﬁcation of the gravity model is:

ln Xijt = β0 + β1 ln (YiYj)t + β2 ln Distij + Zijt κ + γ1 LBij (2.5) + X δ I + X θ J + X φ T + ε , i i i j j j t t t ijt where the dependent variable Xijt denotes the average value of real bilateral trade between country i and country j at time t, mainly inﬂuenced by the “mass” of both economies, indicated by the product of their GDP denoted by Y , and the distance in log miles. Z is a vector of control variables, including population size, geographic characteristics such as sharing a land border, number of landlocked countries, number of island nations in the country-pair (0, 1, or 2), the area of the country (in square kilometers), and colonial relationships. Further, it is controlled for member and nonmember participation in the GATT/WTO (one or both countries), same currency, regional trade agreements, and being a GSP beneﬁciary.11

The main coefficient of interest is γ1. It measures the effect of the different language barriers variables (LB) on international trade. If both countries share a common language,

γ1 should be positive; if instead one of the linguistic distance measures is used, the effect of γ1 on trade should be negative. A comprehensive set of country and year fixed effects is included in the specification to control for any factor affecting trade that is country (e.g., stock of migrants, foreign language knowledge) or time specific (e.g., common shocks and trends).12 The gravity model is estimated by ordinary least squares (OLS) with robust standard errors clustered on the country-pair level.

2.4.2 Results

Table 2.5 summarizes the results of Eq. (2.5). For the sake of brevity, the estimated coefficients for the time- and country-fixed effects are omitted from all tables. In the first column, we reproduce the benchmark specification from Rose (2004) based on his measure of common language augmented with country fixed effects. Rose’ model confirms the hypothesis of a significant positive effect of common language on bilateral trade. Sharing a common language is found to raise trade by about (exp(0.274) − 1 ≈) 31.5%.

11More details are given in Rose (2004), the source for all variables except the linguistic distance measures. In the assignment of GATT/WTO rights and obligation we follow Tomz et al. (2007) and impose the restriction that formal membership has the same effect as nonmember participation. 12Recent empirical work on the determinants of bilateral trade increasingly relies on panel data techniques that account for country-pair instead of exporter and importer specific fixed effects. Country- pair fixed effects control for the impact of any time-invariant country-pair specific determinant such as bilateral distance or common language. However, this comes at the cost of not being able to estimate the effect of the language barrier variables, our variables of interest, on bilateral trade. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 20

Still, this result might be biased by the very broad definition of having a common language. The question we want to answer is whether language barriers affect trade above and beyond the simple effect of sharing a common language. Therefore, our ensuing specifications examine how the results change when we employ the linguistic distance measures instead of the common language variable. The second column shows our preferred model. We replace Rose’ common language variable with our default Levenshtein distance measure. We find significantly lower trade when the Levenshtein distance between both countries in a dyad increases. The coefficient indicates that a country-pair trades about (exp(−0.006)−1 ≈) 0.6% less if the Levenshtein distance increases by one unit. To illustrate the magnitude of this effect, we note that the 75th percentile of the Levenshtein distance in our sample is 99.93 (roughly the distance between English and Japanese) and the 25th percentile is 92.95 (roughly the distance between English and Russian). The estimate in column 2 implies that an increase from the 25th to the 75th percentile in the Levenshtein distance decreases bilateral trade by approximately 4.1%.13 In multi-lingual countries, the assignment of languages to countries is difficult. To show that our findings are not a result of a particular assignment of languages to countries, the estimation results with the Levenshtein distance measure based on lingua francas are presented in column (3). The key result that the Levenshtein distance has a statistically and economically significant negative effect on bilateral trade is robust. However, the effect decreases by 50%, maybe due to the lower variability of the alternative Levenshtein distance. Additionally, lingua francas are purposely chosen to lower transaction costs. Therefore, we should expect a smaller effect on trade when taking the lingua francas into account. Next, column (4) shows the results of Lohmann’s linguistic features index as a measure of common language. Due to a restricted data availability, the linguistic features index is only computable for a subsample of 227,145 country-pairs.14 The coefficient reveals that a pair of countries trade about (exp(−0.618) − 1 ≈) 4.6% less if the linguistic features index increases by 0.1 units (corresponding to a 10% decrease in common linguistic features). To compare the influence of the language features index by Lohmann (2011) to the

13To examine whether the effect of the Levenshtein distance on bilateral trade only builds on the grounds of sharing or not sharing a common language and not on the linguistic distance between different languages, we estimate models 2-4 with subsamples excluding country-pairs with no language barrier in the corresponding measure, i.e., a linguistic distance of zero. The results are available upon request. Regarding both versions of the Levenshtein distance measure, they become even lager in magnitude and are stable in significance, while the linguistic features index becomes distinctly smaller in magnitude and significance. 14To check for sample selection we additionally estimated models 1-3 restricted to the same subsample. The results are available upon request. The estimates regarding the language variables remained stable in magnitude and significance. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 21

Levenshtein distance, we compute the elasticity and the marginal effect multiplied by the interquartile range of the linguistic features index. Increasing the linguistic features index by 1% decreases bilateral trade by about 0.3% compared to a 0.6% decrease in case of the Levenshtein distance. Moving up the distribution of the linguistic features index from the lower to the upper quartile decreases trade between countries by (exp(−1.465) − 1 ≈) 7.7%. The results show a larger effect for the Levenshtein distance with regard to elasticities. Since the distribution of the Levenshtein distance is right-skewed, the value of the interquartile range is smaller compared to the interquartile range of the linguistic features index. As a result, the effect of the linguistic features index becomes larger than the effect of the Levenshtein distance. In summary, the empirical analysis provides evidence that according to both measures linguistic distance has a statistically and economically significant negative effect on bilateral trade flows. The estimated coefficients of the control variables confirm the traditional results of gravity trade equations. The indicators for whether one or both countries in the dyad participated in the GATT/WTO have significantly positive coefficients. The respective coefficients are comparable to those reported by Tomz et al. (2007). Countries that are farther apart trade less, while countries belonging to the same regional trade association, belonging to the same GSP, or sharing a currency trade more. Islands or landlocked countries trade less, while countries sharing a land border trade more. Economically larger and richer countries trade more, as do physically larger countries. A shared colonial history encourages trade as well. These estimation results are both statistically and economically significant and in line with estimates from previous literature. As compared to the first specification, the application of the Levenshtein distance measure does not considerably affect the magnitude or significance of the other independent variables. All variables show the expected results. However, the coefficient of common colonizer increases by about 10 percentage points, indicating that the effect of cultural ties is underestimated in the traditional gravity model. During the colonization period, colonizers created new institutions such as the legal and administrative system in their colonies. These institutions impose policies and law enforcement, thereby determining the formal and informal rules in commerce. Since international transactions between countries with different or poorly developed institutional settings involve high transactions costs, colonial ties between countries that had the same colonial history and therefore established a similar institutional system, facilitate bilateral trade flows. Despite the fact that the colonizers’ languages became the official languages in the colonies and represent one of the official languages in most former colonies even today, a large part of the population failed to achieve an acceptable degree of knowledge in these languages (see, e.g., Lewis, 2009). Hence, using information on common official languages in a country-pair to estimate trade CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 22

flows, in particular between countries with a common colonizer, might underestimate the effect of colonial ties and overestimate the relation of the common colonizers language on trade pattern. In summary, a common colonizer promotes trade between these countries because of establishing a similar institutional setting; an effect that might be hidden when not controlling properly for linguistic heterogeneity.

2.5 Conclusion

This study is concerned with the operationalization of linguistic distance between languages and the estimation of arising costs of linguistic barriers on the micro- and macro-level. Linguistic barriers are strong obstacles in the realization of free worldwide factor movements. The operationalization of linguistic barriers in applied economic studies is not straightforward and makes it necessary to rely on interdisciplinary approaches drawing heavily from linguistic research. Our measure for linguistic distance is based on the Automatic Similarity Judgment Program (ASJP) by the German Max Planck Institute for Evolutionary Anthropology. The linguistic distance is computed as a function of phonetic similarity of words (a Levenshtein distance) from different languages having the same meaning. It can be used as an approximation of the historical difference in languages and is therefore also correlated to differences in other dimensions of dissimilarity, such as grammar or vocabularies. Compared to the previous approach by Chiswick and Miller (1999), which measures linguistic distance by using average test-scores of second language students, the Levenshtein distance has some advantages. It is available for any pair of the world’s languages (instead of being only applicable for the distance toward English). Additionally, it is not influenced by other extrinsic or intrinsic incentives for learning a foreign language, and should deliver an unbiased approximation of the dissimilarity between languages. The measurement of linguistic distance is used in two applications, the language acquisition of immigrants and language barriers in bilateral trade flows. Following the widely accepted rational choice framework of language acquisition (see, e.g., Chiswick and Miller, 1995; Esser, 2006), linguistic distance affects second language skills by lowering the initial efficiency, thereby imposing higher costs of learning a foreign language. Following previous work that shows such a negative relationship for English-speaking countries, we broadened the evidence for other countries by applying the measure in estimations using U.S., German, and Spanish individual micro-data. The results confirm a strong significantly negative effect of linguistic distance on immigrant language skills. The initial disadvantage due to distant linguistic origin is worth several years of additional residence. As such, the linguistic distance is able to explain a large part of language skill heterogeneity CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 23

in immigrant populations. The considerable hurdles for language acquisition on the micro- level might explain the lower migration rates between linguistically distant countries, as analyzed by Adsera and Pytlikova (2012). To additionally look at how these effects on the micro-level accumulate to costs of linguistic barriers on the macro-level, we apply the Levenshtein distance in the setting of international trade. Linguistic proximity is believed to enhance trade flows between countries by lowering costs imposed by language barriers, e.g., translation or information costs. Using a comprehensive dataset of bilateral trade flows by Rose (2004), we estimate a standard gravity model using the Levenshtein distance as an additional explanatory variable and compare this approach to a previous approach based on shared linguistic features by Lohmann (2011). The results provide new and strong evidence indicating that language barriers affect trade above and beyond the simple effect of sharing a common language. Moving up the distribution of the Levenshtein distance from the lower quartile (roughly the distance between English and Russian) to the upper quartile (roughly the distance between English and Japanese) decreases trade between countries by about 4.1%. Taken together, this study suggests an important role of language differences in economic transactions. The results show the significant economic costs of linguistic heterogeneity on the individual and aggregated level. The Levenshtein distance offers a simple and comprehensive way to control for this heterogeneity in a large range of applications in empirical economics and thereby circumvents potential pitfalls by decreasing the degree of unobserved heterogeneity in the data. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 24 Tables

Table 2.1: Closest and Furthest Language Pairs with Respect to the Levenshtein Distance Closest Furthest Language Distance Language Distance Distance to English Afrikaans 62.08 Vietnamese 104.06 Dutch 63.22 Turkmen 103.84 Norwegian 64.12 Hakka (China) 103.10

Distance to German Luxembourgish 42.12 Korean 104.30 Dutch 51.50 Palestinian Arabic 103.72 Westvlaams (Belgium) 57.86 Yoruba (Nigeria) 103.58

Distance to Spanish Galician 54.82 Wolof (Senegal) 103.02 Italian 56.51 Igbo Onitsha (Nigeria) 102.84 Portuguese 64.21 Ewondo (Cameroon) 101.87 Notes: – The table shows the three closest and furthest languages toward English, German and Spanish according to the normalized and divided Levenshtein distance. – Only languages spoken within samples are listed. – Geographic origin of language in parentheses.

Table 2.2: Descriptive Statistics of Dependent and Explanatory Variables – Immigration Sample 2000 U.S. Census SOEP NISS Mean StdD Mean StdD Mean StdD Good language skills 0.58 0.49 0.52 0.50 0.58 0.49 Years of education 11.32 4.28 10.50 2.21 10.82 3.38 Age at entry 26.76 8.72 28.68 8.93 30.00 9.63 Years since migration 12.72 9.91 18.50 11.11 8.33 6.96 Married 0.68 0.47 0.86 0.35 0.58 0.49 One child 0.19 0.39 0.50 0.50 0.24 0.43 Two children 0.19 0.39 0.21 0.41 0.22 0.42 Three or more children 0.14 0.35 0.18 0.38 0.34 0.47 Distance to home country (in 100 km) 57.60 39.95 19.14 14.62 24.41 22.96 Naturalized 0.34 0.47 0.35 0.48 0.07 0.25 Former colony 0.11 0.32 0.10 0.30 0.03 0.18 Southern states 0.29 0.45 –––– Non-metropolitan area 0.01 0.12 –––– Minority language share 0.33 0.25 –––– Abroad ﬁve years ago 0.23 0.42 –––– Refugee 0.12 0.32 0.07 0.25 –– Neighboring country – – 0.12 0.33 –– Family abroad – – 0.30 0.46 –– Political reasons – – – – 0.03 0.16 Notes: – Number of observations: 59,889 in the 2000 U.S. Census, 675 in the SOEP, and 2,513 in the NISS Sample. – The dependent variable “Good language skills” is deﬁned dichotomously, 1 indicates higher language skills. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 25

Table 2.3: Immigrant’s Language Skills – Probit Results Dataset: 2000 U.S. Census SOEP NISS Linguistic distance measure: Test-score LDND LDND LDND ME/StdE ME/StdE ME/StdE ME/StdE Linguistic distance (Test-score-based) −0.003∗∗∗ ––– (0.000) Levenshtein distance (ASJP) – −0.001∗∗∗ −0.002∗ −0.002∗∗ (0.000) (0.001) (0.001) Years of education 0.048∗∗∗ 0.048∗∗∗ 0.054∗∗∗ 0.029∗∗∗ (0.001) (0.001) (0.011) (0.003) Age at entry −0.018∗∗∗ −0.018∗∗∗ −0.039∗∗ 0.006 (0.002) (0.002) (0.015) (0.007) Age at entry2/100 0.012∗∗∗ 0.012∗∗∗ 0.049∗ −0.019∗ (0.002) (0.002) (0.022) (0.010) Years since migration 0.014∗∗∗ 0.014∗∗∗ 0.032∗∗ 0.039∗∗∗ (0.001) (0.001) (0.011) (0.005) Years since migration2/100 −0.021∗∗∗ −0.021∗∗∗ −0.058∗ −0.067∗∗∗ (0.002) (0.002) (0.024) (0.015) Married 0.020∗∗∗ 0.020∗∗∗ −0.008 0.003 (0.005) (0.005) (0.066) (0.026) Children in the HH. (Ref.= 0) One child 0.018∗∗ 0.019∗∗ 0.013 −0.055† (0.006) (0.006) (0.083) (0.030) Two children 0.015∗ 0.015∗ −0.038 0.072† (0.006) (0.006) (0.083) (0.038) Three or more children −0.001 −0.000 −0.094 −0.128∗∗ (0.007) (0.007) (0.087) (0.041) Distance to home country (in 100 km) −0.002† −0.002† 0.006 0.001 (0.001) (0.001) (0.008) (0.004) Distance to home country2/100 0.003∗∗∗ 0.003∗∗∗ −0.002 0.001 (0.001) (0.001) (0.010) (0.002) Naturalized 0.138∗∗∗ 0.138∗∗∗ 0.300∗∗∗ 0.043 (0.006) (0.006) (0.065) (0.049) Former colony 0.108∗∗∗ 0.102∗∗∗ −0.222 0.227∗∗∗ (0.011) (0.011) (0.218) (0.059) Southern states 0.044∗∗∗ 0.045∗∗∗ –– (0.005) (0.005) Non-metropolitan area 0.051∗∗ 0.049∗∗ –– (0.018) (0.018) Minority language share −0.252∗∗∗ −0.280∗∗∗ –– (0.019) (0.020) Abroad five years ago −0.093∗∗∗ −0.091∗∗∗ –– (0.007) (0.007) Refugee −0.233∗∗∗ −0.214∗∗∗ −0.085 – (0.009) (0.009) (0.103) Neighboring country – – 0.310† – (0.166) Family abroad – – −0.069 – (0.056) Political reasons – – – 0.045 (0.063) Region fixed effects yes yes yes yes Pseudo-R2 0.263 0.261 0.160 0.138 Observations 59,889 59,889 675 2,513 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors are reported in parentheses. – The dependent variable is defined dichotomously, 1 indicates higher language skills. – Probit results are reported as marginal effects evaluated at covariate means. – Region controls are not recorded. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 26 average value of realbinary bilateral variable trade which between is countriesLevenshtein unity i distance if and using i j the andLevenshtein at most distance j year prevalent using share native t the a languageindex in most common of which US prevailing language each increases $ lingua and country with francabinary zero decreasing of variable otherwise similarity which each of is country binary a unity variable language-pair if which (values both is betweenbinary i unity 0 variable and if and which j either 1) is are igreat unity GATT/WTO or circle if participants j distance i at is between wasproduct t country a a of i GATT/WTO GSP the participant and beneficiary real atproduct country of GDP’s t j of j of in the or both miles realbinary vice countries GDP’s variable versa in per which at year is capita t t binary unity of variable if both which i countries is andbinary in unity j variable year if which both t i is belong andnumber unity to of j if the landlocked use i same nations the andnumber regional in same of j trade the currency island share agreement country-pair at nations aproduct (0, time in land 1, of t the border or the country-pair 2) landbinary (0, areas variable 1, of which or both is 2) binary unity countries variable if (in which i square is andbinary kilometers) unity j variable if which were i is ever isbinary unity colonies a variable if after which colony i 1945 is of ever with unity j colonized the if at j same i time or colonizer and t vice j or versa remained vice part versa of the same nation during the sample 336 416 453 787 203 476 461 422 809 676 504 120 118 172 466 540 280 300 044 142 017 ...... 062223256 3 063 0 429 22 652 36 307 0 231 0 165 0 881 0 034 0 015 2 014 1 031 0 246 0 341 0 206 0 100 0 002 3 021 0 000 0 0 0 ...... 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 Mean StdD Definitions 10 90 75 47 16 24 Descriptive Sample Statistics and Variable Definitions – International Trade Sample Common language Levenshtein distance Levenshtein distance LF Linguistic features index Both in GATT/WTO One in GATT/WTO General system of preferences Log distance Log product real GDP Log product real GDPRegional p/c FTA Currency union Land border Number landlocked Number islands Log product land area Common colonizer Currently colonized Ever colony Common country Log real trade Notes: – Number of observations: 234,597 in 12,150 country-pair groups with 1 to 52 observations per group. The mean is 19.3 observations per group. – For the linguistic features index there are 227,145 observations in 11,348 country-pair groups. Table 2.4: CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 27

Table 2.5: Effect of Language on Bilateral Trade – OLS Results ComLang LDND I LDND II LingFeat Coef/StdE Coef/StdE Coef/StdE Coef/StdE Common language 0.274∗∗∗ ––– (0.044) Levenshtein distance – −0.006∗∗∗ –– (0.001) Levenshtein distance LF – – −0.003∗∗∗ – (0.000) Linguistic features index – – – −0.618∗∗∗ (0.098) Both in GATT/WTO 0.604∗∗∗ 0.618∗∗∗ 0.609∗∗∗ 0.578∗∗∗ (0.061) (0.061) (0.061) (0.062) One in GATT/WTO 0.277∗∗∗ 0.288∗∗∗ 0.283∗∗∗ 0.247∗∗∗ (0.056) (0.056) (0.056) (0.056) General system of preferences 0.709∗∗∗ 0.733∗∗∗ 0.711∗∗∗ 0.721∗∗∗ (0.032) (0.031) (0.032) (0.032) Log distance −1.313∗∗∗ −1.278∗∗∗ −1.308∗∗∗ −1.293∗∗∗ (0.023) (0.024) (0.023) (0.024) Log product real GDP 0.167∗∗ 0.165∗∗ 0.164∗∗ 0.159∗∗ (0.051) (0.051) (0.051) (0.053) Log product real GDP p/c 0.532∗∗∗ 0.533∗∗∗ 0.535∗∗∗ 0.552∗∗∗ (0.049) (0.049) (0.049) (0.050) Regional FTA 0.941∗∗∗ 0.942∗∗∗ 0.939∗∗∗ 0.975∗∗∗ (0.126) (0.126) (0.126) (0.129) Currency union 1.174∗∗∗ 1.253∗∗∗ 1.169∗∗∗ 1.208∗∗∗ (0.122) (0.125) (0.123) (0.124) Land border 0.280∗∗ 0.283∗∗ 0.284∗∗ 0.292∗∗ (0.108) (0.108) (0.108) (0.113) Number landlocked −1.056∗∗∗ −1.032∗∗∗ −1.014∗∗∗ −0.971∗∗∗ (0.207) (0.205) (0.205) (0.208) Number islands −1.579∗∗∗ −1.575∗∗∗ −1.622∗∗∗ −1.545∗∗∗ (0.188) (0.188) (0.188) (0.190) Log product land area 0.496∗∗∗ 0.501∗∗∗ 0.513∗∗∗ 0.496∗∗∗ (0.041) (0.041) (0.041) (0.041) Common colonizer 0.605∗∗∗ 0.703∗∗∗ 0.592∗∗∗ 0.687∗∗∗ (0.064) (0.062) (0.065) (0.065) Currently colonized 0.743∗∗ 0.744∗∗ 0.753∗∗ 0.719∗∗ (0.263) (0.252) (0.264) (0.262) Ever colony 1.274∗∗∗ 1.272∗∗∗ 1.261∗∗∗ 1.339∗∗∗ (0.114) (0.116) (0.114) (0.113) Common country 0.288 0.090 0.263 0.278 (0.583) (0.658) (0.579) (0.617) Year fixed effects yes yes yes yes Country fixed effects yes yes yes yes Adjusted R2 0.703 0.703 0.703 0.705 RMSE 1.818 1.817 1.817 1.805 F Statistic 274.34∗∗∗ 272.11∗∗∗ 274.96∗∗∗ 265.50∗∗∗ Observations 234,597 234,597 234,597 227,145 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors (clustered at the country-pair lvel) are reported in parentheses. – The dependent variable is defined as log of real bilateral trade in US$. – Intercept, year, and country controls are not recorded. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 28 2.A Appendix

Table 2.A1: 40-Items Swadesh Word List I You We One Two Person Fish Dog Louse Tree Leaf Skin Blood Bone Horn Ear Eye Nose Tooth Tongue Knee Hand Breast Liver Drink See Hear Die Come Sun Star Water Stone Fire Path Mountain Night Full New Name

Source: Bakker et al. (2009). 100 90 80 Levenshtein Distance 70 60 .2 .4 .6 .8 1 Linguistic Distance (Test-Score-Based)

Figure 2.A1: Comparisons of Linguistic Distance Using the Test-Score-Based Measure and the Levenshtein Distance – 2000 U.S. Census CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 29

Table 2.A2: Summary Statistics for the Language Variables – International Trade Sample A. Simple Correlations among Language Distance Measures Common Levenshtein Levenshtein Linguistic features language distance distance LF indexa Common language 1

Levenshtein distance -0.3868 1

Levenshtein distance LF -0.6689 0.4813 1

Linguistic features indexa -0.3533 0.5490 0.4070 1

B. Frequency of Country-pairs with and without the same Language Common Levenshtein Levenshtein Linguistic features language distance distance LF indexa Same language 52,205 11,017 43,229 21,389

Diﬀerent language 182,392 223,580 191,368 205,756 Notes: – Number of observations: 234,597, except a227,145. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 30

0.015

0.01 Density 0.005

0 100 80 20 15 60 10 5 40 0 20 −5 −10 0 −15 Levenshtein Distance Log Bilateral Trade

Figure 2.A2: Bivariate Kernel Density Estimation of Log Bilateral Trade and Levenshtein Distance – International Trade Sample CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 31

Sensitivity Analysis

As mentioned above, we perform a number of sensitivity analyses which, in each case, find similar results to those reported above. First, we estimate our model of immigrant’s language skills using an expanded sample which also includes individuals speaking the language of the host country as mother tongue. The results are reported in Table 2.A3. Second, we repeat our estimations using the four or fivefold information of language fluency as dependent variable in Ordered Probit models. The marginal effects across the different categories indicate a comparable effect as in the dichotomous case. The signs of the marginal effects change at the same threshold we use for the dichotomization. The results are available upon request. Third, Tables 2.A4 and 2.A5 report estimation results for two subsamples of the trade sample. The results are quite similar in magnitude and significance level to those for the whole sample. Table 2.A4 examines the sensitivity of the results with respect to the measurement of linguistic distance. Therefore, we exclude dyad-observations with the same language from the sample, thereby including only country-pairs with a language barrier greater than zero. This tests the idea that country-pairs speaking or not speaking a common language delivering the results of the language barrier, rather than an effect of linguistic distance per se. Table 2.A5 analyzes the sensitivity of our results when we restrict our sample to the slightly smaller one of Lohmann’s linguistic features index. Finally, we add for both countries in a country-pair country-by-time interaction terms, P P it ηit (I × T )it and jt ψjt (J × T )jt, to the models of Table 2.5. These interaction terms capture any exporter and importer specific time-variant effects such as each country’s business cycle or its institutional characteristics. The findings for the key variables (available upon request) are quite similar in magnitude and significance level to those for the models with country and year specific fixed effects. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 32

Table 2.A3: Immigrant’s Language Skills – Probit Results, Including Native Speakers ME/StdE ME/StdE ME/StdE ME/StdE Linguistic distance (Test-score-based) −0.008∗∗∗ ––– (0.000) Levenshtein distance (ASJP) – −0.007∗∗∗ −0.003∗ −0.004∗∗∗ (0.000) (0.001) (0.000) Years of education 0.040∗∗∗ 0.040∗∗∗ 0.054∗∗∗ 0.015∗∗∗ (0.001) (0.001) (0.011) (0.002) Age at entry −0.014∗∗∗ −0.015∗∗∗ −0.038∗ 0.002 (0.001) (0.001) (0.015) (0.003) Age at entry2/100 0.009∗∗∗ 0.010∗∗∗ 0.047∗ −0.008 (0.002) (0.002) (0.021) (0.005) Years since migration 0.013∗∗∗ 0.013∗∗∗ 0.029∗∗ 0.022∗∗∗ (0.001) (0.001) (0.011) (0.003) Years since migration2/100 −0.020∗∗∗ −0.022∗∗∗ −0.053∗ −0.038∗∗∗ (0.002) (0.002) (0.024) (0.008) Married 0.015∗∗ 0.018∗∗∗ 0.007 0.003 (0.005) (0.005) (0.066) (0.014) Children in the HH. (Ref.= 0) One child 0.012∗ 0.016∗∗ 0.008 −0.030† (0.005) (0.005) (0.082) (0.017) Two children 0.012∗ 0.014∗ −0.039 0.035∗ (0.005) (0.005) (0.082) (0.018) Three or more children −0.005 0.001 −0.092 −0.065∗∗ (0.006) (0.006) (0.086) (0.022) Distance to home country (in 100 km) −0.000 0.000 0.007 0.001 (0.001) (0.001) (0.008) (0.002) Distance to home country2/100 0.002∗∗∗ 0.002∗∗∗ −0.003 0.000 (0.000) (0.000) (0.010) (0.001) Naturalized 0.108∗∗∗ 0.106∗∗∗ 0.268∗∗∗ 0.017 (0.004) (0.005) (0.066) (0.024) Former colony 0.123∗∗∗ 0.095∗∗∗ −0.203 0.240∗∗∗ (0.008) (0.008) (0.202) (0.030) Southern states 0.049∗∗∗ 0.064∗∗∗ –– (0.004) (0.004) Non-metropolitan area 0.039∗∗ 0.021 –– (0.015) (0.015) Minority language share −0.340∗∗∗ −0.567∗∗∗ –– (0.016) (0.017) Abroad five years ago −0.084∗∗∗ −0.078∗∗∗ –– (0.006) (0.006) Refugee −0.266∗∗∗ −0.214∗∗∗ −0.096 – (0.009) (0.008) (0.102) Neighboring country – – 0.313∗ – (0.159) Family abroad – – −0.085 – (0.056) Political reasons – – – 0.032 (0.029) Region fixed effects yes yes yes yes Pseudo-R2 0.304 0.299 0.165 0.347 Observations 70,201 70,201 689 3,986 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors are reported in parentheses. – The dependent variable is defined dichotomously, 1 indicates higher language skills. – Probit results are reported as marginal effects evaluated at covariate means. – Region controls are not recorded. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 33

Table 2.A4: Effect of Language on Bilateral Trade – OLS Results, Subsample Language Barrier > 0 LDND I LDND II LingFeat Coef/StdE Coef/StdE Coef/StdE Levenshtein distance −0.008∗∗∗ –– (0.002) Levenshtein distance LF – −0.008∗∗∗ – (0.002) Linguistic features index – – −0.266∗ (0.123) Both in GATT/WTO 0.506∗∗∗ 0.435∗∗∗ 0.454∗∗∗ (0.067) (0.070) (0.069) One in GATT/WTO 0.219∗∗∗ 0.193∗∗ 0.174∗∗ (0.061) (0.064) (0.063) General system of preferences 0.754∗∗∗ 0.612∗∗∗ 0.682∗∗∗ (0.031) (0.032) (0.032) Log distance −1.281∗∗∗ −1.211∗∗∗ −1.271∗∗∗ (0.025) (0.027) (0.027) Log product real GDP 0.060 −0.070 −0.005 (0.053) (0.058) (0.056) Log product real GDP p/c 0.646∗∗∗ 0.793∗∗∗ 0.727∗∗∗ (0.051) (0.056) (0.054) Regional FTA 0.828∗∗∗ −0.252∗ 0.190 (0.147) (0.118) (0.155) Currency union 1.312∗∗∗ 1.138∗∗∗ 1.200∗∗∗ (0.133) (0.275) (0.203) Land border 0.281∗ 0.391∗∗ 0.301∗ (0.120) (0.130) (0.133) Number landlocked −1.232∗∗∗ −1.330∗∗∗ −1.307∗∗∗ (0.206) (0.207) (0.214) Number islands −1.837∗∗∗ −2.492∗∗∗ −1.993∗∗∗ (0.192) (0.199) (0.203) Log product land area 0.551∗∗∗ 0.668∗∗∗ 0.581∗∗∗ (0.042) (0.043) (0.044) Common colonizer 0.697∗∗∗ 0.887∗∗∗ 0.662∗∗∗ (0.063) (0.092) (0.069) Currently colonized 0.322 1.149∗∗ 0.442 (0.292) (0.391) (0.390) Ever colony 1.517∗∗∗ 1.041∗∗∗ 1.182∗∗∗ (0.131) (0.193) (0.152) Common country 1.203∗∗∗ –– (0.346) Year fixed effects yes yes yes Country fixed effects yes yes yes Adjusted R2 0.702 0.706 0.708 RMSE 1.827 1.792 1.801 F Statistic 860.28∗∗∗ 267.50∗∗∗ 275.73∗∗∗ Observations 223,580 191,368 205,756 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors (clustered at the country-pair level) are reported in parentheses. – The dependent variable is defined as log of real bilateral trade in US$. – Intercept, year, and country controls are not recorded. – In column (2) and (3) common country is omitted from the equations because of collinearity. CHAPTER 2. LINGUISTIC DISTANCE IN APPLIED ECONOMICS 34

Table 2.A5: Effect of Language on Bilateral Trade – OLS Results, Subsample Linguistic Features Index ComLang LDND I LDND II Coef/StdE Coef/StdE Coef/StdE Common language 0.292∗∗∗ –– (0.045) Levenshtein distance – −0.006∗∗∗ – (0.001) Levenshtein distance LF – – −0.004∗∗∗ (0.001) Both in GATT/WTO 0.585∗∗∗ 0.600∗∗∗ 0.592∗∗∗ (0.062) (0.062) (0.062) One in GATT/WTO 0.255∗∗∗ 0.267∗∗∗ 0.264∗∗∗ (0.057) (0.056) (0.057) General system of preferences 0.707∗∗∗ 0.732∗∗∗ 0.708∗∗∗ (0.032) (0.031) (0.032) Log distance −1.305∗∗∗ −1.268∗∗∗ −1.297∗∗∗ (0.023) (0.024) (0.023) Log product real GDP 0.166∗∗ 0.163∗∗ 0.164∗∗ (0.053) (0.053) (0.052) Log product real GDP p/c 0.546∗∗∗ 0.547∗∗∗ 0.548∗∗∗ (0.050) (0.050) (0.050) Regional FTA 0.980∗∗∗ 0.980∗∗∗ 0.975∗∗∗ (0.129) (0.128) (0.128) Currency union 1.185∗∗∗ 1.268∗∗∗ 1.172∗∗∗ (0.123) (0.126) (0.124) Land border 0.285∗ 0.290∗∗ 0.293∗∗ (0.113) (0.112) (0.112) Number landlocked −1.033∗∗∗ −1.009∗∗∗ −0.987∗∗∗ (0.209) (0.207) (0.207) Number islands −1.602∗∗∗ −1.599∗∗∗ −1.655∗∗∗ (0.191) (0.190) (0.190) Log product land area 0.493∗∗∗ 0.498∗∗∗ 0.510∗∗∗ (0.042) (0.041) (0.041) Common colonizer 0.595∗∗∗ 0.701∗∗∗ 0.570∗∗∗ (0.067) (0.065) (0.068) Currently colonized 0.734∗∗ 0.734∗∗ 0.744∗∗ (0.268) (0.256) (0.269) Ever colony 1.255∗∗∗ 1.251∗∗∗ 1.222∗∗∗ (0.115) (0.117) (0.114) Common country 0.307 0.103 0.281 (0.596) (0.678) (0.592) Year fixed effects yes yes yes Country fixed effects yes yes yes Adjusted R2 0.705 0.706 0.706 RMSE 1.805 1.804 1.805 F Statistic 268.30∗∗∗ 265.73∗∗∗ 269.23∗∗∗ Observations 227,145 227,145 227,145 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors (clustering at the country-pair level) are reported in parentheses. – The dependent variable is defined as log of real bilateral trade in US$. – Intercept, year, and country controls are not recorded. 35

Chapter 3

Linguistic Barriers in the Destination Language Acquisition of Immigrants∗

3.1 Introduction

Already the biblical description of the fall of the Tower of Babel acknowledged the fact that differences and diversity between languages impose major obstacles for human communication. A range of empirical studies have shown that linguistic barriers constitute distinctive hurdles for international factor flows, e.g., in international trade (Isphording and Otten, 2013; Lohmann, 2011) or international migration flows (Adsera and Pytlikova, 2012; Belot and Ederveen, 2012). On the individual level, language skills have been analyzed as being a crucial determinant for the economic and social integration of immigrants in their destination country, starting with early work by Carliner (1981) and McManus et al. (1983) and more recently estimating strong wage effects for destination language proficiency (Bleakley and Chin, 2004; Chiswick and Miller, 1995; Dustmann and van Soest, 2002). These wage effects arise from the role of language as a medium of everyday and working life, constituting an important productive trait of individuals (Crystal, 2010). Furthermore, low proficiency may also act as a signal of foreignness, facilitating discrimination and differentiation (Esser, 2006). Apart from wages, language proficiency is related to further economic outcomes, such as employment status (Dustmann and Fabbri, 2003), occupational

∗Co-authored with Ingo E. Isphording (IZA). This chapter is published in the Journal of Economic Behavior & Organization, 105, 2014. A preliminary version of this chapter is available as Ruhr Economic Paper #274 and IZA Discussion Paper No. 8090. The authors are grateful to Thomas K. Bauer, John P. Haisken-DeNew, Julia Bredtmann, Carsten Crede, Michael Kind, Jan Kleibrink, Maren Michaelsen, William Neilson, the participants of the EEA 2011, the EALE 2011, and the International German Socio-Economic Panel User Conference 2012 for helpful comments and suggestions. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 36

choice (Chiswick and Miller, 2007), and locational choice (Bauer et al., 2005). Language skills are not randomly distributed: rather, they display the outcome of a systematic human capital investment decision influenced by costs and expected benefits (Chiswick and Miller, 1995). This study is concerned with the analysis of a specific cost factor of language acquisition related to the origin of an immigrant. The degree of difficulty learning a new language depends on the degree of dissimilarity of the mother tongue of immigrants to the language of the destination country. This linguistic distance, denoting differences between vocabularies, phonetic inventories, grammars, scripts, etc., is expected to crucially affect the efficiency of language learning and to raise the costs of human capital investment. In spite of the strong impact of the skills of immigrants in the destination language on their integration process, the literature on the determinants of the acquisition of the language of their destination remains surprisingly scarce. The systematic analysis of the determinants of language proficiency started with the early work by Evans (1986) comparing immigrants in Germany, the US, and Australia. More recently, Chiswick and Miller (1999, 2002, 2005) provide a comprehensive analysis of the language acquisition of immigrants in the US. For Germany, Dustmann (1999) analyzes the language proficiency as a jointly determined outcome along with migration duration. Dustmann and van Soest (2001) takes into account potential misclassification in self-reported language proficiency and Danzer and Yaman (2010) analyze German language proficiency as a function of enclave density. Still, the influence of characteristics related to the country of origin, such as the linguistic distance faced by immigrants, remains an under-researched area (Esser, 2006). The major challenge in analyzing the effect of linguistic barriers on the language acquisition of immigrants is to operationalize the linguistic distance for use in large scale micro data studies. We propose drawing from comparative linguistics and using an innovative linguistically based operationalization of linguistic distance, the so-called normalized and divided Levenshtein distance calculated by the Automated Similarity Judgment Program (ASJP). The ASJP approach offers advantages in terms of transparent computation and general applicability. We compare its benefits to those of three other approaches previously used in further applications in the economic literature to measure linguistic distance: (i) The WALS measure, which uses differences in language characteristics, (ii) the TREE measure, which is based on a priori knowledge on language families, and (iii) a measure based on average test scores of native US foreign language students (SCORE). Combining this information on language differences with US and German micro data, we provide a comprehensive analysis of the influence of the linguistic origin on the acquisition of the destination language proficiency. The US and Germany are excellent examples for analyzing the language acquisition of immigrants. Both countries have a long history as CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 37 significant immigration hubs, receiving immigrants from a large variety of source countries. The present study contributes to the literature of the determinants of language proficiency in several ways. First, we provide a comprehensive overview of the different methods of deriving a measure of language differences applicable to the analysis of the role of languages in economic behavior. Second, we introduce the ASJP approach as an easily and transparently computed measure of linguistic dissimilarity between languages. Moreover, this new approach to measuring linguistic distance is applicable to any of the world’s languages, and offers specific advantages compared to other linguistic and non-linguistic approaches used in the previous literature. We apply the derived methods to explain the language acquisition of immigrants in the US using the American Community Survey (ACS) as a very recent data source. Finally, we contribute to the literature by taking advantage of the general applicability of the linguistically based methods and extend our analysis beyond the case of Anglophone countries using data from the German Socio-Economic Panel (SOEP). Our results suggest that the linguistic barriers raised by language differences play a crucial role in the determination of the destination-country language proficiency of immigrants. Regardless of the method employed, we estimate large initial disadvantages by linguistic distance for immigrants both in the US and in Germany. In Germany, these initial differences in language skills decrease with a moderate convergence over time. Contrarily, in the US, the initial disadvantages increase over time. The gap between immigrants from different linguistic groups becomes larger with the time of residence. A potential explanation for the opposing results might be found in the higher prevalence of linguistic enclaves in the US, leading to different long-term incentives for investment in language skill in the US and Germany. The estimated differences by linguistic origin witness to the great influence of linguistic background on the economic integration of immigrants. This role should be accounted for in the design of integration policy measures. The results allow the identification of potential target groups for policy intervention. Typical measures aiming at increasing the average language proficiency of immigrants have relied on lump sum payments or fixed classroom hours for language classes. Public spending for language acquisition support might be more effective when a priori information about the expected difficulties is taken into account to specifically address target groups prone to insufficient levels. The remainder of the paper is organized as follows. In Section 3.2 we provide an overview of the measurement of linguistic differences employed in our analysis. Section 3.3 describes the data, Section 3.4 outlines our empirical model. The findings obtained from our empirical analysis are presented and discussed in Section 3.5, and Section 3.6 concludes. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 38 3.2 Measuring Linguistic Distance

The massive increase in migration flows during the last decades have shaped previously homogeneous populations into linguistically and culturally diverse melting pots. Immigrants face very different costs of language acquisition, associated with their linguistic origin. The influence of the first language (L1) on the acquired language (L2) is a common research topic in linguistics: A larger linguistic distance between L1 and L2 is believed to hamper any potential language transfer (the application of knowledge in the mother tongue to second languages) and to make it more difficult to differentiate between different sounds and words. Linguistic studies typically analyze the effect of linguistic distance employing small samples or case studies. An overview and notable exception can be found in Van der Slik (2010). The effect of linguistic distance on language acquisition can also be interpreted within an economic framework. The acquisition of language skills is an investment in a type of human capital with a high degree of specificity. Analogously to the restricted portability of source-country education (Friedberg, 2000), language skills are restricted in their portability across borders. The value of language skills outside a certain country can be very low, and immigrants have to invest in destination language skills as a prerequisite for successful integration. The imperfect portability of source-country language proficiency is a cost factor in the acquisition of the destination language. The linguistic distance indicates this portability of source-country language skills to the destination country. The larger the linguistic distance, the lower is the applicability of source-country language knowledge in the acquisition of the destination language. This leads, ceteris paribus, to greater difficulties and higher costs in the language acquisition (Chiswick and Miller, 1999). The difficulty in analyzing the relation between linguistic distance and language skills in a large scale micro data setting lies in the operationalization of the concept of linguistic distance. While specialized linguists have dedicated their whole career to studying the difference between two specific languages, our research question requires a simple standardized and continuous measure of differences between a large set of origin and destination languages. We propose to use a measure of linguistic distance relying on the phonetic dissimilarity between languages based on linguistic research by the so-called Automated Similarity Judgment Program (ASJP). This measure, the normalized and divided Levenshtein distance, offers a continuous measure of linguistic differences and is easily computed for any pair of the world’s languages. We compare this measure with two linguistic approaches and a test-score based method that have been applied in different settings in the economic literature. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 39

The test score measure

The only work we are aware of that addresses the effect of linguistic distance on language proficiency using large micro data sets are the studies by Chiswick and Miller (1999, 2001, 2005). The construction of that measure is based on average exam scores of US American English native speakers in standardized tertiary education language courses after a fixed amount of class hours. Assuming symmetry in the difficulty of learning languages, the authors state that the difficulty of English native speakers’ learning a foreign language resembles the difficulty of speakers of this foreign language in learning English. This symmetry assumption allows using these test scores as a summary statistic for the dissimilarity between languages. The necessary classroom assessments of test scores are provided by Hart-Gonzalez and Lindemann (1993), Chiswick and Miller (1999) report the respective averages by foreign language. For example, US students learning Norwegian reached an average score of 3.0 (the highest potential score). Using this score the linguistic distance for a Norwegian native speaker learning English is defined as the

inverse: LDSCORE = 1/Score = 0.33. Since Icelandic and Faroese are assumed to be close languages to Norwegian, the same distance is assigned to these languages. Unfortunately, this test-score based measure of linguistic distance is restricted to differences of a finite set of languages from English. An excerpt of the scores and resulting distances provided by Chiswick and Miller (1999) can be found in Table 3.1. The approach, especially the underlying symmetry assumption, has been widely dis- puted in the linguistic literature (see, e.g., Van der Slik, 2010). A further disadvantage of such a test-score based approach is a potential bias by incentives and motivations to learn a foreign language that cannot be separated from the effect of differences between languages. These incentives can include different economic prospects from learning a language (differences in the applicability in the labor market), or the prestige from learning new, difficult or “hip” languages. These potential biases might lead to rather counter-intuitive assessments, such as the similarly low distance between Swahili and English or Dutch and English.

Linguistic approaches: The TREE and the WALS measure

Comparative linguistics, a branch of linguistics that is concerned with the analysis of family ties and similarities within language families, provides alternatives to the test-score based method. To retrace the historical development of languages, language trees have been developed to arrange languages into diﬀerent families. These language trees depict the “genealogical” relations between languages and allow of tracing back the development of languages to likely extinct common ancestors. Most prominently, the Ethnologue Project CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 40

(see, Lewis, 2009) aims at evaluating the family relations between all known languages in the world. Using this information about the family relations between languages, it is possible to derive a measure of the linguistic distance between languages by counting the number of branches between the languages. While offering a convenient and continuous measure of linguistic distance, although with a comparably low number of increments, the resulting measure is build on strong and arbitrary assumptions of cardinality along the language tree and makes it difficult to include isolated languages (such as Korean or Basque) in the analysis. Two recent studies apply this approach to measure the effect of linguistic distance in a macroeconomic framework. Desmet et al. (2009) use a measure based on steps through the branches of a language tree to assess the effect of linguistic diversity on redistribution. Adsera and Pytlikova (2012) use a language tree approach to analyze the role of linguistic barriers in migration flows. Using the Ethnologue information, they define a language proximity index that takes on the value of 0 for languages without any family language relation, and 1 for being the same language. Between these extreme values, the language proximity indicator takes on values of 0.1, 0.25, 0.45 and 0.7 for sharing up to four levels of family relations. As both approaches by Desmet et al. (2009) and Adsera and Pytlikova (2012) rely on more or less arbitrarily chosen assumptions on cardinality and functional form, we employ the one by Adsera and Pytlikova (2012) due to its straightforward computation. Figure 3.1 illustrates a subset of a language tree to outline its computation. Since Portuguese and Spanish share the first four common branches: Indo-European, Italic, Romance, and Italo-Western, this is coded as a linguistic proximity of 0.7. English and German only share three branches: Indo-European, Germanic, and West. Therefore, the approach leads to a proximity indicator for this language pair of 0.45. The linguistic distance is again defined as the inverse of this proximity indicator:

LDTREE = 1/P roximity. Apart from Ethnologue, a second information source about languages is the World Atlas of Language Structure (WALS). The WALS offers an online database of the structural properties of languages, such as the phonological, grammatical and lexical features of more than 2,500 different languages. The 144 different characteristics include, for example, different cases, word order or syntax. Specific grammatical features from WALS have been used recently to analyze the relation between language structure and economic behavior, such as the encoding of present and future savings behavior (Chen, 2013) or gender systems and female political participation (Santacreu-Vasut et al., 2013). Panel C of Table 3.2 lists some examples of English and German WALS features. Both languages share a low consonant–vowel ratio, but while English possesses a vowel nasalization, German does not. Using the full information on all features offered by WALS, Lohmann (2011) derives an index of linguistic dissimilarity between 0 and 1 by counting and averaging shared CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 41 characteristics between languages to explain international trade flows. While conveniently summarizing linguistic differences in a number of different dimensions, the approach relies on the more or less arbitrary assumption of the equal importance of each linguistic feature. More importantly, the WALS database suffers from highly unbalanced data, since not every WALS characteristic is assessed for every language. This leads to the fact that the distance between some languages relies on a very small subsets of the commonly assessed WALS features, which potentially generates a large measurement error in the variable. To reduce this measurement error (with the trade-off of losing observations), we only include distances between languages that are based on at least 20 out of the 144 available characteristics.

The Automatic Similarity Judgment Program

The main focus of our analysis is the application of a new and innovative way of measuring linguistic distance, the so-called Automatic Similarity Judgment Program developed by the German Max Planck Institute for Evolutionary Anthropology.1 This project aims at developing an automatic procedure to evaluate the phonetic similarity between all of the world’s languages and offers a convenient way of deriving a continuous measure of linguistic differences that is purely descriptive in nature. As such, it might be used to derive language trees (which is its original purpose) but does not rely on any prior expert opinion on language families, as does the TREE approach. The basic idea behind the ASJP is the automatic comparison of the pronunciation of words across languages. A more similar pronunciation proxies the number of cognates, word pairs between languages with common ancestors, which then again indicates a closer relation between the languages. Petroni and Serva (2010) and Brown et al. (2008) demonstrate that the language relations predicted by the ASJP coincide closely with expert opinions on language relations taking into account any available language characteristics, despite the fact that it is only based on simple comparisons of word lists. To implement this “lexicostatistical” approach, the ASJP uses a core set of vocabulary for each language, describing common things and environments, called the Swadesh word list (Swadesh, 1952). The Swadesh list consists of words which are deductively chosen according to their availability in as many languages as possible, so that synonyms for these words exist in almost any potential language. Panel A of Table 3.2 lists the words used, which comprise parts of the human body, environmental descriptions, and basic words of human communication such as classifiers or personal pronouns. To focus on the pronunciation instead of the written word, these words are transcribed into a phonetic script, the ASJP code. The ASJP code uses all available characters on a standard QWERTY

1Further information can be found at http://www.eva.mpg.de. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 42 keyboard to represent sounds of human communication. For example, the English word mountain is transcribed in the ASJP code as maunt3n, while its German counterpart Berg is transcribed as bErk. The English word you is transcribed as yu, the respective German du is the same in the ASJP code, du. In the following, we go through the algorithm that leads to the continuous measure of language dissimilarity. In the first step, all word pairs from the transcribed 40-word list are compared with regard to their similarity in pronunciation. For each word pair, the minimum distance between the transcribed phonetic strings is measured as the Levenshtein distance, a measure of distance between string variables. The Levenshtein distance counts how many additions and/or subtractions are necessary to transform the string of the pronunciation of a word in language A into the string of the pronunciation of the respective word in Language B. For example, to transform the English yu into the very similar German du, only the first sound has to be changed. Whereas for the very dissimilar words mountain transcribed as maunt3n and Berg (bErk), all of the seven sounds of maunt3n have to be changed or removed. This first step results in a word-by-word absolute distance 2 D(αi, βi) between item i of two languages α and β. Examples for the transcription and determination of the word-by-word minimum distance are listed in Panel B of Table 3.2.

Taking a simple average across all M word pairs αi and βi, i = 1, .., N results in the normalized Levenshtein distance (LDN):

1 X LDN(α, β) = D(αi, βi). (3.1) M i This simple normalized Levenshtein distance might indicate a closeness between languages if languages shared the same set of commonly used sounds in communication. These potential similarities in phonetic inventories (the sum of speech sounds used in a particular language) between two compared languages do not conclusively hint at a genealogical relation between the languages, but might rather produce a similarity by chance. To filter out similarities by common phonetic inventories, a global average distance Γ(α, β) between all non-related items of the languages α and β is defined by comparing each word of the first language with all non-related words from the second language. This distance takes into account the overall similarity in phonetic inventories irrespective of the meanings of the words:

1 X Γ(α, β) = D(αi, βj). (3.2) M(M − 1) i6=j The ﬁnal normalized and divided linguistic distance is then deﬁned as the quotient between

2We draw in our notations from Petroni and Serva (2010). CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 43 the normalized linguistic distance and the global distance between α and β:

LDN(α, β) LDND(α, β) = . (3.3) Γ(α, β) The resulting continuous measure can be broadly interpreted as a percentage measure of dissimilarity between languages, with lower numbers indicating a closer relation. In a few cases, the resulting numbers are bigger than 100%, indicating a dissimilarity that exceeds a potentially incidental similarity between languages that would be expected due to similar phonetic inventories. The ASJP algorithm allows including or excluding loan words from different languages, e.g., the predominance of former Latin words in many of the European languages. While it makes sense to exclude these loan words in the analysis of the long-term development of languages, we include these loan words in our analysis, as they lead to certain similarities of languages that might ease the later language transfer in the acquisition process.3 The normalized and divided Levenshtein distance offers some advantages compared to previous measures of linguistic distance, which lead to more precise and efficient results in economic and social science applications. First, the measure is easily and transparently computed and is purely descriptive in nature, as such it does not rely on any a priori expert information on language relations. Second, due to this purely descriptive nature, it is not likely to be biased by economic incentives. Third, it offers a high variation as it is not restricted to certain parameter values. Lastly, it is comprehensive (all relevant languages are covered by the ASJP database) and can be used for any destination-country language included in the ASJP database. Therefore, it not only allows the analysis of important immigration countries such as the US, the United Kingdom, Canada, Germany, and France, but also permits the analysis of immigrants from rather “exotic” countries with typically few observations that are otherwise excluded from datasets. The comprehensiveness of the database further allows analyses concerning South–South migration, including rather seldom analyzed languages. This is a major advantage compared to the test-score based approach of Chiswick and Miller (1999), which is restricted to distances from English.

Identiﬁcation issues

We rely in our estimations on four measures of linguistic diﬀerences between the destination- and source-country language that diﬀer in their ranges of availability and in the restrictive- ness of their necessary assumptions. The test-score based measure (SCORE) is compelling with its encompassing nature, but relies on a strong symmetry assumption and is poten-

3The necessary software to compute the distance matrix is available at http://www.eva.mpg.de. The complete distance matrix used in our analysis is available upon request. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 44 tially biased by differences in incentives to learn a specific language. Most importantly, it is restricted to distances from English. In our US sample, it is available for up to 56 different source-country languages, when expert opinions on close language relations are taken into account to maximize the scope of the measure (Chiswick and Miller, 1999). Compared to this test-score measure, linguistically based approaches offer a more general framework to asses the distance between languages. The tree approach (TREE) derives a measure of distance by counting the number of shared branches in language trees, relying on prior knowledge of language family relations. It is based on strong assumptions on functional form and cardinality. Due to the completeness of the language family classifications by Ethnologue, this approach is available for the distances of 85 languages toward English and 83 languages toward German. Using external databases on language characteristics and pronunciation, the WALS and the ASJP approach offer ways to assess the differences between languages in a more descriptive manner. Neither approach relies on a priori expert knowledge of language families. However, the WALS approach has to make assumptions on cardinality. The data restrictions of the WALS database reduce the number of available languages to 67 languages in the case of the US sample and 68 languages in the case of the German sample. The ASJP database does not suffer from these restrictions, offering sufficient information for almost any language in the samples. We can rely on information for 85 languages in the US sample, and 83 languages in the German sample, providing the same applicability as that of the TREE approach. Because of its general applicability and descriptive nature, we argue that the ASJP approach, based on simple comparisons of pronunciations of word lists, offers the most appropriate way to measure linguistic distance and is superior for the application at hand. Although the ASJP approach includes much broader information on source-country languages, for the sake of comparability we restrict our estimations to immigrants from those source countries for which we have common information using all four approaches. Table 3.3 summarizes the three closest and the three most distant languages from En- glish and German according to the four different measures of linguistic distance. Consistent across the different measures, the closest languages consist of members of the Germanic language family. Some advantages and disadvantages of the measures employed are already apparent in this table. Due to the low number of increments within the measurement scale, both the TREE and the SCORE approach show only a small variation between the closest and furthest languages. Therefore, a range of languages shares the closest and the most distant position, respectively. In contrast, the ASJP and the WALS measure offer a high variation in the data. The comprehensiveness of the ASJP database allows including more remote languages, such as the Caribbean Creole languages, in the analysis, which are not covered by the other approaches. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 45

Regardless of the approach employed, the identification of an effect of linguistic barriers on the language acquisition might be biased by a correlation between linguistic distance and unobservable further cultural differences in habits and behavior (Chen et al., 1995). These unobserved cultural differences might hamper the identification of language barriers in terms of an omitted variable bias. To address this identification issue, we additionally control for the geographic distance between the destination and the immigrant’s source country. Moreover, we use a measure of genetic differences between populations as a proxy for cultural differences. Spolaore and Wacziarg (2009) combine the frequencies of gene manifestations in populations sampled by Cavalli-Sforza et al. (1994) and the ethnicity composition of countries compiled by Alesina et al. (2003) to derive a measure of the average genetic distance between countries. The change in genes, the emergence of new alleles, happens randomly at an almost constant rate. This constant rate of change over time makes it a reasonable proxy for the time populations spent separated, making the genetic distance an “excellent summary statistic capturing divergence in the whole set of implicit beliefs, customs, habits, biases, conventions, etc. that are transmitted across generations—biologically and/or culturally—with high persistence.” (Spolaore and Wacziarg, 2009, p. 471). Including this measure of genetic distance as a proxy for cultural distance and assuming a reasonable correlation between the measured genetic differences and any unobservable cultural differences should allow the identification of the isolated effect of linguistic distance in the estimations.4 The linguistic, genetic, and geographic distance are, due to their parallel emergence over time, likely to be highly correlated. High pairwise correlations could lead to difficulties in the identification of single effects, but the pairwise rank correlations in Table 3.4 are far from perfect. The linguistic distance measures are highly correlated among each other, increasing our confidence in these measures. The correlation between linguistic and geographic and especially between linguistic and genetic distance is distinctively lower.5

3.3 Data

To assess how the diﬀerent approaches to measuring linguistic dissimilarities fare in explaining the diﬀerences in the language acquisition of immigrants, two sources of

4The data on genetic differences was originally gathered by Cavalli-Sforza et al. (1994) for 42 subpopulations. Spolaore and Wacziarg (2009) extended this data to genetic differences between 180 countries by weighting it using data on the composition of ethnicities of countries compiled by Alesina et al. (2003). It is stressed again at this point that the measure of genetic distance focuses solely on genetic distance based on neutral change, not caused by evolutionary pressure, and therefore does not explain differences in language acquisition due to superior skills or ability. 5Due to the lag of normally distributed measures, we report the rank correlations instead of the Pearson correlation coefficients. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 46 individual data are employed for the estimations. Large scale micro data from the American Community Survey (ACS) offers a comprehensive representative sample of the American immigrant population. Furthermore, using a dataset from an English speaking destination country allows us to compare the linguistically based approaches with the test-score measure by Chiswick and Miller (1999) due to its restriction to distances toward English. To take advantage of the comprehensive nature of the linguistically based measures of linguistic distance, we further use data from the German Socio-Economic Panel (SOEP) to analyze the influence of the linguistic origin on the language acquisition in a non English-speaking country. Besides this new application, the SOEP offers the benefit of a broader range of individual characteristics that are unobservable in census-like data such as the ACS. The ACS data is taken from the 2006–2010 Public Use File and used as a pooled cross section. The dataset includes a self-reported measure of language skills which indicates English proficiency on a four point scale ranging from “Not at all/Bad” to “Very Well,” which constitutes our dependent variable. To focus on the potential workforce, the sample is restricted to immigrants between 17 and 65 years of age. As we want to concentrate our analysis on immigrants who acquire a destination language as an additional language, we restrict the sample to immigrants arriving at an age of 17 or older, and who originate from a non-English speaking country. After excluding observations with missing information, the pooled sample consists of 514,874 observations. A disadvantage of using the ACS is that it only offers scarce background information. As explanatory variables in our model, we use information on the time of residence, the age at arrival, individual education, sex, and marital status.6 We also include indicators of the source countries’ geopolitical world region and the year of observation to control for region- and time-fixed effects.7 To bring the analysis beyond the case of English-speaking destination countries, we use the German SOEP as a long-run panel which is an excellent data source for immigration- and integration-specific research, due to its over-sampling of immigrants and a migration- specific background questionnaire.8 The sample used in this study covers the period

6We recode the information on highest degree to compute years of schooling using a modified version of the definition proposed by Jaeger (1997) adapted to the categories of the ACS. Specifically, we recode: No schooling completed = 0, Nursery school to grade 4 = 4, Grade 5 or grade 6 = 6, Grade 7 or grade 8 = 8, Grade 9 = 9, Grade 10 = 10, Grade 11 = 11, Grade 12, no diploma = 12, High school graduate = 12, Some college, but less than one year = 13, One or more years of college, no degree = 13, Associate’s degree = 14, Bachelor’s degree = 16, Master’s degree = 18, Professional school degree = 18, Doctoral degree = 18. 7The geopolitical regions are defined following the MAR project, see http://www.cidcm.umd.edu/mar. 8The SOEP is a panel survey conducted since 1984 covering more than 20,000 individuals per wave. For more information, see Haisken-DeNew and Frick (2005). The data used in our analysis was extracted using the Add-On package PanelWhiz for Stata. PanelWhiz (http://www.PanelWhiz.eu) was written by John P. Haisken-DeNew ([email protected]). See Haisken-DeNew and Hahn (2006) for details. The PanelWhiz generated DO file to retrieve the data used here is available upon request. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 47 between 1997 and 2010. Until 2007, questions concerning the language proficiency of immigrants were included in every second wave, and on an annual basis after 2007. Analogously to the ACS sample, we restrict the SOEP sample to immigrants between 17 and 65 years of age who were at least 17 years of age when migrating to Germany, and who were born in a non-German speaking country. Furthermore, we exclude Ethnic Germans and asylum seekers from the sample. After excluding observations with missing values, we end up with a sample of 5,803 person-year observations which we use in a pooled cross-section. Similarly to the ACS, the SOEP offers information on self-reported German (oral) proficiency. The self-reported measure of language proficiency is fivefold, but because of the small number of individuals indicating the category “Not at all,” we recode this information to derive an analog fourfold ordinal measure ranging from “Not at all/Bad” to “Very Well.” The survey character of the SOEP offers a broader range of information about the individual characteristics shaping the language acquisition process. The factors influencing the language acquisition of immigrants can be divided into three groups: the exposure to the destination-country language, the efficiency of their learning ability, and the economic incentives of learning the new language (Chiswick and Miller, 1995). Our main variable of interest—the linguistic distance—affects the efficiency in acquiring the new language, decreasing the potential of any lexical transfer or portability of their proficiency in the source-country language. The efficiency of learning a new language is further controlled for by individual years of education, an indicator of good proficiency in the source-country language (as a proxy for literacy) and the age at entry, related to neurobiological research demonstrating a decreased efficiency for older arrivers (Newport, 2002). We model the effect of exposure to the destination-country language by including five variables in our estimation model. The simple ‘learning by doing’ effect is captured by a function of the years since migration. Moreover, we account for family composition characteristics captured by the number of children, marital status, and the German nationality of the spouse. The relation of these factors to the language acquisition process is ambiguous, because they lead to a social exclusion or inclusion of immigrants. Finally, an indicator for neighboring countries of Germany serves as a proxy for the probability of pre-migration exposure to the German language. The economic incentives for learning a new language are primarily influenced by the expected length of stay, shaping the time horizon of the expected benefits. An indicator variable for having family ties abroad captures potential return plans that might alter the economic incentives to invest in the destination language. Our estimation model also includes an indicator for immigrant’s sex and controls for the source country’s geopolitical CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 48 world region and the year of observation using region- and time-fixed effects. We augment both individual datasets—the ACS and the SOEP—with a set of aggregated country characteristics. These characteristics capture aspects of the relation between the immigrant’s country of birth and the country of residence that might be correlated with the linguistic distance. First, we include the share of immigrants from the migrant’s source country among the destination country’s population to capture potential network and enclave effects. The data on bilateral migrant stocks are taken from United Nations (2012). Ethnic enclaves may reduce the incentives for immigrants to acquire destination-country specific abilities such as proficiency in the official language. Although the share of immigrants of the same source country is only a raw proxy for the immigrant’s neighborhood, it might still provide some insights into the role of networks and enclaves in the acquisition of foreign language skills. Second, we control for the geographic distance, which serves as a proxy for the individual costs of migration. The geographic distance is defined as the geodesic distance between the capitals of the source and the destination country in 100 kilometers.9 Lastly, we include a measure of the genetic distance between the source and the destination country as discussed in Section 3.2, which serves as a proxy for cultural differences.10 As neither of our micro data sources (the ACS and the SOEP) offer information on the mother tongue of an immigrant, the linguistic distance is assigned by the predominant language of the country of birth. In multi-lingual countries, languages are assigned as the most prevalent native language (excluding lingua francas, i.e., commonly known foreign languages used for trade and communication across different mother tongues), which is identified using a multitude of sources, including factbooks, encyclopedias, and Internet resources.11 To allow easier comparison between the differently defined measures, we standardize each measure to have a mean of zero and a standard deviation of one.

3.4 Method

This data setup, the ACS and SOEP micro data combined with the measures of linguistic distance, allows us to estimate the language proﬁciency L as a function of the linguistic distance and the control variables, both on an aggregated and on the individual level. To get a ﬁrst glimpse into the relationship between linguistic barriers and the language

9The geographic distance data are compiled by researchers at Centre d’Etudes Prospectives et d’Informations Internationales (CEPII) and available at http://www.cepii.fr/anglaisgraph/bdd/distances.htm. 10Descriptive statistics for the ACS and the SOEP sample are presented in Table 3.A1 in the Appendix. Table 3.A2 in the Appendix provides a description of the variables used in our estimations. 11A comprehensive index of assigned languages with further explanations is available upon request. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 49 acquisition of migrant groups, we start with estimations on the aggregated level. In these estimations, we explain the average language proficiency by source country and year of observation. As dependent variable, we use predictions from a first stage explaining the S individual language proficiency Lit by a fully interacted set of source-country (cj ) and time indicators (Tk) and a set of individual characteristics Xit (gender, marital status, years since migration and age at entry):

J K 0 X X S Lit = β0 + Xitβ + γcj Tk + εit. (3.4) j=1 k=1

From this ﬁrst stage, we derive averages of the predicted language proﬁciency by source country and year of observation (Ldjt). In the second step, we then explain these predicted values by the respective linguistic distance (LDj) and a set of aggregated source-country and country-pair characteristics (Zjt):

0 Ldjt = δ0 + δ1LDj + Zjtη + εjt. (3.5)

Although this specification on an aggregated country-of-origin level provides some first insights in the relation between linguistic barriers and the language acquisition, it ignores further available information on individual migration experience and potential interactions between the linguistic barriers and individual characteristics. Therefore, in a second step we change to the individual level and model the destination language proficiency as:

0 Lit = β0 + β1LDi + β2YSMit + Xitγ + εit. (3.6)

Here, LD depicts the linguistic distance between the source- and destination-country languages, YSM represents the years since migration, and X is a vector including the control variables. In the following, we refer to the model depicted by Equation 3.6 as Model 1.12

In Model 1, β1 represents an average effect of linguistic origin for all immigrants. However, it is likely that the linguistic distance not only imposes an initial barrier to language acquisition, but also affects the steepness of the language acquisition. Two different profiles are imaginable. On the one hand, recent immigrants with a distant linguistic background might have higher incentives to invest in language skills than

12For the sake of brevity, we present here and in the following only the linear notation of our estimation models. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 50 linguistically close immigrants due to decreasing returns to invested effort. This would lead to a convergence over time. On the other hand, the hurdles imposed by language barriers can discourage investments and might lead to flatter acquisition profiles for distant immigrants. This would then lead to a divergence for immigrants from different linguistic origins, leaving linguistically distant immigrants worse off. To address this potential convergence or divergence, we allow the disadvantage by the linguistic distance to vary with the years since migration. We include an interaction of both variables LD × YSM in Equation 3.7. We will refer to this specification as Model 2:

0 Lit = β0 + β1LDi + β2YSMit + β3LDi × YSMit + Xitγ + εit. (3.7)

In Model 2 the main effect indicated by β1 shows the effect of linguistic distance on language ability at the time of immigration and β3 depicts the change in the steepness of the assimilation profile. A convergence in skill levels over time should be represented in a positive coefficient β3, indicating a catching up to immigrants with a lower linguistic distance. A negative β3 would imply a divergence. Linguistically more distant immigrants would then face flatter assimilation profiles than immigrants with a lower linguistic distance. We start our analysis by estimating our models using Ordinary Least Squares (OLS), separately for the four measures of linguistic distance in the US case and three measures in the German case. To interpret the OLS results using the ordinal language proficiency variable quantitatively, we have to impose strong cardinality assumptions. To take into account this ordinal character of the dependent variable and to derive quantitatively interpretable results, we repeat the estimations using Ordered Logit regressions and use graphical representations to interpret the interaction between linguistic distance and years since migration. Throughout all specifications in our analysis, we use (cluster)-robust standard errors to correct for possible heteroskedasticity in the data.

3.5 Results

A first descriptive look at the relation between language proficiency and the different measures of linguistic distance is provided in Table 3.5. The distribution of language skills in the US and Germany is quite different. While in the US about 37% of all immigrants report a “Very Well” proficiency, in Germany only 15% report the highest category. The expected negative relation between linguistic distance and language proficiency does not appear in the unconditional means reported in Table 3.5 in the US sample, ASJP, WALS and SCORE even suggest a marginally positive relation. In the German sample, the CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 51

relation between linguistic distance and language proficiency is distinctively negative on the descriptive level: across all three available measures, we observe a decrease in the linguistic distance as the language skills increases, with the lowest average distance in the “Very Well” category. However, it remains to be seen how potentially correlated individual characteristics change this first descriptive picture. The results of the estimations of equation 3.5 on the aggregated source country level are summarized in Table 3.6.13 We find a strong negative relationship between linguistic barriers and the average language proficiency. This result is robust, and can be observed both in the US and the German data, but differs distinctively in magnitude across different measures of linguistic distance. Assuming cardinality in our dependent variable, the coefficient of linguistic distance measured by the ASJP approach indicates that an increase of the linguistic distance by one standard deviation (roughly the difference in the distance to English between German and Romanian) decreases the average language proficiency by 0.17 points on the 0–3 scale in the US sample and 0.19 points in the German sample. Using the WALS or TREE approach shows a decrease by only 0.11 points in the US sample, while the TREE approach indicates a decrease of 0.2 points in the German sample, comparable to the ASJP sample. Concerning the control variables, migrant stocks are negatively related to the average language proficiency, hinting at potential negative influences of ethnolinguistic enclaves, see also Chiswick and Miller (2002); Dustmann and Fabbri (2003) and Cutler et al. (2008). We further find positive relationships between geographic and genetic distance (as a proxy for cultural differences) which we interpret as indirect evidence for selection on unobservable motivation and ability, while the positive coefficients of GDP per capita capture potential differences in pre-migration language exposure and education. The results provided in Table 3.7 bring the analysis to the individual level. Table 3.7 summarizes the results of the OLS estimations for the ACS sample, separately for the different measures of linguistic distance. As already seen in the aggregated results, the estimations of the effect of the linguistic distance remain very volatile to the choice of employed measure. This highlights the importance of applying different available measures, rather than relying on only one approach, to get a comprehensive insight into the relation of linguistic barriers and the language acquisition. The results of Model 1 are summarized in Panel A. Across all different methods, the effect is highly significant and negative. Similar to the aggregated results, the ASJP approach indicates the strongest influence of the linguistic origin on the language acquisition: an increase by one standard deviation is related to a lower language proficiency by 0.24 points on the 0–3 scale, while estimations using the TREE, the WALS and the SCORE approach indicate a decrease by 0.10 to 0.12

13We generated all estimation output tables using the Stata routine estout by Ben Jann (see, Jann, 2007). CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 52 points. The coefficients in Panel A represent an average effect of linguistic distance for immigrants sharing a common linguistic background. To analyze whether this disadvantage is increasing or decreasing over time of residence, Model 2 includes an interaction term between the years since migration and the linguistic distance. The respective results are included in Panel B. In this specification, the main effect of linguistic distance is to be interpreted as an initial disadvantage at the time of immigration. Compared to Model 1, this initial disadvantage is smaller than the average difference by linguistic origin in Panel A. This results from a negative interaction between linguistic distance and the years since migration. Although we observe an overall positive language assimilation over time, the language assimilation profile becomes flatter with increased linguistic distance. Linguistically distant immigrants not only experience a higher initial disadvantage in their language acquisition, but also seem to experience a slower acquisition of English as destination language. After immigration, the initial gap between the immigrants from close and from distant linguistic origins increases over time. This pattern is robust across all four different models, while again the effect is strongest for the ASJP approach.14 To drop the cardinality assumption and to take the ordinal character of the self-reported language proficiency into account, we estimate both models using Ordered Logit regressions instead of OLS. Table 3.8 provides the marginal effects of the linguistic distance on the probability of reporting specific categories of language proficiency in Model 1.15 Increasing the linguistic distance quantified by the ASJP approach by one standard deviation decreases the probability of reporting “Very Well” language skills in English by about 20 percentage points. Due to the non-linear Ordered Logit model and the inclusion of an interaction term, the marginal effects of linguistic distance in Model 2 are best interpreted in a graphical manner. Figure 3.2 depicts predicted probabilities of reporting the highest category of language proficiency by different levels of linguistic distance over the time of residence. Linguistically close immigrants in the 1st percentile of the distance distribution face a initially steeper assimilation profile, linguistically distant migrants are outpaced. While this pattern sheds some light on the effect of the heterogeneity in linguistic origin on the language acquisition of immigrants in the US, the large differences in immigration policy regimes and differences in selection patterns make it difficult to generalize the results to other countries. Previous analyses using the SCORE approach have been restricted to English-speaking destination countries (Chiswick and Miller, 1999). However, English, as

14Regarding the influence of the control variables, Model 1 and Model 2 do not differ much, neither does the influence of the control variables vary with the measure of linguistic distance applied. For the sake of brevity, we do not further discuss the influence of the control variables. The respective coefficients can be found in Table ?? in the Appendix. 15The underlying coefficients and the marginal effects of Model 1 and 2 of the Ordered Logit regressions are presented in Tables 3.B2–3.B4 in the Appendix. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 53 a lingua franca in international trade, the Internet, and communication technology, might enjoy very different incentives for being learned, compared to languages which lack this worldwide predominance. Against this restriction, a major advantage of the linguistically based methods of language differences is the general applicability to any pair of languages. Taking advantage of this general applicability, we are able to extend the analysis beyond immigration to English-speaking countries. Specifically, we turn to Germany, one of the most important non English-speaking destinations for international migration. The German SOEP sample allows a similar analysis as that of the ACS data, but with a richer set of control variables including the number of children, literacy, family ties abroad, and having a native spouse.16 Table 3.9 lists the respective OLS results of Model 1 and Model 2.17 Again, we find a negative effect of the linguistic distance between the mother tongue and the destination language on the language acquisition process which differs strongly by the employed approach. To derive a quantitative interpretation, we again turn to the results of an Ordered Logit model in Table 3.8. The marginal effects of Model 1 show a negative effect of linguistic distance on reporting “Very Well” German proficiency by 1.9 to 4.4 percentage points, which is moderate compared to the US results.18 However, the results for Model 2 draw a very different picture for the German SOEP sample compared to the US results. The interaction term between the linguistic distance and the years since migration in Model 2 turns out to have a positive sign but is insignificant across all different estimations in the German case (see Table 3.9, Panel B). This slight convergence is more distinctive in the Ordered Logit results, which are illustrated in Figure 3.3 in terms of predicted probabilities of reporting “Very Well” proficiency. Immigrants from a more distant linguistic origin therefore face a steeper assimilation profile than immigrants with a close linguistic background. Instead of observing a divergence by linguistic origin, we find a convergence in language skills. Over the time of residence, the gap between the linguistically close and distant immigrants closes, linguistically distant immigrants are able to catch up. We might speculate about the driving factors of the difference between the divergence and convergence patterns in the US and in Germany. English and German are very closely related Germanic languages. This raises doubts that the differences are simply driven by purely linguistic reasons, such as that one language possesses particularly strong obstacles, e.g., by very special grammatical features, that would lead to the observed divergence. A more economically based potential explanation are differences in

16Following Dustmann (1999), literacy is assumed for individuals reporting being able to write in their mother tongue. 17The coefficients, omitted in Table 3.9, are included in Table 3.B5 in the Appendix. 18The underlying coefficients and the marginal effects of Model 1 and 2 of the Ordered Logit regressions are presented in Tables 3.B6–3.B8 in the Appendix. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 54

unobservable characteristics by different selection patterns of the migrants in the US and Germany. A perceived higher difficulty of German compared to English could lead to a self-selection of immigrants with superior skills in the acquisition of foreign languages into Germany. If this selection pattern were stronger for linguistically more distant immigrants (where the initial returns to the language acquisition would be higher), the observed competing patterns of divergence in the US and convergence in Germany could arise. However, as we control in both samples for individual education, which is expected to be correlated with unobserved ability, we should at least partially capture such a selection process. A second, in our opinion more plausible, explanation might be related to enclave effects in language acquisition. A range of studies have addressed the potentially discouraging effects of linguistical enclaves on investments in language skills (e.g., Chiswick and Miller, 2002; Cutler et al., 2008; Dustmann and Fabbri, 2003). Living in a linguistic enclave reduces the need for and potential advantages of learning the destination language, as immigrants can communicate in daily life in their mother tongue. Danzer and Yaman (2010) argue that the probability of moving into a neighborhood dominated by speakers of their own mother tongue is positively related to the own learning costs. The initial learning costs are strongly related to the linguistic distance between the mother tongue and the destination language, making it more likely for linguistically distant immigrants to move into segregated neighborhoods. Neighborhood segregation needs time to take place: due to its longer migration history, the ethnic segregation within cities is much more pronounced in the US than in Germany with its comparably short-running migration history. Therefore, the observed differences in assimilation patterns are potentially driven by the larger prevalence of linguistic enclaves in the US (e.g., the famous Chinatowns and Little Italy’s in US cities). In order to test the robustness of our results, we use different subsamples of our datasets. In doing so, we split the sample: (i) by gender, (ii) between low-skilled and high- skilled immigrants, and (iii) by excluding the majority immigrant groups, i.e., Mexican immigrants in the US and Turkish immigrants in Germany, from our regressions. A summary of these sensitivity checks is provided in Tables 3.A3 and 3.A4 in the Appendix. The underlying pattern of initial disadvantage and divergence in the US is stable across all subsamples. Linguistic barriers seem to play a larger role in the case of low-skilled immigrants (having a high school degree or less) than for high-skilled immigrants. The results for Germany are less robust, likely due to the low number of observations in the SOEP data. The negative main effect of linguistic distance remains robust across all different subsamples. The interaction term between linguistic distance and the years since migration becomes positively significant for high-skilled immigrants, who seem to drive the CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 55 observed convergence in Germany. However, the convergence profile becomes insignificant when we split the sample by gender, by skill-level, and when Turks are excluded. To summarize, our results highlight the importance of linguistic origin as a factor of typically unobservable heterogeneity in the integration process of immigrants. The initial disadvantages only marginally disappear over time of residence, but linguistic barriers remain even after a long period of stay. Given the large impact of language proficiency on labor market outcomes (Bleakley and Chin, 2004; Chiswick and Miller, 1995; Dustmann and van Soest, 2002), it is likely that these differences are transferred into labor market disadvantages. Disadvantages in the language acquisition process prevent the social integration of immigrants by reducing their ability to communicate with natives. In addition, imperfect language skills can act as a signal for foreignness, opening the way to discriminatory behavior of employers and decreasing the productiveness of individuals, leading to lower employment probabilities and wages. Against the background of immigration policy design, our results hint at a way to identify target groups for supportive integration policy measures. Immigrants obviousy differ strongly in their costs of language acquisition, dependent on their linguistic background. This heterogeneity is seldomly accounted for in the design of integration policies. Policies aiming at the support of immigrant language aquisition, as currently practiced in Germany with the “Integrationskurse” system (“integration classes”), often include a lump sum payment for public language classes. This lump sum payment, restricting class hours irrespective of the actual need for support, is likely to lead to a inefficient spending of public money. In a class system that does not distinguish language students by their actual need for support, linguistically close immigrants are provided too many class hours, while linguistically distant immigrants might be outpaced. A means-tested voucher taking into account the expected costs by linguistic origin might lead to a more efficient spending of public moneys than a lump sum policy measure.

3.6 Conclusion

International labor migration is a worldwide and steadily growing phenomenon. According to UN estimates, in 2010 roughly 215 million individuals lived in a country different from their country of birth (World Bank, 2011). On a first glimpse, this is a massive number but it still accounts for only around 3% of the world’s population, a surprisingly low number given the large differentials in economic conditions. While technological progress in transportation and communication have led to a significant decrease in the initial costs of migration, cultural and linguistical borders continue to play an important role for international migration flows (Adsera and Pytlikova, 2012; Belot and Ederveen, 2012). In CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 56 this study, we provide a in-depth analysis and quantification of the linguistic barriers in destination language acquisition in Germany and the US. For immigrants, proficiency in the destination-country language leads to substantial economic returns (Bleakley and Chin, 2004; Dustmann and van Soest, 2002). However, large fractions of the immigrant population possess only insufficient levels of proficiency in the destination language. While the investment in language capital has been thoroughly analyzed in human capital frameworks (Chiswick and Miller, 1995), our knowledge of the influence of typically unobservable heterogeneity in the linguistic origin of immigrants remains limited. The linguistic distance between languages is a concept that is difficult to operationalize for its implementation in empirical models. In this study, we demonstrate four different methods providing continuous measures of linguistic differences and compare their specific advantages and shortcomings. More specifically, we draw from linguistic research and propose using a measure of linguistic distance based on comparisons of pronunciation between word lists. This method, referred to as the ASJP approach, offers a convenient way to derive a continuous measure of linguistic differences. Given its purely descriptive measurement and general applicability to any potential pair of languages, it provides an advantageous measure for the application at hand. We compare its performance with further linguistic approaches using information about language relations (TREE measure) and language characteristics (WALS measure) and a measure based on average test scores (SCORE) by Chiswick and Miller (1999). All four measures of language differences are applied to the analysis of the destination language acquisition of immigrants in the US using data of the American Community Survey (ACS). To take advantage of the general applicability of the linguistically based methods beyond the analysis of English-speaking destination countries, we extend the analysis to German microdata from the German Socio-Economic Panel (SOEP). In both scenarios, we use the different measures of linguistic distance to explain differences in self-reported measures of immigrant’s destination language proficiency. Our results indicate that the linguistic distance, the dissimilarity between the origin and destination languages, has a distinctively negative average effect on the language acquisition of immigrants. Immigrants with a distant linguistic origin face higher costs in the language acquisition than immigrants with a closer linguistic background. Furthermore, we analyze differences in the slope of the language assimilation curve that can be attributed to differences in the linguistic origin. We find different assimilation patterns for the US and Germany. In Germany, immigrants with a more distant source-country language display a steeper language assimilation profile. Initial disadvantages are reduced over time, leading to a convergence in average proficiency for immigrants from different linguistic origins. For the US, we estimate the opposite picture of diverging profiles. Gaps in the CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 57 proficiency of linguistically close and distant immigrants tend to increase over time of residence. We interpret this difference in assimilation patterns as a potential outcome of stronger enclave effects in the US. This crucial difference highlights the importance of extending the analysis beyond the case of Anglophone countries. The initial disadvantages and differences in assimilation patterns attributable to linguistic distance are able to explain a large fraction of the explained variation in the destination language proficiency. This highlights the importance of linguistic differences for the analysis of the skill acquisition of immigrants, as an influencing factor that was previously part of the “black box” of culture in the economic literature (see, Epstein and Gang, 2010). This additionally explained variation might play an important role in the design of integration policy measures. Lump sum payments for language classes might turn out to be inefficient in the presence of a high heterogeneity in the actual need for language acquisition support and compared to means-tested vouchers taking into account the expected costs of language acquisition. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 58 Figures and Tables

Table 3.1: Average Test Scores of US Language Students Average Linguistic Languages Test Scores Distance (Examples) 1.00 1.00 Laotian, Korean, Japanese 1.25 0.80 Cantonese, Mien, Hakka 1.50 0.67 Syriac, Vietnamese, Arabic 1.75 0.57 Bengali, Nepali, Greek 2.00 0.50 Serbo-Croatian, Turkish, Finnish 2.25 0.44 Spanish, Danish, Yiddish 2.50 0.40 Italian, Portuguese, French 2.75 0.36 Dutch, Swahili, Bantu 3.00 0.33 Norwegian, Swedish, Afrikaans Notes: Average test scores of American students learning foreign languages. Numbers provided by Hart-Gonzalez and Lindemann (1993), reproduced from Chiswick and Miller (1999), Appendix B.

R RR R R

Figure 3.1: Language Relations in the TREE Approach CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 59

Table 3.2: 40-Items Swadesh Word List with Computational Examples A. Swadesh Word List I You We One Two Person Fish Dog Louse Tree Leaf Skin Blood Bone Horn Ear Eye Nose Tooth Tongue Knee Hand Breast Liver Drink See Hear Die Come Sun Star Water Stone Fire Path Mountain Night Full New Name

B. ASJP Computation Word English German Distance fish fiS fiS 0 breast brest brust 1 hand hEnd hant 2 tree tri baum 4 Mountain maunt3n bErk 7

C. Examples for WALS Features Feature English German Consonant-vowel ratio Low Low Vowel Nasalization Present Absent Number of cases 2 4 Notes: Panel A displays the 40-item Swadesh sublist used in computing linguistic distance. – Panel B: Examples of computation for linguistic distance between English and German. – Panel C: Examples for diﬀerences in WALS features between English and German. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 60 667 681 685 688 00 627 632 67 00 ...... 0 0 0 0 0 0 0 1 1 ∗ Burmese (Burma) Guarani (Paraguay) Somali (Somalia) Vietnamese Guarani (Paraguay) Zulu (South Africa) Palestinian Arabic Korean Japanese 302 338 348 283 302 310 33 33 33 ...... 0 0 0 0 0 0 0 0 0 SCORE Approach Russian French English Russian Romanian –––– –––– –––– German French Norwegian Bokmal Swedish 03 43 06 00 00 30 72 58 00 00 00 00 ...... 1 1 1 1 1 1 103 103 104 104 103 103 Due to the small number of increments, a range of languages shares the “top” ∗ ∗ Amoy Minnan Chinese Palestinian Arabic Somali (Somalia) Twi Asante (Ghana) Vietnamese Korean Palestinian Arabic Yoruba (Nigeria) Vietnamese Korean Palestinian Arabic Yoruba (Nigeria) 55 55 34 95 44 12 34 50 55 30 55 55 ...... 0 0 0 0 0 0 48 51 41 41 47 42 ASJP Approach WALS Approach TREE Approach Closest and Furthest Languages to English and German Closest Furthest Closest Furthest Closest Furthest Closest Furthest Table 3.3: LanguageSt. Vincent Creole Jamaican Creole Krio (Sierra Leone) Luxembourgish DistanceBernese German LanguageDutch Language DistanceWestvlaams (Belgium) LanguageDutch Bernese German Luxembourgish Distance Distance Language Language Distance Distance Language Distance Language Distance English Westvlaams (Belgium) Distance to German Distance to German positions. Distance to English Distance to English spoken within samples are listed. – Geographic origin of language in parentheses. – Notes: – The table shows the three closest and furthest languages toward English and German according to diﬀerent measures of linguistic distance. – Only languages CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 61

Table 3.4: Rank Correlations among Linguistic, Geographic, and Genetic Distance Measures Linguistic Linguistic Linguistic Linguistic Geographic Genetic Distance Distance Distance Distance Distance Distance ASJP WALS TREE SCORE (100 km) ACS Sample Linguistic distance ASJP 1 Linguistic distance WALS 0.84 1 Linguistic distance TREE 0.84 0.83 1 Linguistic distance SCORE 0.81 0.87 0.73 1 Geographic distance (in 100 km) 0.79 0.72 0.66 0.72 1 Genetic distance 0.48 0.54 0.70 0.35 0.25 1 SOEP Sample Linguistic distance ASJP 1 Linguistic distance WALS 0.75 1 Linguistic distance TREE 0.88 0.85 1 Geographic distance (in 100 km) 0.61 0.47 0.61 – 1 Genetic distance 0.69 0.81 0.80 – 0.64 1 Notes: – Spearman Rank correlations reported. – Number of observations in the ACS sample: 514,874. – Num- ber of observations in the SOEP sample: 5,803. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 62

Table 3.5: Distribution of Language Skills across Samples Mean Linguistic Distance Proﬁciency Observations ASJP WALS TREE SCORE ACS Sample 58,523 Bad 93.79 0.40 0.91 0.46 0.11

128,384 Not bad 94.71 0.42 0.92 0.49 0.25

139,223 Well 95.62 0.45 0.93 0.52 0.27

188,744 Very well 95.20 0.46 0.92 0.52 0.37

Total 514,874 SOEP Sample 1,057 Bad 96.21 0.50 0.95 – 0.18

1,918 Not bad 95.14 0.48 0.94 – 0.33

1,946 Well 92.53 0.45 0.90 – 0.34

882 Very well 88.01 0.39 0.84 – 0.15

Total 5,803 Notes: – Column 2 shows the absolute number of observations and the relative frequencies for each category. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 63

Table 3.6: Language Ability and Linguistic Distance – Aggregated Results

ACS Sample SOEP Sample

ASJP WALS TREE SCORE ASJP WALS TREE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

Linguistic distance −0.173∗∗∗ −0.105∗∗∗ −0.105∗∗∗ −0.128∗∗∗ −0.188∗∗∗ −0.094† −0.195∗∗∗ (0.022) (0.014) (0.013) (0.012) (0.031) (0.048) (0.032) Migrant stock (% of population) −0.207∗∗∗ −0.204∗∗∗ −0.218∗∗∗ −0.213∗∗∗ −0.341∗∗∗ −0.329∗∗∗ −0.341∗∗∗ (0.020) (0.021) (0.022) (0.019) (0.044) (0.045) (0.045) Geographic distance (in 100 km) 0.006∗∗∗ 0.006∗∗∗ 0.005∗∗∗ 0.005∗∗∗ −0.001 −0.000 −0.003 (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) Genetic distance 0.003 −0.003 0.004 −0.007 0.054∗∗∗ 0.039∗∗ 0.063∗∗∗ (0.006) (0.006) (0.006) (0.005) (0.014) (0.014) (0.015) ln GDP per capita (in USD) 0.141∗∗∗ 0.136∗∗∗ 0.149∗∗∗ 0.175∗∗∗ 0.120∗∗ 0.135∗∗ 0.101∗ (0.016) (0.016) (0.015) (0.015) (0.045) (0.048) (0.047) Constant 0.622∗∗ 0.759∗∗∗ 0.562∗∗ 0.497∗ 0.716 0.723 0.929† (0.195) (0.198) (0.193) (0.197) (0.496) (0.525) (0.503) Region dummies yes yes yes yes yes yes yes Year ﬁxed eﬀects yes yes yes yes yes yes yes

Adjusted R2 0.609 0.579 0.602 0.600 0.244 0.202 0.241 F Statistic 103.0 75.6 79.1 75.4 19.4 13.3 19.0 Observations 395 395 395 395 423 423 423

Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Robust standard errors are reported in parentheses. – Depen- dent variable: predicted average language skills by source country and year. – Level of observation is the source country, destination-country language ability evaluated as source-country averages of the predicted language proﬁciency. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 64

Table 3.7: OLS Results of Linguistic Distance – ACS Sample

ASJP WALS TREE SCORE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE

Panel A: Model 1 Linguistic distance −0.241∗∗∗ −0.103∗∗∗ −0.107∗∗∗ −0.124∗∗∗ (0.004) (0.003) (0.002) (0.002)

Adjusted R2 0.406 0.404 0.405 0.409 F Statistic 16,051.0 15,351.8 16,102.4 15,706.5 Observations 514,874 514,874 514,874 514,874

Panel B: Model 2 Linguistic distance −0.180∗∗∗ −0.038∗∗∗ −0.086∗∗∗ −0.075∗∗∗ (0.004) (0.004) (0.002) (0.003) Years since migration 0.015∗∗∗ 0.014∗∗∗ 0.014∗∗∗ 0.014∗∗∗ (0.000) (0.000) (0.000) (0.000) LD × YSM −0.004∗∗∗ −0.005∗∗∗ −0.001∗∗∗ −0.003∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.407 0.405 0.405 0.410 F Statistic 14,965.9 14,489.8 15,076.1 14,881.2 Observations 514,874 514,874 514,874 514,874

Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Ro- bust standard errors are reported in parentheses. – The dependent variable is defined on a scale of 0 to 3 such that higher values indicate a higher level of language proficiency.

Table 3.8: Ordered Logit Marginal Effects of Linguistic Distance – Model 1 ACS & SOEP Sample

Bad Not bad Well Very well ME/StdE ME/StdE ME/StdE ME/StdE

ACS Sample Linguistic distance ASJP 0.063∗∗∗ 0.175∗∗∗ −0.034∗∗∗ −0.204∗∗∗ (0.001) (0.003) (0.001) (0.003) Linguistic distance WALS 0.021∗∗∗ 0.057∗∗∗ −0.012∗∗∗ −0.066∗∗∗ (0.001) (0.002) (0.000) (0.002) Linguistic distance TREE 0.032∗∗∗ 0.090∗∗∗ −0.018∗∗∗ −0.105∗∗∗ (0.001) (0.002) (0.000) (0.002) Linguistic distance SCORE 0.021∗∗∗ 0.058∗∗∗ −0.012∗∗∗ −0.067∗∗∗ (0.000) (0.001) (0.000) (0.001)

SOEP Sample Linguistic distance ASJP 0.017∗ 0.035∗ −0.033∗ −0.019∗ (0.008) (0.016) (0.015) (0.009) Linguistic distance WALS 0.038∗∗∗ 0.080∗∗∗ −0.074∗∗∗ −0.044∗∗∗ (0.011) (0.024) (0.022) (0.013) Linguistic distance TREE 0.027∗∗ 0.056∗∗ −0.052∗∗ −0.031∗∗ (0.009) (0.020) (0.018) (0.011)

Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – (Cluster- )robust standard errors are reported in parentheses. – The dependent variable is defined on a scale of 0 to 3 such that higher values indicate a higher level of language proficiency. – Marginal effects are reported at the mean of the covariates vector. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 65

Table 3.9: OLS Results of Linguistic Distance – SOEP Sample

ASJP WALS TREE

Coef/StdE Coef/StdE Coef/StdE

Panel A: Model 1 Linguistic distance −0.079∗∗ −0.160∗∗ −0.117∗∗ (0.030) (0.051) (0.037)

Adjusted R2 0.373 0.375 0.376 F Statistic 34.55 35.80 34.59 Observations 5,803 5,803 5,803

Panel B: Model 2 Linguistic distance −0.121∗ −0.187∗∗ −0.182∗∗∗ (0.050) (0.059) (0.054) Years since migration 0.017∗∗∗ 0.017∗∗∗ 0.017∗∗∗ (0.004) (0.004) (0.004) LD × YSM 0.002 0.002 0.003 (0.002) (0.002) (0.002)

Adjusted R2 0.373 0.375 0.377 F Statistic 34.41 34.69 34.59 Observations 5,803 5,803 5,803

ASJP Approach WALS Approach 1 .8 .8 .6 .6 .4 .4 .2 .2 Pr(Language skills = Very well) Pr(Language skills = Very well) 0 0

0 10 20 30 40 50 0 10 20 30 40 50 Years since migration Years since migration

TREE Approach SCORE Approach 1 .8 .8 .6 .6 .4 .4 .2 Pr(Language skills = Very well) Pr(Language skills = Very well) .2 0

0 10 20 30 40 50 0 10 20 30 40 50 Years since migration Years since migration

LD = 1st percentile LD = 50th percentile LD = 99th percentile

Figure 3.2: Predicted Language Assimilation Profiles for the ACS Sample Notes: – The predicted assimilation proﬁles are based on Ordered Logit regressions of Model 2. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 67

ASJP Approach WALS Approach .5 .3 .4 .2 .3 .2 .1 .1 Pr(Language skills = Very well) Pr(Language skills = Very well) 0 0

0 10 20 30 40 50 0 10 20 30 40 50 Years since migration Years since migration

TREE Approach .4 .3 .2 .1 Pr(Language skills = Very well) 0

0 10 20 30 40 50 Years since migration

LD = 1st percentile LD = 50th percentile LD = 99th percentile

Figure 3.3: Predicted Language Assimilation Profiles for the SOEP Sample Notes: – The predicted assimilation proﬁles are based on Ordered Logit regressions of Model 2. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 68 3.A Appendix

Table 3.A1: Descriptive Statistics – ACS & SOEP Sample ACS Sample SOEP Sample Mean StdD Mean StdD English proficiency 1.890 1.029 –– Oral German proficiency – – 1.457 0.957 Linguistic distance ASJP 95.030 5.118 93.376 8.437 Linguistic distance WALS 0.439 0.090 0.458 0.107 Linguistic distance TREE 0.921 0.065 0.915 0.101 Linguistic distance SCORE 0.507 0.121 –– Years since migration 14.804 9.980 23.688 11.268 Age at entry 27.323 8.398 25.331 6.885 Years of education 12.159 4.459 10.192 2.555 Female 0.437 0.496 0.534 0.499 Married 0.630 0.483 0.850 0.357 Children in the HH. No children – – 0.564 0.496 One child – – 0.207 0.405 Two children – – 0.148 0.355 Three or more children – – 0.081 0.273 Native German partner/spouse – – 0.218 0.413 Family abroad – – 0.263 0.440 Proficiency home language – – 0.876 0.330 Desired stay (years) – – 14.722 11.266 Neighboring country – – 0.079 0.270 Migrant stock (% of population) 1.485 1.682 1.621 1.264 Geographic distance (in 100 km) 68.951 43.260 19.434 17.727 Genetic distance 8.813 2.686 3.734 2.923 ln GDP per capita (in USD)a 8.294 1.055 9.029 0.970 Observations 514,874 5,803 Notes: – Unweighted means and standard deviations reported. – English and Ger- man proficiency is defined on a scale of 0 to 3, corresponding to the classifications “Bad”, “Not bad”, “Well”, “Very well”. – ln GDP per capita is based on 510,220 observations in the ACS sample and 5,801 observations in the SOEP sample. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 69

Table 3.A2: Variables Description – ACS & SOEP Sample Variable Description English proficiency Self-reported English proficiency German proficiency Self-reported oral German proficiency Linguistic distance ASJP Levenshtein distance normalized divided Linguistic distance WALS Measure based on structural language features by Lohmann (2011) Linguistic distance TREE Language tree measure based on Adsera and Pytlikova (2012) Linguistic distance SCORE Test-score measure by Chiswick and Miller (1999) Years since migration Years of residence in destination country Age at entry Age at entry into destination country Years of education Years of education Female Dummy = 1 if female Married Dummy = 1 if married Children in the HH. No children Dummy = 1 if no children live in the household One child Dummy = 1 if one child lives in the household Two children Dummy = 1 if two children live in the household Three or more children Dummy = 1 if three or more children live in the household Native German partner/spouse Dummy = 1 if partner/spouse is native German Family abroad Dummy = 1 if family lives abroad Proficiency home language Dummy = 1 if written proficiency in mother tongue is well or very well Desired stay (years) Years desired to stay in Germany Neighboring country Dummy = 1 if country of origin is a neighboring country of Germany Migrant stock (% of population) Migrant stock by source country as percentage of the total population Geographic distance (in 100 km) Geodesic distance between capitals in 100 km Genetic distance Weighted FST genetic distance, divided by 100 (Spolaore and Wacziarg, 2009) ln GDP per capita (in USD) Logarithm of GDP per capita in constant 2000 US dollars Region dummies 6 region dummies, indicating the source countries’ geopolitical world region. The geopolitical regions are defined following the MAR project as: 1) Western democra- cies and Japan, 2) Eastern Europe and the former Soviet Union, 3) Asia, 4) North Africa and the Middle East, 5) Sub-Saharan Africa, and 6) Latin America and the Caribbean. Year fixed effects Time dummy variables for each sample year Notes: – Geodesic distances are calculated following the great circle formula, which uses the geographic coordinates of the capital cities for calculating the distance to the capital of the US and Germany, respectively. Geographic distance reports the calculated distance divided by 100. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 70

Table 3.A3: Robustness Checks: OLS Results of Linguistic Distance – Model 2 ACS Sample

ASJP WALS TREE SCORE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

Male Subsample Linguistic distance −0.165∗∗∗ −0.056∗∗∗ −0.081∗∗∗ −0.051∗∗∗ (0.006) (0.005) (0.003) (0.004) LD × YSM −0.007∗∗∗ −0.006∗∗∗ −0.003∗∗∗ −0.006∗∗∗ (0.000) (0.000) (0.000) (0.000) Years since migration 0.015∗∗∗ 0.014∗∗∗ 0.014∗∗∗ 0.014∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.394 0.393 0.392 0.396 F Statistic 8,976.37 8,850.42 9,019.55 8,889.42 Observations 290,127 290,127 290,127 290,127

Female Subsample Linguistic distance −0.199∗∗∗ −0.026∗∗∗ −0.093∗∗∗ −0.100∗∗∗ (0.006) (0.006) (0.003) (0.004) LD × YSM −0.002∗∗∗ −0.004∗∗∗ −0.000∗∗∗ −0.001∗∗∗ (0.000) (0.000) (0.000) (0.000) Years since migration 0.013∗∗∗ 0.013∗∗∗ 0.013∗∗∗ 0.013∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.429 0.427 0.428 0.433 F Statistic 6,976.10 6,664.67 7,023.96 6,985.29 Observations 224,747 224,747 224,747 224,747

Low-skilled Subsample Linguistic distance −0.381∗∗∗ −0.059∗∗∗ −0.178∗∗∗ −0.097∗∗∗ (0.011) (0.010) (0.006) (0.007) LD × YSM −0.001 −0.002∗∗∗ 0.000 −0.001∗ (0.000) (0.000) (0.000) (0.000) Years since migration 0.020∗∗∗ 0.019∗∗∗ 0.020∗∗∗ 0.019∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.208 0.203 0.206 0.204 F Statistic 3,149.23 2,813.24 3,116.19 2,848.01 Observations 268,271 268,271 268,271 268,271

High-skilled Subsample Linguistic distance −0.142∗∗∗ −0.049∗∗∗ −0.072∗∗∗ −0.096∗∗∗ (0.004) (0.004) (0.002) (0.003) LD × YSM −0.002∗∗∗ −0.003∗∗∗ −0.001∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) Years since migration 0.008∗∗∗ 0.008∗∗∗ 0.007∗∗∗ 0.007∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.238 0.236 0.236 0.252 F Statistic 2,302.87 2,201.65 2,330.84 2,436.70 Observations 246,603 246,603 246,603 246,603

Subsample excluding Mexican immigrants Linguistic distance −0.205∗∗∗ −0.057∗∗∗ −0.089∗∗∗ −0.095∗∗∗ (0.004) (0.004) (0.002) (0.003) LD × YSM −0.002∗∗∗ −0.003∗∗∗ −0.000∗∗∗ −0.002∗∗∗ (0.000) (0.000) (0.000) (0.000) Years since migration 0.011∗∗∗ 0.011∗∗∗ 0.010∗∗∗ 0.010∗∗∗ (0.000) (0.000) (0.000) (0.000)

Adjusted R2 0.403 0.399 0.400 0.406 F Statistic 7,676.43 7,256.76 7,620.57 7,450.67 Observations 344,686 344,686 344,686 344,686

Table 3.A4: Robustness Checks: OLS Results of Linguistic Distance – Model 2 SOEP Sample

ASJP WALS TREE Coef/StdE Coef/StdE Coef/StdE

Male Subsample Linguistic distance −0.115 −0.223∗ −0.185∗ (0.085) (0.096) (0.083) LD × YSM 0.002 0.002 0.003 (0.004) (0.004) (0.004) Years since migration 0.016∗∗ 0.015∗∗ 0.015∗∗ (0.005) (0.005) (0.005)

Adjusted R2 0.302 0.307 0.308 F Statistic 13.55 14.18 13.32 Observations 2,703 2,703 2,703

Female Subsample Linguistic distance −0.085 −0.139† −0.114† (0.058) (0.077) (0.066) LD × YSM 0.000 −0.000 0.000 (0.002) (0.003) (0.002) Years since migration 0.020∗∗∗ 0.021∗∗∗ 0.020∗∗∗ (0.005) (0.005) (0.005)

Adjusted R2 0.448 0.449 0.449 F Statistic 30.74 31.52 31.00 Observations 3,100 3,100 3,100

Low-skilled Subsample Linguistic distance −0.101 −0.178∗∗ −0.158∗ (0.062) (0.062) (0.070) LD × YSM −0.000 −0.001 0.000 (0.002) (0.003) (0.002) Years since migration 0.021∗∗∗ 0.022∗∗∗ 0.020∗∗∗ (0.004) (0.004) (0.004)

Adjusted R2 0.334 0.337 0.338 F Statistic 24.68 25.13 26.07 Observations 5,067 5,067 5,067

High-skilled Subsample Linguistic distance −0.208∗∗ −0.150 −0.286∗∗∗ (0.064) (0.142) (0.069) LD × YSM 0.011∗∗ 0.008† 0.014∗∗∗ (0.004) (0.005) (0.004) Years since migration 0.007 0.005 0.011 (0.007) (0.008) (0.007)

Adjusted R2 0.294 0.279 0.313 F Statistic 5.11 4.94 5.83 Observations 736 736 736

Subsample excluding Turkish immigrants Linguistic distance −0.076 −0.131† −0.136∗ (0.053) (0.069) (0.061) LD × YSM −0.000 −0.001 0.001 (0.002) (0.002) (0.002) Years since migration 0.014∗∗ 0.015∗∗ 0.014∗∗ (0.005) (0.005) (0.005)

Adjusted R2 0.281 0.281 0.285 F Statistic 17.46 17.75 17.51 Observations 4,032 4,032 4,032

Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is defined on a scale of 0 to 3 such that higher values indicate a higher level of language proficiency. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 72 3.B Supplementary Appendix ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 013) 008 000) 008 003) 106 001) 031 001) 040 003) 097 000) 028 000) 003 000) 000) 014 018 410 003) 075 ...... 0 1 0 0 0 0 0 0 0 0 0 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 013) (0 008 000)014 (0 003) (0 031 108 040 001) (0 001) (0 097 000)028 003) (0 (0 000)018 (0 000) (0 014 409 0 124 002) (0 ...... 0 1 0 0 0 0 0 0 0 0 − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 015) (0 010 000)023 595 (0 003)039 (0 107 001) (0 001) (0 027 097 000)003) (0 (0 001 018 000)000) (0 000) (0 (0 014 086 405 0 002) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 023 015) (0 039 010 000)595 (0 003) (0 027 107 001) (0 001) (0 018 097 000)003) (0 (0 000) (0 000) (0 107 014 405 0 002) (0 ...... 0 0 0 0 0 0 0 0 0 0 − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 025 017) (0 000) (0 036 011 655 003) (0 001) (0 026 001) (0 102 000) (0 018 097 003) (0 005 000) (0 000) (0 000) (0 038 014 405 0 004) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 026 017) (0 000) (0 036 011 686 003) (0 001) (0 024 001) (0 105 000) (0 097 003) (0 018 000) (0 000) (0 103 014 404 0 003) (0 ...... 0 0 0 0 0 0 0 0 0 0 − − − − − 10% level. – Robust standard errors are reported in parentheses. – The dependent variable † ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 000)023 (0 016) (0 011 504 001) (0 037 003) (0 028 001) (0 106 000) (0 018 096 003) (0 004 000) (0 000) (0 000) (0 180 015 407 0 004) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 (0 5% level; − − − − − − ∗ OLS Results – Model 1 & 2 ACS Sample ASJP WALS TREE SCORE ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 1% level; 000)024 (0 016) (0 037 011 516 003) (0 001) (0 026 001) (0 106 000) (0 018 096 003) (0 000) (0 000) (0 241 014 406 0 004) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 ∗∗ (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE − − − − − 0.1% level; Table 3.B1: ∗∗∗ 2 YSM – × Region dummiesYear fixed effects yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes Genetic distance Migrant stock (% of population) Geographic distance (in 100 km) Constant Female Married Years of education LD Age at entry F StatisticObservations 514,874 16,051.0 514,874 14,965.9 15,351.8 514,874 14,489.8 514,874 16,102.4 514,874 15,076.1 514,874 15,706.5 514,874 14,881.2 514,874 Linguistic distance Adjusted R Years since migration Notes: – Significant at: is defined on a scale of 0 to 3 such that higher values indicate a higher level of language proficiency. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 73 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 087 002) 038) 496 000) 039) 008 000) 045 000) 231 001) 079 240 007) 007) 079 018 003) 000) 234 186 909 006) 035 038) 195 ...... 0 0 0 0 0 0 0 1 0 0 2 0 0 − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 007) (0 078 018 003)000) (0 (0 087 002) (0 344 200 890 004)035 (0 038)038)476 (0 (0 194 0 000) (0 039) (0 045 000)231 (0 001)080 245 (0 007) (0 ...... 0 0 0 0 2 0 0 0 0 0 0 1 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 076 233 007)007)081 (0 (0 025 003)000) (0 (0 021 003) (0 374 453 548 011)035 (0 041)041)126 (0 (0 194 0 000) (0 042) (0 016 001)045 (0 000)228 (0 001) (0 ...... 0 0 0 0 0 0 2 0 4 0 0 0 0 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 000)228 (0 001)240 (0 073 007)007)079 (0 (0 024 003)000) (0 (0 028 003) (0 533 398 484 011)036 (0 040)041)059 (0 (0 193 0 000) (0 041) (0 044 ...... 0 0 0 0 0 0 2 0 4 0 0 − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 003)027 000) (0 (0 064 002) (0 198 003 0 010)034 (0 044)092 044)664 (0 (0 191 0 000) (0 044) (0 011 000)045 (0 000)228 (0 001) (0 227 073 007)007)067 (0 (0 ...... 3 0 0 0 0 0 0 0 0 0 0 0 2 − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 007)234 007) (0 (0 066 003)026 000) (0 (0 068 002) (0 061 339 009)035 (0 043)023 044)593 (0 (0 190 0 000) (0 044) (0 044 000) (0 228 001) (0 070 ...... 0 0 0 0 0 0 2 3 0 0 0 0 − − − − − − 10% level. – Robust standard errors are reported in parentheses. – The dependent variable is † ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 227 001) (0 078 007)007) (0 (0 231 074 003)000) (0 (0 029 042 002) (0 602 719 017) (0 041) (0 195 0 039 700 041)283 (0 000) (0 042) (0 028 001)045 (0 000) (0 ...... 0 0 0 0 0 0 0 0 2 4 0 0 0 (0 5% level; − − − − − − ∗ ASJP WALS TREE SCORE ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 1% level; 047 002) (0 041) (0 184 000) (0 041) (0 044 000) (0 227 001) (0 074 007)007) (0 (0 237 072 003)000) (0 (0 028 035 516 194 0 017) (0 041) (0 036 604 ...... Ordered Logit Results – Model 1 & 2 ACS Sample ∗∗ 0 0 0 0 0 0 1 0 0 0 2 4 0 126,505.3 125,775.2 129,153.7 128,591.9 125,523.4 124,729.3 129,363.1 129,271.9 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE − − − − − 0.1% level; ∗∗∗ Table 3.B2: 2 2 YSM – χ × Geographic distance (in 100 km) Genetic distance Region dummiesYear fixed effects yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes Years since migration Threshold 2 Threshold 3 Wald Log-likelihoodObservations -13,780,358 -13,758,860 514,874 -13,841,570 -13,824,642 514,874 -13,795,170 -13,778,354 514,874 -13,771,129 -13,761,387 514,874 514,874 514,874 514,874 514,874 LD Age at entry Years of education Female Married Migrant stock (% of population) Linguistic distance Threshold 1 Pseudo-R Notes: – Significant at: defined on a scale of 0 to 3 such that higher values indicate a higher level of language proficiency. CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 74 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 066 067 002) 007 001) 007 009 009 000) 045 000) 045 000) 014 000) 016 000) 046 000) 048 001) 013 001) 015 001) 005 001) 004 001) 013 001) 017 000) 000) 000) 000) ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 012 012 000)001 (0 000)001 (0 002 002 000) (0 008 000) (0 008 000) (0 002 000) (0 003 000) (0 008 000) (0 009 000) (0 002 000) (0 003 000) (0 001 000) (0 001 000) (0 002 000) (0 003 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 057 058 002)006 (0 001)006 (0 008 008 000) (0 039 000) (0 039 000) (0 012 000) (0 014 000) (0 040 000) (0 042 001) (0 011 001) (0 013 001) (0 004 001) (0 003 000) (0 011 000) (0 015 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 021 021 001)002 (0 000)002 (0 003 003 000) (0 014 000) (0 014 000) (0 000) (0 004 005 000) (0 014 000) (0 015 000) (0 000) (0 004 005 000) (0 002 000) (0 001 000) (0 000) (0 004 005 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 204 003)007 (0 105 002)007 (0 000)009 (0 045 000)009 (0 045 000) (0 000) (0 000)014 (0 047 000)014 (0 047 001) (0 001) (0 001)014 (0 005 001)016 (0 005 001) (0 001) (0 000)009 (0 000)006 (0 000) (0 001) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 034 001) (0 018 000) (0 001 001 000)001 (0 000)001 (0 008 008 000) (0 000) (0 000)002 (0 000)002 (0 008 008 000) (0 000) (0 000)002 (0 000)003 (0 001 001 000) (0 000) (0 000)002 (0 000)001 (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − 10% level. – Robust standard errors are reported in parentheses. – The dependent † ASJP WALS TREE SCORE ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 003) (0 002) (0 175 090 006 006 000) (0 000) (0 008 008 000)038 (0 000)039 (0 000) (0 000) (0 012 012 001)040 (0 001)041 (0 001) (0 001) (0 012 013 000)005 (0 000)004 (0 000) (0 000) (0 008 005 000) (0 000) (0 ...... 5% level; 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ∗ − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 1% level; 001) (0 001) (0 063 002 032 002 000) (0 000) (0 003 000) (0 003 000) (0 014 014 000) (0 000) (0 004 000) (0 004 000) (0 014 015 000) (0 000) (0 004 000) (0 005 000) (0 002 001 000) (0 000) (0 003 000) (0 002 000) (0 Bad Not bad Well Very well Bad Not bad Well Very well Bad Not bad Well Very well Bad Not bad Well Very well ...... ∗∗ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE − − − − − − − − Ordered Logit Marginal Effects – Model 1 ACS Sample 0.1% level; ∗∗∗ YSM no no noYSM no no no no no no no no no no no no no Table 3.B3: × × Linguistic distance Years since migration Linguistic distance Years since migration Age at entry Age at entry Years of education Years of education Female Female Married Married Migrant stock (% of population) Migrant stock (% of population) Geographic distance (in 100 km) Geographic distance (in 100 km) Genetic distance Genetic distance LD Region dummiesYear fixed effects yes yes yes yes yes yes yesLD yesRegion dummiesYear fixed effects yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes Notes: – Significant at: variable is definedmean on of a the scale of covariates vector. 0 to 3 such that higher values indicate a higher level of language proficiency. – Marginal effects are reported at the CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 75 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 069 066 002) 007 001) 007 009 009 000) 045 000) 045 000) 014 000) 015 000) 045 000) 047 001) 013 001) 015 001) 005 001) 004 001) 013 001) 017 000) 000) 000) 000) ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 012 012 000)001 (0 000)001 (0 002 002 000) (0 008 000) (0 008 000) (0 003 000) (0 003 000) (0 008 000) (0 008 000) (0 002 000) (0 003 000) (0 001 000) (0 001 000) (0 002 000) (0 003 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 060 058 002)006 (0 001)006 (0 008 008 000) (0 039 000) (0 039 000) (0 012 000) (0 013 000) (0 039 000) (0 041 001) (0 011 001) (0 013 001) (0 005 001) (0 003 000) (0 011 000) (0 015 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 022 021 001)002 (0 000)002 (0 003 003 000) (0 014 000) (0 014 000) (0 000) (0 004 005 000) (0 014 000) (0 015 000) (0 000) (0 004 005 000) (0 002 000) (0 001 000) (0 000) (0 004 005 000) (0 000) (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 217 003)007 (0 116 002)007 (0 000)009 (0 045 000)009 (0 045 000) (0 000) (0 000)015 (0 045 000)015 (0 046 001) (0 001) (0 001)015 (0 006 001)016 (0 005 001) (0 001) (0 000)008 (0 000)004 (0 000) (0 001) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 036 001) (0 019 000) (0 001 001 000)001 (0 000)001 (0 007 007 000) (0 000) (0 000)003 (0 000)002 (0 007 007 000) (0 000) (0 000)002 (0 000)003 (0 001 001 000) (0 000) (0 000)001 (0 000)001 (0 000) (0 000) (0 ...... 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 − − − − − − − − − − 10% level. – Robust standard errors are reported in parentheses. – The dependent † ASJP WALS TREE SCORE ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 003) (0 002) (0 187 100 006 006 000) (0 000) (0 008 008 000)039 (0 000)039 (0 000) (0 000) (0 013 013 001)039 (0 001)040 (0 001) (0 001) (0 013 014 000)005 (0 000)004 (0 000) (0 000) (0 007 004 000) (0 000) (0 ...... 5% level; 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ∗ − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 1% level; 001) (0 001) (0 066 002 035 002 000) (0 000) (0 003 000) (0 003 000) (0 014 014 000) (0 000) (0 005 000) (0 005 000) (0 014 014 000) (0 000) (0 004 000) (0 005 000) (0 002 001 000) (0 000) (0 002 000) (0 001 000) (0 Bad Not bad Well Very well Bad Not bad Well Very well Bad Not bad Well Very well Bad Not bad Well Very well ...... ∗∗ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 (0 ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE ME/StdE − − − − − − − − Ordered Logit Marginal Effects – Model 2 ACS Sample 0.1% level; ∗∗∗ YSM yes yes yesYSM yes yes yes yes yes yes yes yes yes yes yes yes yes Table 3.B4: × × Linguistic distance Years since migration Linguistic distance Years since migration Age at entry Age at entry Years of education Years of education Female Female Married Married Migrant stock (% of population) Migrant stock (% of population) Geographic distance (in 100 km) Geographic distance (in 100 km) Genetic distance Genetic distance LD Region dummiesYear fixed effects yes yes yes yes yes yes yesLD yesRegion dummiesYear fixed effects yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes Notes: – Significant at: variable is definedmean on of a the scale of covariates vector. 0 to 3 such that higher values indicate a higher level of language proficiency. – Marginal effects are reported at the CHAPTER 3. LINGUISTIC BARRIERS IN LANGUAGE ACQUISITION 76

Table 3.B5: OLS Results – Model 1 & 2 SOEP Sample

ASJP WALS TREE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

Linguistic distance −0.079∗∗ −0.121∗ −0.160∗∗ −0.187∗∗ −0.117∗∗ −0.182∗∗∗ (0.030) (0.050) (0.051) (0.059) (0.037) (0.054) Years since migration 0.017∗∗∗ 0.017∗∗∗ 0.017∗∗∗ 0.017∗∗∗ 0.016∗∗∗ 0.017∗∗∗ (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) LD × YSM – 0.002 – 0.002 – 0.003 (0.002) (0.002) (0.002) Age at entry −0.018∗∗∗ −0.018∗∗∗ −0.020∗∗∗ −0.020∗∗∗ −0.018∗∗∗ −0.019∗∗∗ (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Years of education 0.103∗∗∗ 0.103∗∗∗ 0.102∗∗∗ 0.101∗∗∗ 0.100∗∗∗ 0.099∗∗∗ (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) Female −0.081† −0.080† −0.076† −0.076† −0.081† −0.079† (0.047) (0.047) (0.046) (0.046) (0.047) (0.047) Married −0.179∗∗ −0.178∗∗ −0.198∗∗∗ −0.196∗∗∗ −0.184∗∗ −0.182∗∗ (0.059) (0.059) (0.057) (0.057) (0.058) (0.059) Children in the HH. (Ref.= 0) One child −0.035 −0.035 −0.036 −0.040 −0.029 −0.030 (0.056) (0.055) (0.055) (0.055) (0.055) (0.055) Two children −0.036 −0.037 −0.030 −0.031 −0.038 −0.039 (0.065) (0.065) (0.065) (0.065) (0.065) (0.065) Three or more children −0.157† −0.154† −0.173∗ −0.171∗ −0.160∗ −0.156† (0.080) (0.080) (0.080) (0.080) (0.081) (0.081) Native German partner/spouse 0.297∗∗∗ 0.299∗∗∗ 0.312∗∗∗ 0.314∗∗∗ 0.286∗∗∗ 0.293∗∗∗ (0.069) (0.069) (0.066) (0.067) (0.068) (0.068) Family abroad −0.073 −0.070 −0.051 −0.049 −0.075 −0.073 (0.064) (0.064) (0.061) (0.061) (0.063) (0.063) Proficiency home language 0.279∗∗∗ 0.279∗∗∗ 0.294∗∗∗ 0.293∗∗∗ 0.277∗∗∗ 0.278∗∗∗ (0.059) (0.059) (0.058) (0.058) (0.058) (0.059) Desired stay (years) 0.004 0.005 0.003 0.003 0.004 0.004 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Migrant stock (% of population) −0.208∗∗∗ −0.210∗∗∗ −0.155∗∗ −0.158∗∗∗ −0.202∗∗∗ −0.209∗∗∗ (0.053) (0.052) (0.048) (0.047) (0.050) (0.049) Neighboring country 0.286∗∗ 0.283∗∗ 0.238∗ 0.237∗ 0.266∗ 0.253∗ (0.107) (0.107) (0.108) (0.108) (0.107) (0.107) Geographic distance (in 100 km) 0.002 0.002 0.002 0.002 −0.000 −0.001 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Genetic distance 0.041∗ 0.041∗∗ 0.042∗∗ 0.043∗∗ 0.048∗∗ 0.050∗∗ (0.016) (0.016) (0.015) (0.015) (0.016) (0.016) Constant 0.352 0.357 0.309 0.337 0.432 0.444 (0.286) (0.284) (0.273) (0.276) (0.275) (0.273) Region dummies yes yes yes yes yes yes Year fixed effects yes yes yes yes yes yes

Adjusted R2 0.373 0.373 0.375 0.375 0.376 0.377 F Statistic 34.55 34.41 35.80 34.69 34.59 34.59 Observations 5,803 5,803 5,803 5,803 5,803 5,803

Table 3.B6: Ordered Logit Results – Model 1 & 2 SOEP Sample

ASJP WALS TREE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

Linguistic distance −0.210∗ −0.331∗ −0.478∗∗∗ −0.546∗∗∗ −0.335∗∗ −0.525∗∗ (0.093) (0.165) (0.139) (0.162) (0.118) (0.174) Years since migration 0.043∗∗∗ 0.043∗∗∗ 0.043∗∗∗ 0.043∗∗∗ 0.040∗∗∗ 0.041∗∗∗ (0.011) (0.011) (0.010) (0.010) (0.010) (0.010) LD × YSM – 0.006 – 0.004 – 0.008 (0.007) (0.006) (0.007) Age at entry −0.047∗∗∗ −0.048∗∗∗ −0.056∗∗∗ −0.056∗∗∗ −0.050∗∗∗ −0.051∗∗∗ (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) Years of education 0.271∗∗∗ 0.270∗∗∗ 0.267∗∗∗ 0.266∗∗∗ 0.264∗∗∗ 0.262∗∗∗ (0.029) (0.029) (0.028) (0.028) (0.029) (0.028) Female −0.197 −0.197 −0.179 −0.180 −0.199 −0.197 (0.124) (0.124) (0.121) (0.120) (0.124) (0.124) Married −0.443∗∗ −0.444∗∗ −0.495∗∗∗ −0.490∗∗∗ −0.454∗∗ −0.454∗∗ (0.154) (0.153) (0.146) (0.147) (0.151) (0.151) Children in the HH. (Ref.= 0) One child −0.096 −0.098 −0.095 −0.104 −0.085 −0.089 (0.149) (0.149) (0.147) (0.146) (0.147) (0.147) Two children −0.125 −0.128 −0.109 −0.112 −0.133 −0.137 (0.173) (0.173) (0.174) (0.174) (0.173) (0.173) Three or more children −0.422† −0.413† −0.471∗ −0.467∗ −0.432∗ −0.420† (0.217) (0.217) (0.218) (0.218) (0.218) (0.218) Native German partner/spouse 0.765∗∗∗ 0.773∗∗∗ 0.807∗∗∗ 0.815∗∗∗ 0.735∗∗∗ 0.753∗∗∗ (0.189) (0.189) (0.183) (0.184) (0.189) (0.187) Family abroad −0.180 −0.174 −0.120 −0.117 −0.191 −0.190 (0.171) (0.172) (0.163) (0.163) (0.171) (0.172) Proficiency home language 0.730∗∗∗ 0.731∗∗∗ 0.780∗∗∗ 0.775∗∗∗ 0.730∗∗∗ 0.734∗∗∗ (0.160) (0.160) (0.159) (0.159) (0.159) (0.160) Desired stay (years) 0.010 0.011 0.005 0.005 0.010 0.010 (0.010) (0.010) (0.009) (0.009) (0.009) (0.009) Migrant stock (% of population) −0.513∗∗∗ −0.518∗∗∗ −0.371∗∗ −0.379∗∗ −0.503∗∗∗ −0.522∗∗∗ (0.151) (0.149) (0.131) (0.128) (0.141) (0.136) Neighboring country 0.750∗ 0.743∗ 0.601∗ 0.600∗ 0.685∗ 0.647∗ (0.302) (0.301) (0.300) (0.300) (0.305) (0.300) Geographic distance (in 100 km) 0.009 0.009 0.008 0.007 0.002 0.001 (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) Genetic distance 0.102∗ 0.103∗ 0.110∗∗ 0.111∗∗ 0.124∗∗ 0.129∗∗ (0.046) (0.046) (0.041) (0.041) (0.047) (0.046) Region dummies yes yes yes yes yes yes Year fixed effects yes yes yes yes yes yes

Threshold 1 0.763 0.745 0.880 0.791 0.562 0.502 (0.794) (0.791) (0.737) (0.753) (0.761) (0.760) Threshold 2 2.893∗∗∗ 2.879∗∗∗ 3.018∗∗∗ 2.933∗∗∗ 2.693∗∗∗ 2.640∗∗∗ (0.804) (0.800) (0.746) (0.762) (0.770) (0.769) Threshold 3 5.270∗∗∗ 5.254∗∗∗ 5.402∗∗∗ 5.315∗∗∗ 5.088∗∗∗ 5.034∗∗∗ (0.806) (0.803) (0.753) (0.767) (0.775) (0.775)

Pseudo-R2 0.180 0.180 0.182 0.182 0.182 0.183 Wald χ2 572.3 584.4 577.0 576.5 569.7 584.5 Log-likelihood -22,897,234 -22,887,498 -22,830,347 -22,824,832 -22,833,466 -22,812,425 Observations 5,803 5,803 5,803 5,803 5,803 5,803

Chapter 4

Language and Cultural Barriers in International Trade and Investment∗

4.1 Introduction

In a number of studies, economists have shown the importance of trust and information in overcoming informal barriers to international factor movements. For example, Guiso et al. (2009) find that lower bilateral trust leads to less trade between countries and less cross- border portfolio investments. Informal barriers may arise because of weak international legal institutions or inadequate information of opportunities for advantageous economic exchange (Rauch and Trindade, 2002). In particular, crossing cultural and language boundaries, as is often the case in international factor movements, reduces the level of trust between economic agents and makes obtaining information more costly by raising hurdles in communication. Both economic theory and empirical evidence suggest that sharing an official or speaking a common language provide an important stimulus in international economic exchange. From international trade, for instance, it is known that having a common language increases trade flows by about 40%.1 However, previous empirical literature leaves open the question whether differences between countries’ native languages affect international factor mobility above and beyond the effect of sharing a common language. Furthermore, the possibility

∗A revised version of this chapter is available from the author. The author is grateful to Thomas K. Bauer, Julia Bredtmann, Ingo Isphording, Christoph M. Schmidt, the participants of the Economics Seminar at the University of Wollongong, the Brown Bag Seminar at the Melbourne Institute of Applied Economic and Social Research, the 16th Workshop in International Economics, the IEA World Congress 2014, and the CEPR Conference on: Economic Inequality, Labor Markets and International Trade for helpful comments and suggestions. I am also very thankful to Andrew K. Rose for providing large parts of the datasets. 1In a meta-analysis based on 701 language effects collected from 81 academic articles, Egger and Lassmann (2012) find that a common language increases bilateral trade flows by 44% on average. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 81

that economic agents might use a non-native language, a lingua franca, to communicate is disregarded in most empirical work. Language and cultural barriers are also possible candidates to illuminate the “home bias puzzle” in cross-border portfolio investment. Unlike trade in goods, assets are “weightless”. Therefore, there are no physical transportation costs, which might explain investors’ preferences for their own country. Moreover, standard finance theory based on portfolio choice models suggests diversifying investment risks by holding a variety of assets from countries whose business cycles are uncorrelated with each other (Portes and Rey, 2005). The international financial integration over the last thirty years, which was fostered by, e.g., capital account liberalizations, electronic trading, increasing exchanges of information across borders, and falling transaction costs, should reduce the disproportionate high share of national assets in portfolios toward a higher share of foreign assets (Coeurdacier and Rey, 2013; Lane and Milesi-Ferretti, 2003). However, despite better financial integration, the observable market segmentation in country portfolios contradicts the assumed view of frictionless global financial markets. Rather, it points to the impact of informal barriers such as information asymmetries and cultural differences in international capital flows (Portes and Rey, 2005).2 Knowledge about these barriers is important in understanding the “home bias puzzle” in investment decisions. However, empirical evidence on the effect of language and cultural differences on investors’ preferences to invest in their home country (or in this regard close countries) is scarce. In this paper, I evaluate the extent to which language and cultural barriers affect different types of international factor movements, i.e., international trade flows, cross- holdings of assets, and consolidated international banking claims by addressing three major issues. First, I investigate the effect of differences between countries’ native languages on the three types of bilateral factor movements using a novel measure of linguistic distance. This measure allows for directly testing whether and to what extend dissimilarities between countries’ native languages affect international factor mobility. Next, I examine the impact of cultural differences on bilateral economic exchange by using a measure of genetic distance as proposed and applied by Guiso et al. (2009). My particular interest lies in a comparison of the effects of genetic and linguistic distance, which might vary substantially over the aforementioned factor movements. Finally, I follow the idea that economic agents might use English, the lingua franca in international business, for communication. In doing so, I examine how the results of differences between countries’ native languages change

2This view is supported by Coeurdacier and Rey (2013), who state that in 2007, US investors still held more than 80% of the US equities compared to 92% in 1989 documented by French and Poterba (1991). In both cases, this is a much higher proportion than the share of US equities in the world market portfolio. For literature surveys on the phenomenon of home bias, see Lewis (1999) and more recently Coeurdacier and Rey (2013). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 82

when including a measure for countries’ linguistic distance toward English in the empirical models. The impact of language and cultural differences is likely to be especially important in transactions where individuals have to coordinate with each other. This holds true for less transparent and standardized transactions in markets with non-trustworthy institutions and a difficult bureaucratic and commercial environment. Cross-border portfolio investments and consolidated international banking claims are relatively standardized transactions, generally taking place between financial intermediaries, e.g., investment banks, pension funds, mutual funds, etc. In contrast to these transactions, international trades of goods are less organized exchanges between mostly heterogeneous firms. Differences between native languages or insufficient proficiencies in English, the lingua franca in international business, induce communication barriers affecting both bilateral trade as well as international investment flows. Furthermore, a greater cultural distance could decrease the likelihood of finding an adequate trading partner due to larger differences in the supplied and demanded products. The opposite effect, however, might exist for international investments. A greater cultural distance between two countries should be associated with a lower correlation of their stock markets and thereby increase the diversification potential of the respective assets for the portfolios of foreign investors. To test these hypotheses, the empirical analysis is based on three different data sets. The first data set comprises international bilateral trade flows from 1950 to 2006 between 180 exporter countries and 187 importer countries. The second database covers cross- holdings of asset stocks between 62 source and 173 host countries for the years 2001 till 2003. To check the sensitivity of this relative small sample, I further use data of consolidated international banking claims covering 24 source and 182 host countries between 1983 and 2006. In the empirical analysis, I apply a gravity model as first proposed by Tinbergen (1962) and since then applied in numerous empirical studies on factor mobility.3 These models explain the intensity of bilateral relations by some elements of mass, e.g., economic size (often approximated by GDP) and geographic distance. I follow this approach, but measure distance not only in the geographical dimension, but further incorporate measures for cultural and language dissimilarities as described above. This study attempts to quantify the effect of informal barriers on international factor movements by focusing on language and cultural barriers. It directly relates to two strands in the recent empirical literature which seek to explain the determinants of bilateral trade

3For recent applications in international trade, see, e.g., Mélitz (2008), Head et al. (2010) and Aviat and Coeurdacier (2007) for an application in bilateral asset holdings. An overview of the estimation and interpretation of gravity models can be found in Head and Mayer (2013). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 83 and international portfolio investment, respectively. The first strand emphasizes the importance of language and cultural barriers in international trade. To incorporate language-related barriers into a gravity model, common empirical practice is to use an indicator variable that covers countries which share the same official language (see, e.g., Anderson and van Wincoop, 2004). While most studies follow this approach, Mélitz (2008) goes beyond the consideration of official languages and develops two different measures. The first measure reflects the probability that two randomly chosen individuals from either country share a common language. The second measure indicates whether the same language is spoken by at least 20% of the populations in both countries. These measures share the shortcoming that they only consider countries that share the same language, but do not account for the dissimilarity between the countries’ languages. To the best of my knowledge, there are only four studies that take differences between native languages into account. Hutchinson (2005) relies on the measure by Chiswick and Miller (1999) and his approach is therefore restricted to distances toward English. In contrast, Lohmann (2011) uses an index of shared linguistic features within language pairs for about 200 countries. Two recent studies by Isphording and Otten (2013) and Mélitz and Toubal (2012) directly focus on the differences between native languages using the same measure of linguistic distance applied in this paper. Certainly, a range of dominant regional languages (e.g., English in the Western countries, Russian in Eastern Europe, French in Africa, and Spanish in Latin America) play a major role in international trade (Isphording and Otten, 2013). Especially, proficiency in English as the leading language in international business is essential for most trade relations. However, the only study I am aware of that addresses issues of the ability to communicate in English is the one by Ku and Zussman (2010). To measure the level of English proficiency in more than 100 countries, they use a country’s mean score in the “Test of English as a Foreign Language” (TOEFL). This paper also relates to a line of research that studies the role of cultural differences in trade flows. Felbermayr and Toubal (2010) draw on bilateral score data from the Eurovision Song Contest, a popular pan-European television show, to construct a measure of cultural proximity and employ this measure into the framework of a gravity model of trade. The studies by Guiso et al. (2009) and Giuliano et al. (2014) use a measure of genetic distance to reveal the effect of cultural differences on trade. Specifically, Guiso et al. (2009) examine the relationship between bilateral trust and bilateral trade, using genetic distance as an instrument for trust. In contrast, Giuliano et al. (2014) directly analyze the effect of genetic distance on trade. They find that the negative effect of genetic distance on bilateral trade flows decreases substantially, or even disappears, once they CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 84

control for further geographical barriers other than the simple geographic distance between capital cities. The second strand in the literature emphasizes the role of language and cultural barriers in cross-border portfolio investment. One set of studies focuses on the impact of information costs raising informational asymmetries and transaction costs between domestic and foreign investors (Ahearne et al., 2004; Daude and Fratzscher, 2008; Portes and Rey, 2005, amongst others). In this regard, e.g., Tesar and Werner (1995), Grinblatt and Keloharju (2001), Chan et al. (2005) and more recently Lane and Milesi-Ferretti (2008), Beugelsdijk and Frijns (2010), and Aggarwal et al. (2012) explicitly underline the influence of language barriers on investment decisions. For example, Grinblatt and Keloharju (2001) find that investors in Finland are more likely to trade stocks of firms that share the investor’s language and cultural background. To incorporate language-related barriers, these studies again use a variable that indicates whether two countries or investors share the same official language. As mentioned above, this approach has the shortcoming that the estimated effect only relies on the difference of sharing or not sharing a common language, but does not account for the dissimilarity between the languages. The only study that takes differences between languages into account is the one by Guiso et al. (2009), who use a measure for linguistic common roots created by Fearon and Laitin (2003).4 A second set of studies examines the relationship between “familiarity effects”, i.e., cultural proximity and cross-border investments (e.g., Aggarwal et al., 2012; Beugelsdijk and Frijns, 2010; Chan et al., 2005; Grinblatt and Keloharju, 2001). The only study that uses genetic distances to uncover the effect of cultural differences on cross-border portfolio investment is again the one by Guiso et al. (2009). This paper makes three important contributions to the existing literature. First, I provide some first evidence on the effect of language and cultural differences on international trade flows. Using a novel measure of linguistic distance and incorporating genetic distance as a proxy for cultural differences, this is the first study that allows disentangling the impact of language and cultural dissimilarities on trade flows. Given the established correlation between geographic distance and language and cultural differences, the paper further sheds light on how an appropriate modeling of linguistic and genetic distance helps to solve the “distance puzzle” in international trade. Second, I extend the finance literature by providing first evidence on the effect of language and cultural dissimilarities on cross-border portfolio investment and international banking claims. Finally, this is the first comprehensive analysis of the effect of English proficiency on international factor movements, an impact factor so far neglected in the literature. The only notably exception

4For an overview of the approaches employed in the economic literature to measure the dissimilarity between languages, see Isphording and Otten (2011). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 85

I am aware of is the study by Ku and Zussman (2010). In contrast to their investigation, however, I use a measure of linguistic distance to model English proficiency. Thereby, I am able to extend the underlying sample to almost all of the world’s countries. Furthermore, employing linguistic distance instead of TOEFL scores has the advantage of using an unbiased proxy for English proficiency. In addition, I extend their analysis by further examining financial flows, i.e., cross-holdings of asset stocks and international banking claims, providing first evidence in this line of research. The results show that linguistic and genetic distance have varying effects on the examined factor movements. While controlling for a number of other possible determinants, I find strong evidence that a higher linguistic distance between two countries reduces cross-border trade and investment holdings between these countries. The size of this effect is substantial. For example, an increase in linguistic distance by one standard deviation decreases bilateral exports by 7.4%. This finding suggests that the further away a country’s language is from the rest of the world’s languages, the higher the information costs for trading partners and foreign investors and thus the smaller bilateral factor flows. In contrast to these findings, the results for genetic distance are mixed. Cultural differences have significant effects only on bilateral trade, while no significant effects are found for international investments. In particular, the results show that a greater genetic distance between two countries reduces the volume of bilateral trade and, if anything, increases the level of cross-border investments. When including the country’s linguistic distance toward English in the models, the estimates indicate a significant negative impact of a higher linguistic distance toward English on international factor mobility. Taken together, these findings are fully in line with the theoretical arguments outlined above and provide supportive evidence that language differences contribute to higher informational frictions across countries, thereby reducing bilateral trade and cross-border investment flows, respectively. As hypothesized above, the results for cultural differences are asymmetric for international trade and investment holdings. The remainder of the paper is structured as follows. The next section briefly outlines the theoretical framework and specifies a gravity model including country-year fixed effects. Moreover, I formalize the identification strategy underlying the estimations of English proficiency on international factor mobility. Section 4.3 describes the bilateral trade and financial data, the measures of linguistic and genetic distance, and the extensive set of control variables. The detailed specifications used in the estimations of the different models and the respective empirical results are presented and discussed in Section 4.4. The concluding section summarizes and interprets the main results. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 86 4.2 Empirical Model

4.2.1 Theoretical Background: The Structural Gravity Model

In estimating the effects of language and cultural barriers on international factor mobility, I follow common empirical practice and build on the gravity model of international trade, which has become the standard model for estimating the determinants of bilateral factor flows.5 The gravity model can be derived from a number of theoretical frameworks, relying on different modeling assumptions (see Anderson and van Wincoop, 2003, 2004; Eaton and Kortum, 2002). The first theoretical foundations of the gravity model have been laid by Anderson (1979) and Bergstrand (1985). The following equations build on this work and are in particular based on the canonical model by Anderson and van Wincoop (2003, 2004). In this setting, bilateral trade is given by

k k k !1−σk k Ej Yi tij Xij = k k k , (4.1) Y Pj Πi

k where Xij is deﬁned as nominal exports from country i into country j in product class k. k k Ej denotes the expenditure in country j for product class k, Yi is the value of production k k in country i for product class k, and Y is world output in sector k. tij are the trade related

costs between i and j, the parameter σk is the elasticity of substitution among brands, k k and Pj and Πi denote inward and outward multilateral resistance, respectively, which are k k 6 functions of trade barriers and the entire set of Ej ,Yi . The main contribution of the “structural” gravity equation is that bilateral trade is determined by relative trade barriers, k namely the bilateral trade barrier tij divided by multilateral resistance, which depend on the average trade barriers the exporter and importer face with all their trading partners (Anderson and van Wincoop, 2004). Assuming (i) aggregation to one-sector economies7, thereby omitting the superscripts

k, (ii) expenditures to be equal to the value of production (Ei = Yi), and (iii) symmetric

bilateral trade costs (tij = tji), which results in Pi = Πi, Eq. (4.1) can be rewritten as:

5Gravity equations have been applied not only to estimate the determinants of trade flows, see Anderson (2011) and Bergstrand and Egger (2011) for recent surveys, but also to migration, equity, and FDI flows. For a recent study on migration flows, see, e.g., Adsera and Pytlikova (2012) and Martin and Rey (2004) for the first application to international portfolio investment. Brenton et al. (1999) apply the gravity equation to FDI flows. 6See Anderson and van Wincoop (2004) for the derivation and a general discussion of this equation. Anderson and Yotov (2010b) refer to the second ratio of the right-hand side product as a measure of “constructed home bias”, which is a useful quantification with respect to the exploration of possible explanations of the “home bias puzzle” in research on international portfolio investment. 7Anderson and Yotov (2010a,b) show that the bias resulting from aggregation is large. However, for the sake of simplicity, I follow the conventional approach and use data on aggregated goods. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 87

!1−σ Yit Yjt tijt Xijt = W . (4.2) Yt Pjt Πit Since most of the recent gravity papers rely on panel data, it has to be assumed that the cross-section equation derived from Anderson and van Wincoop holds in all time units of the panel (Hornok, 2012). I follow this literature and denote time-varying components in

Eq. (4.2) with the subscript t. Let Yit, Yjt represent nominal GDP in country i and j in W year t and let Yt deﬁne world GDP in year t. The ﬁnal step in the derivation of the “structural” gravity equation is to model the trade costs tijt. Here, Anderson and van Wincoop follow the literature and assume that trade costs are a log-linear function of a set of observable proxies for various types of trade costs along with some direct measurable components of trade barriers:

M Y m γm tijt = zijt . (4.3) m=1 m Common examples for zijt from the literature are geographic distance between the capitals of country i and j, adjacency, and membership in a customs or monetary union. Substituting Eq. (4.3) into Eq. (4.2) and log-linearizing the resulting equation yields

M W X m ln Xijt = ln Yit + ln Yjt − ln Yt + λm ln zijt − (1 − σ)[ln (Πit) + ln (Pjt)] , (4.4) m=1

W where λm = (1 − σ) γm. Xijt,Yit,Yjt, and Yt are observable, whereas the multilateral resistance terms Πit and Pjt are unobservable. The “structural” gravity equation can be estimated using different approaches. An- derson and van Wincoop (2003) employ nonlinear least squares (NLS) after solving for the multilateral resistance terms, which leads to unbiased estimators if the error term is uncorrelated with the regressors. However, most of the subsequent studies control for the unobservable multilateral resistance terms and the observable production variables using country-specific fixed effects and apply ordinary least squares (OLS) (see, e.g., Anderson, 2011; Eaton and Kortum, 2002). The following empirical analysis uses mainly directional (i.e., source and destination) country-year fixed effects to control not only for the multilateral resistance terms, but also for all country-specific time-varying factors. The gravity equation to be estimated becomes:

I J 0 X Ex X Im ln Xijt = Zijt β + δit cit + θjt cjt + ijt , (4.5) i=2 j=2 CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 88

P Ex P Im where δitcit and θjtcjt denote full sets of exporter-year and importer-year dummy variables, respectively, while Zijt is a vector of observable proxy variables for bilateral trade costs. The error term ijt consists of two components. One component is a random error, which simply reﬂects measurement error associated with export ﬂows. The second component, however, comprises all unobservable bilateral factors that are correlated with exports between country i and j and are not appropriately controlled for in the estimation.8 In case these factors are correlated with the included regressors, the estimates will be biased due to omitted variables.9

4.2.2 Empirical Strategy

In the following analysis, different types of factor movements are examined by estimating the gravity model in Eq. (4.5) using three different dependent variables. A first set of equations is estimated employing the logarithm of bilateral annual exports. Following previous literature, export volume instead of total trade volume is used, because the direction of trade volumes matters for the parameter estimation. A second set of models is estimated by using the logarithm of cross-holdings of asset stocks as the dependent variable, while I investigate the logarithm of consolidated international banking claims in a third set of models. The derived gravity model in Eq. (4.5), including the full sets of directional country- year fixed effects, can be estimated using OLS.10 Due to the panel nature of the data, leading to a repeated presence of country-pairs, all bilateral observations of country i and country j may have common disturbance elements. For this reason, the standard errors

8To avoid bias accruing from the correlation between included determinants of bilateral trade and the unobservable bilateral components of the error term, Baltagi et al. (2003) and Baldwin and Taglioni (2006) suggest the additional inclusion of time-invariant country-pair fixed effects in the estimation model. The downside of this approach, however, is that time-invariant parameters would no longer be identifiable as the fixed effects net out all time-constant variation. Since the regressors of interest in this analysis are time-invariant, I do not follow this approach. 9Another source of bias leading to non-orthogonality of the error term are endogenous regressors resulting from reverse causality. Since the causal direction of the variables of interest in the following analysis is considered as unambiguous, issues regarding this source of endogeneity are not further discussed. For an overview of how the literature has addressed these problems, see Anderson and van Wincoop (2004). 10The standard OLS procedure has been criticized in the recent literature, because of dropping country- pairs with zero trade volumes (see, e.g., Helpman et al., 2008; Santos Silva and Tenreyro, 2006). There are a number of possible methods to incorporate zeros in the gravity model. For an overview, see Anderson (2011) and Head and Mayer (2013). However, I do not use these methods in the following analysis. Instead, I follow the conventional method described in Section 4.2.1 and take the natural logarithm of annual exports as the dependent variable, thereby dropping observations where exports are recorded as zero. To verify that the results obtained in the analysis are robust to this specification, I conduct a robustness check in which ln (Xijt + 1) serves as dependent variable. The robustness analysis yields similar results and is available upon request. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 89 are adjusted to be robust to clustering of residuals by country-pairs.11 To examine the impact of language and cultural barriers on international factor movements, the analysis proceeds in three steps. While the first two steps rely on the gravity model derived in Eq. (4.5), country-pair specific random effects instead of the two sets of country-year fixed effects are applied in the third step.

Identifying Language and Cultural Barriers

In the first step, interest is solely directed toward identifying the effect of bilateral language and cultural differences on international factor mobility. As the analysis of these factors lies outside of the standard gravity model, Eq. (4.3) is augmented by variables measuring bilateral language and cultural differences. Since it is difficult to incorporate these factors in an empirical model, I use observable measures for linguistic and genetic distance, respectively, as proxy variables. The bilateral trade costs12 are then given by

h 0 i tijt = exp γ LDij + κ GDij + Zijt β . (4.6) The set of dyadic variables in Eq. (4.6) can be divided into two groups. The first group includes the variables of primary interest, LDij and GDij, where the former denotes the linguistic and the latter the genetic distance between country i and j. The second group, represented by the vector Zijt, comprises a set of observable control variables typically used in gravity regressions. The controls include (unless otherwise specified) the logarithm of the geographic distance between i and j, an indicator variable that equals one if two countries share a border, an indicator that equals one if both countries share the same legal origin, and an indicator that equals one if two countries were once, or are still, in a colonial relationship. Furthermore, two indicator variables are added, where the first equals one if both countries belong to a common regional trade arrangement (RTA) in year t and the second equals one if both countries are GATT/WTO participants in year t.

The final control variable included in the vector Zijt is an indicator that equals one if the two countries share a common currency. The essential idea is that the variation in the country-pair dimension of the bilateral linguistic and genetic distance identifies the effect of language and cultural dissimilarities on factor mobility. However, simultaneity between the variables of interest and the dependent variable or omitted variables that are correlated with both the dependent variable and one

11To check the robustness of the results, I also reestimate all models using two-way clustering by country and year employing the method of Cameron et al. (2011). The results are similar and available upon request. 12In conformity with the standard gravity model, the term “trade costs” is used to refer to the costs of bilateral factor movements, although diﬀerent factor movements are considered (i.e., trade, portfolio investments, and banking claims). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 90

of the variables of interest would cause endogeneity and bias the parameter estimates. Since linguistic and genetic distance are used as proxies for language and cultural differences, reverse causality is not considered as a source of endogeneity, as it is unlikely that these measures are influenced by international factor movements.13 The main concern in this analysis is the possibility of omitted variables. However, to bias the estimates, a possible confounder has to be correlated with the measure of linguistic or genetic distance. I argue that the existence of such a variable is unlikely.14 Nonetheless, to address this possible bias, the empirical strategy is to include a number of variables in Eq. (4.6) that affect factor mobility, thereby controlling for as many theoretical causes of bilateral factor movements as possible. Since the variables of interest are time-invariant, it is not possible to include country-pair fixed effects in Eq. (4.6). The primary interest is how an increase in the linguistic and genetic distance affects bilateral factor mobility. Therefore, the baseline specifications are as follows: First, I

estimate Eq. (4.5) using Eq. (4.6) without the genetic distance variable GDij to analyze the eﬀect of linguistic distance on bilateral factor movements. Second, I estimate the same speciﬁcation, but use genetic distance rather than linguistic distance as a regressor. Finally, a horse race between linguistic and genetic distance is conducted by including both variables in Eq. (4.5).

Identifying the Eﬀect of English Proﬁciency

In the second and third step, the analysis turns to the investigation of the impact of English proficiency on international factor mobility. To analyze this effect, Eq. (4.6) has to be augmented by variables that capture English proficiency. Since it is difficult to measure a country’s average English skills, each country’s linguistic distance toward English (henceforth LDE) is used as a proxy for the average English ability of the population living in the respective country.15 However, in the presence of directional country-year fixed effects, as in Eq. (4.5), regressors that vary in the same dimensions as these fixed effects cannot be identified, since the fixed effects net out all variation from the data except for the bilateral variation. This precludes, for example, the estimation of separate exporter and importer country-specific effects. A solution to get around this problem is to create a new variable that varies bilaterally and include it in the estimation. One possibility to transform two monadic variables describing country i and j, respectively,

13For a description of these measures and the underlying data, see Section 4.3.1. 14Spolaore and Wacziarg (2009, 2012) argue in the same direction by pointing out that genetic distance acts as an exogenous regressor in their models of development diffusion and technology adoption, respectively. 15Isphording and Otten (2011) provide empirical evidence that the linguistic distance between individuals’ mother tongue and their second language is a strong predictor of their language proficiency. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 91 into a dyadic variable is to multiply both variables, e.g., one can create a dyadic area variable by multiplying the area of country i times the area of country j. The created bilateral-varying variable is identifiable even when directional country fixed effects are included in the estimation model. However, the identification solely rests on functional form assumptions and is therefore a sort of constructed identification. A second possibility is to use the logarithm of the average, sum, or difference of country-specific variables. In these cases the created bilateral variable is identified because of the change in the functional form. Without taking the logarithm, the new variable is only a linear transformation of the monadic variables and cannot be identified (Head and Mayer, 2013). Since the identification assumptions that have to be made to generate the dyadic variables are strong, there are reasons to be cautious in accepting the parameter estimates based on these methods. In the next steps of the analysis, I therefore do not use the described approaches to examine the effect of a country’s LDE. In the second step, I investigate the effect of a country’s LDE in relation to the bilateral linguistic distance on international factor movements. In order to do so, notice that there are two types of countries: (i) countries with a higher LDE than the respective bilateral linguistic distance and (ii) countries with a lower LDE than the respective bilateral linguistic distance. In the following, two different identification assumptions are considered to construct separate indicator variables. Formally, the first indicator variable is defined as:

  (1) 1 if (LDEi > LDij) ∧ (LDEj > LDij) LDEij = , (4.7) 0 otherwise where LDEi and LDEj denote the linguistic distance toward English of country i and j, respectively. Under the first identification assumption, the indicator variable captures the effect of both trading partners facing a higher LDE than their bilateral linguistic distance. The second indicator variable is defined as below:

  (2) 1 if (LDEi > LDij) ∨ (LDEj > LDij) LDEij = , (4.8) 0 otherwise

(2) where LDEij captures the effect of at least one trading partner facing a higher LDE than their bilateral linguistic distance. The identification assumptions above determine the exact formulation of the indicator variables. However, it is reasonable to assume that bilateral factor mobility is unaffected by the LDE if both countries in a pair share a common language, i.e., if the bilateral linguistic distance is zero. Therefore, both indicator variables are multiplied with a trigger CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 92 variable denoting one if two countries in a pair do not share a common language and zero otherwise. The aim of using these indicator variables is to estimate the effect of a high LDE in relation to the bilateral linguistic distance on bilateral factor mobility while controlling for the bilateral linguistic distance and each country’s LDE by employing directional country fixed effects. The essential idea is that if the LDE is higher than the bilateral linguistic distance, the importance of the bilateral linguistic distance in the communication process increases, because the hurdles in the acquisition of English language skills are higher compared to the hurdles in the acquisition of the trading partner’s native language. Therefore, the probability that the trading partners communicate in one of the native languages increases. The estimation of specifications including these indicator variables implicitly assume that a country’s LDE and the bilateral linguistic distance affect bilateral factor movements above and beyond the inclusion of the direct effects. To test these hypotheses, Eq. (4.6) is extended by separately adding an interaction term of the trigger variable and each indicator variable. The resulting specifications are as follows: First, I estimate Eq. (4.5) (1) using Eq. (4.6) including the interaction term TRij × LDEij . Second, I estimate the (2) same specification, but use TRij × LDEij as a regressor.

Identiﬁability and the Random Eﬀects Estimator

In the third step of the analysis, I attempt to explore the direct effect of the LDE on international factor movements. Considering the gravity model in Eq. (4.5), which includes the full sets of directional country-year fixed effects, reliable estimates for the effect of a country’s LDE on international factor mobility are not possible. For instance, it is not possible to estimate separate country-specific effects without relying on strong identification assumptions, because the fixed effects net out all variation from the data except for the bilateral variation. In the following, a random effects approach is therefore presented as an alternative method to account for the unobservable multilateral resistance terms in the gravity equation (see Eq. (4.4)). The focus of this step is on estimating the effects of both the country-specific LDE and the bilateral linguistic distance on international factor mobility. To identify the impact of the country-specific LDE, Eq. (4.6) is augmented by the LDE of country i, LDEi, and the LDE of country j, LDEj. The trade cost function is then given by:

h 0 i tijt = exp η LDEi + ϕ LDEj + γ LDij + κ GDij + Zijt β . (4.9) Substituting Eq. (4.9) into the “structural” gravity equation (Eq. (4.2)) and log-linearizing CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 93 the resulting equation yields:

W ln Xijt = ln Yit + ln Yjt − ln Yt + η LDEi + ϕ LDEj + γ LDij 0 + κ GDij + Zijt β − (1 − σ)[ln (Πit) + ln (Pjt)] . (4.10)

To estimate the “structural” gravity equation using random effects (GLS), the population of country i and j in year t is added to Eq. (4.10) to control for further country-specific characteristics. World GDP and other common time effects are captured by a full set of year dummy variables, which are also added to Eq. (4.10). The gravity equation to be estimated becomes:

ln Xijt = φ ln P OPit + ν ln P OPjt + ϑ ln Yit + ρ ln Yjt + η LDEi + ϕ LDEj T 0 X + γ LDij + κ GDij + Zijt β + τt Tt + ζij + ijt , (4.11) t=2 where Yit and Yjt represent GDP per capita in country i and j in year t, and ζij denotes a 16 directional country-pair speciﬁc random eﬀect, assuming E (ζij) = 0.

Under the identification assumption that the error term ijt and the bilateral random effect ζij are uncorrelated with the variables of interest, the random effects estimator delivers unbiased parameter estimates. The main concern in this regard is again the possibility of omitted variable bias. However, as discussed above, to bias the estimates, a possible confounder has to be correlated with the measure of linguistic or genetic distance. I argue that the existence of such a variable is unlikely. Furthermore, the random effects estimator has two advantages over the fixed effects estimator in this analysis. First, it allows identifying the effect of variables that only vary in the same dimension as the bilateral fixed effects and are as such not identifiable in the fixed effects model. Second, the estimates do not suffer from potential over-specification due to the high number of fixed effects, which implies that much less data information will be used for the estimation of the parameters of interest (Matyas et al., 2013).

16A country-pair random eﬀect is used, since this is a more general model compared to the inclusion of separate random eﬀects for each country i and j. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 94 4.3 Data

4.3.1 Measuring Linguistic and Genetic Distance

As the data on linguistic and genetic distance are not commonly used in the economic literature, they are in the following described in some detail. Linguistic distance is the dissimilarity of languages in a multitude of dimensions, such as vocabulary, grammar, pronunciation, scripture, and phonetic inventories. This multidimensionality of linguistic distance makes it difficult to find an appropriate empirical operationalization to be used in applied economic studies. Isphording and Otten (2011) give an overview of the different approaches used in the economic literature to measure the linguistic distance between two languages and compare their advantages and disadvantages. In this analysis, the so-called Levenshtein distance is used to measure linguistic distance.

The Levenshtein Distance

As described by Isphording and Otten (2011, 2013), it is possible to derive an operationalization of linguistic distance without strong identification assumptions that underlie previous approaches. The so-called Automatic Similarity Judgment Program (ASJP) developed by the German Max Planck Institute for Evolutionary Anthropology17 aims at developing an automatic procedure to evaluate the phonetic similarity between all of the world’s languages and offers a convenient way of deriving a continuous measure of linguistic differences that is purely descriptive in nature. The basic idea is the automatic comparison of the pronunciation of words having the same meaning across languages. The average similarity across a specific set of words is then taken as a measure for the linguistic distance between the languages (Bakker et al., 2009). This distance can be interpreted as an approximation of the number of cognates between languages. The linguistic term cognates denotes common ancestries of words. A higher number of cognates indicates closer common ancestries. Although restricting its computation to differences in pronunciation, a lower Levenshtein distance therefore also indicates a higher probability of sharing other language characteristics such as grammar (see Serva, 2011). The language acquisition of second language learners is crucially affected by such differences in pronunciation and phonetic inventories, as these determine the difficulty in discriminating between different words and sounds.18 The algorithm calculating the distance between words relies on a specific phonetic

17Further information can be found at http://www.eva.mpg.de. 18For a recent overview of the linguistic literature on language background and language acquisition, see Llach (2010). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 95

alphabet, the ASJPcode. The ASJPcode uses the characters within the standard ASCII19 alphabet to represent common sounds of human communication. The ASJPcode consists of 41 different symbols representing 7 vowels and 34 consonants. Words are then analyzed as to how many sounds have to be substituted, added, or removed to transfer the one word in one language into the same word in a different language (Holman et al., 2011). The words used in this approach are taken from the so-called 40-item Swadesh list, a list including 40 words that are common in nearly all the world’s languages, including parts of the human body or expressions for common things of the environment. The Swadesh list is deductively derived by Swadesh (1952), its items are believed to be universally and culture independently included in all world’s languages.20 The ASJP program judges each word pair across languages according to their similarity in pronunciation. For example, to transfer the phonetic transcription of the English word you, transcribed as yu, into the transcription of the respective German word du, one simply has to substitute the first consonant. But to transfer maunt3n, which is the transcription of mountain, into bErk, which is the transcription of the German Berg, one has to remove or substitute each of the 7 consonants and vowels, respectively. The following formalization of the computation follows Petroni and Serva (2010). To normalize the distance according to the word length, the resulting number of changes is divided by the word length of the longer word. Denoting this normalized distance between

1 X LDN(α, β) = D(αi, βi). (4.12) M i To additionally account for potential similarities in phonetic inventories which might lead to a similarity by chance, a global distance between languages is deﬁned as the average Levenshtein distance of words with diﬀerent meanings:

1 X Γ(α, β) = D(αi, βj). (4.13) M(M − 1) i6=j The ﬁnal normalized and divided linguistic distance (LDND) is then deﬁned as the quotient of the normalized linguistic distance and the global distance between α and β:

LDN(α, β) LDND(α, β) = . (4.14) Γ(α, β)

19American Standard Code for Information Interchange, keyboard-character-encoding scheme. 20A list of the 40 words is available upon request. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 96

The resulting continuous measure can broadly be interpreted as a percentage measure of dissimilarity between languages, with lower numbers indicating a closer relation. In a few cases, the resulting numbers are bigger than 100%, indicating a dissimilarity that exceeds a potentially incidental similarity between languages that would be expected due to similar phonetic inventories.21 The Levenshtein distance is computed for every country-pair in the data sets. In mono-lingual countries, the respective native language is assigned to the country. In multi- lingual countries, the most prevalent native language is assigned, which was identiﬁed using a multitude of sources, including CIA’s World Factbook, encyclopedias, and Internet resources.22 Table 4.1 lists the closest and furthest languages toward English and the closest and furthest language-pairs worldwide. The results show that the measurement via the normalized and divided Levenshtein distance is in line with an intuitive guessing about language dissimilarities.

The Genetic Distance

For an exogenous measure of the cultural differences between two countries, it is important that the measure does not reflect bilateral relationships. For instance, religious differences are rooted in past history. However, this history is relatively recent, therefore a measure of religious dissimilarities could reflect past relationships between countries (Guiso et al., 2009). To measure deeper cultural differences between two countries, the genetic distance between the dominant population groups in these countries is used. The data on genetic differences was originally gathered by Cavalli-Sforza et al. (1994) for 42 subpopulations. Spolaore and Wacziarg (2009) extended this data to genetic differences between 180 countries.23 In order to do so, Spolaore and Wacziarg combine the frequencies of gene manifestations in populations sampled by Cavalli-Sforza et al. (1994) and the ethnicity composition of countries compiled by Alesina et al. (2003) to derive a measure of the genetic distance between countries.

21The ASJP algorithm allows including or excluding loan words from diﬀerent languages, e.g., the predominance of former Latin words in many of the European languages. While it makes sense to exclude these loan words in the analysis of the long-term development of languages, these words are included in this analysis, as they lead to certain similarities of languages that might ease the later language transfer in the acquisition process. The necessary software to compute the distance matrix is available at http://www.eva.mpg.de. The complete distance matrix used in this analysis is available upon request. 22For example, English is used as the native language in the United Kingdom, because it is a monolingual country and English is the national language. In a multi-lingual country such as Canada, English instead of French is employed, because English is the most prevalent native language. A comprehensive index of assigned languages with further explanations is available upon request. 23For a more detailed description of the genetic distance data and the construction of the measure, see Spolaore and Wacziarg (2009). The data set is available at http://www.anderson.ucla.edu/faculty_pages/romain.wacziarg. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 97

The genetic distance is measured as the difference in allele frequencies. Alleles are the specific manifestation of a gene, which might differ between individuals. The genetic distance measure as defined by Cavalli-Sforza et al. (1994) is related to the inverse probability that groups of alleles are the same for two populations. Hence, the lower the common frequency of alleles in two populations, the longer these populations have been separated. Changes in genes, hence the emergence of new alleles, happen randomly at an almost constant time.24 This constant rate of change over time makes it a reasonable proxy for the time populations spent separated, making the genetic distance an “excellent summary statistic capturing divergence in the whole set of implicit beliefs, customs, habits, biases, conventions, etc. that are transmitted across generations—biologically and/or culturally—with high persistence.” (Spolaore and Wacziarg, 2009, p. 471). Using this measure of genetic distance as a proxy for cultural differences and assuming a reasonable correlation between the measured genetic distance and any unobservable cultural differences should allow the identification of the effect of cultural dissimilarities between two countries on international factor mobility.

4.3.2 Data on Trade Flows, Portfolio Investment, and Banking Claims

The analysis further relies on bilateral trade data from the International Monetary Fund’s Direction of Trade Statistics (DOTS), which reports annual aggregate import and export ﬂows. FOB exports and CIF imports are measured in millions of nominal US dollars.25 The sample covers the 1950–2006 period for 180 exporter and 187 importer countries. For some country-pairs, the trade data are unavailable, thus around 530,000 non-zero observations of exports are used in the baseline regression. The bilateral trade data are merged with the bilateral linguistic and genetic distance data at the country-pair level. Since both the linguistic and the genetic distance data are symmetric, the direction of trade does not have to be considered. The data on bilateral asset holdings is gathered by the International Monetary Fund in its Coordinated Portfolio Investment Survey (CPIS), which reports year-end data on

24As evolutionary pressure might direct the random change into certain directions, the genetic distance measure focuses on neutral genes, which are not prone to evolutionary pressure. 25The terms FOB and CIF are abbreviations for “Free on Board” and “Cost including Insurance and Freight”, respectively, and describe the method used by countries to report the value of imports and exports. For further details on the data set, see http://www.imf.org/external/data.htm. I use the data set provided by Rose and Spiegel (2011). CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 98

stocks of cross-border assets measured in millions of nominal US dollars.26 The sample covers the years 2001–2003 and contains information on portfolio investments of 62 source in 173 host countries, leading to 8,596 non-zero observations of asset stocks used in the baseline regression. As the bilateral trade data, this data set is merged with the bilateral linguistic and genetic distance data at the country-pair level. The source of the second ﬁnancial data set is the Consolidated Banking Statistics published by the Bank of International Settlements (BIS), which reports data on bilateral consolidated international banking claims measured in millions of nominal US dollars on a quarterly basis.27 The sample covers annual data for the years 1983–2006 and contains information on banking claims of 24 source and 182 host countries, leading to 34,838 non-zero observations of international banking claims used in the baseline regression. Again, this data set is merged with the bilateral linguistic and genetic distance data at the country-pair level. The data on the control variables are obtained from the CEPII “square” gravity data set compiled by Head et al. (2010).28 This database contains a set of standard gravity variables which is used in all estimations, such as the geographical distance29, colonial ties, and information on regional trade agreements.

4.4 Empirical Findings

4.4.1 The Eﬀect of Linguistic and Genetic Distance on Trade

This section analyzes the effect of linguistic and genetic distance on bilateral export flows. Linguistic and genetic distance are important determinants in explaining bilateral exports by approximating language and cultural barriers. In order to examine systematically how linguistic and genetic distance affect export flows, Table 4.2 reports results for six OLS regressions employing exporter-year and importer-year fixed effects. The country-year specific fixed effects capture the effects of all monadic variables (e.g., GDP per capita, population, and multilateral resistance terms) and remove these variables from the specifications. In each column, standard errors of estimates are robust to heteroskedasticity and correlation of error terms within

26Further details on the data and the data set itself are available at http://cpis.imf.org. I use the data set provided by Rose and Spiegel (2009). 27Further details on the data and the data itself are available at http://www.bis.org/statistics/consstats.htm. I use the data set provided by Rose and Spiegel (2009). 28For further details on the variables and the data set, see http://www.cepii.fr/CEPII/en/bdd_modele/presentation.asp?id=8. 29In all estimations, the population-weighted great circle distance between large cities of the two countries is used. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 99

country-pairs. The analysis starts by estimating bilateral exports on linguistic distance and a host of control variables using the gravity model in Eq. (4.5) (column (1)). The results show that the coefficient of the linguistic distance is negative and significant at the 0.1% level. The estimated coefficient of −0.078 implies that a one standard deviation increase in the linguistic distance lowers a country’s exports by exp (−0.078) − 1 ≈ 7.5%, which is a sizeable impact. In column (3), linguistic distance is substituted with genetic distance. The effect is similar, but weaker. A one standard deviation increase in the genetic distance lowers bilateral exports by 4.9%. The results are robust in magnitude and in significance when both factors are simultaneously included in the regression (column (5)). This is not surprising, given the relative low correlation between these two variables, and shows that both capture different aspects in the determination of trade flows.30 The effect of linguistic and genetic distance is identified by assuming a linear relationship between these variables and bilateral exports. However, imposing wrong functional forms may lead to misleading results for trade barrier estimates (Anderson and van Wincoop, 2004). Eaton and Kortum (2002) assume heterogeneous trade barriers for six different geographic distance intervals. They generalize the treatment of geographical distance with a flexible functional form using a spline approach, which is likely to be more robust to specification errors. To avoid imposing a functional form on the relationship between exports and linguistic and genetic distance, respectively, I follow Eaton and Kortum (2002) and rerun the models using indicator variables. To capture the effect of linguistic distance, a set of five indicator variables is employed. The first indicator comprises all country- pairs speaking the same language, i.e., their linguistic distance is zero. The remaining four indicators turn on for the first up to the fourth quartile of the positive values of the linguistic distance distribution. The same approach is used for the genetic distance. However, instead of employing four indicators for the non-zeros values, the distribution is separated only into two groups. The three dummies indicate the effect on exports for country-pairs with zero, low, and high genetic distance, respectively. The results are shown in columns (2), (4), and (6). The estimated coefficients have the same signs and similar levels of significance. Since the reference group is comprised of country-pairs that speak the same language and have genetically the same dominant populations, respectively, the effects become stronger in magnitude.31 The estimated coefficients of the traditional gravity control variables are in line with estimates from the literature. These effects all bear the expected signs, are large in size, economically substantial, and statistically highly significant. For instance, geographical distance reduces trade with an estimated elasticity of about 1.3%. The other dyadic control

30The correlation coefficient is 0.175. 31To check the robustness of these specifications, the estimations were tested with different sets of indicator variables. The results are broadly comparable to the results presented in Table 4.2. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 100

variables – shared border, shared legal origin, colonial relationship, common regional trade arrangement, both countries GATT/WTO members, and common currency – promote bilateral exports as expected. Throughout the rest of the analysis of bilateral exports, both linguistic and genetic distance are simultaneously included in the estimation models. Table 4.3 reports the results showing whether a country’s LDE and the bilateral linguistic distance affect bilateral factor movements above and beyond the inclusion of the direct effects. In doing so, the indicator variables defined in Eq. (4.7) and (4.8) are added separately to the baseline regressions from Table 4.2 (column (5) and (6)). Both indicator variables enter the regressions with a positive sign, but are statistically insignificant except for column (3). The coefficient of LDE(2) reveals that if at least one trading partner in a pair faces a higher LDE than the bilateral linguistic distance, exports between these countries increase by 6.1%. The coefficients for the linguistic and genetic distance variables are robust in magnitude and in significance when the indicator variables are included in the regressions. As expected, the coefficients of the gravity control variables remain also unchanged. The results described above present OLS estimates employing exporter-year and importer-year fixed effects, representing the standard way gravity models have been estimated in the literature. The downside of using this specification is that all monadic effects are captured by the fixed effects. Under the stronger assumption that the error term and the random effect are uncorrelated with the variables of interest, the random effects estimator delivers unbiased parameter estimates. Table 4.4 presents results obtained by estimating the random effects model in Eq. (4.11). Due to the unavailability of GDP per capita data for some countries, the sample is reduced to 495,286 country-pair observations. For a proper comparison, the analysis starts by re-estimating the baseline regressions from Table 4.2 (column (5)) using the smaller sample. The five subsequent columns employ the random effects model. Throughout all estimations, year-specific fixed effects are added to take common time effects into account. While column (2) shows the results without further controls, a comprehensive set of export and import continent dummy variables is added to the regression model in column (3). In Column (4), the continent dummy variables are replaced by sets of world region fixed effects to control for heterogeneous effects within the continents. The final two specifications add the indicator variables defined in Eq. (4.7) and (4.8) to the model in column (4). The coefficients of the income and population variables reveal the expected results. Throughout all specifications, they are positively related to bilateral exports and statistically significant. The parameters of particular interest are η and ϕ, representing the effect CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 101 of the exporter’s and importer’s LDE. The magnitudes of these effects are substantial. A one standard deviation increase in the exporter LDE reduces bilateral exports by between 17.5 and 22.7%, depending on the specification. A similar effect shows up for the importer LDE. The results for the bilateral linguistic and genetic distance are robust in sign and significance, but larger in magnitude. The results reported in Tables 4.2-4.4 demonstrate that linguistic and genetic distance play an import role in explaining international export flows. I hypothesized above that the effect of linguistic distance in the determination of international investments, i.e., cross-border asset stocks and international banking claims, should be comparable to the impact on bilateral exports. However, the role of cultural difference on international investment decisions is theoretically ambiguous. These conjectures are analyzed in the next section.

4.4.2 The Eﬀect of Linguistic and Genetic Distance on Invest- ment

This section analyzes the effect of linguistic and genetic distance on international investment. In order to examine different types of investments, two data sets are considered. First, the gravity models displayed in Tables 4.2-4.4 are re-estimated using data on cross-border asset stocks. Second, data on consolidated international banking claims are employed to rerun the gravity equations shown in Tables 4.2-4.4. Consistent with prior expectations, Tables 4.5 and 4.6 shows that linguistic distance enters with negative sign and is highly significant. The estimated coefficients are larger compared to the estimates in Table 4.2. As expected, the results for the genetic distance are ambiguous. The specifications applying the linear measure of genetic distance report insignificant coefficients (column (3) and (5)). In contrast, employing the splines approach yields negative significant results, but only for country-pairs with a low genetic distance. This result might indicate that investors prefer culturally familiar countries for their investments. However, if they invest in culturally unfamiliar countries, they prefer countries with a relative high difference due to the lower correlation of the stock markets and the higher diversification potential for the portfolios of foreign investors. Tables 4.7 and 4.8 show the results of whether a country’s LDE and the bilateral linguistic distance affect bilateral factor movements above and beyond the inclusion of the direct effects. In doing so, the indicator variables defined in Eq. (4.7) and (4.8) are added separately to the baseline regressions from Tables 4.5 and 4.6, respectively. Both indicator variables are statistically insignificant in all regressions. The coefficients of the variables of interest as well as the gravity control variables remain unchanged as compared to the CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 102

prior results. Turing to the results of the random effects model, Tables 4.9 and 4.10 reveal that both the exporter’s and importer’s LDE enter with the expected negative signs and are statistically significant at the 0.1% level. The sizes of the estimated coefficient are substantial. The results of the bilateral linguistic distance are robust in sign and significance, whereas the estimated coefficients of the genetic distance are ambiguous.

4.5 Conclusion

Empirical examinations of international factor movements, e.g., bilateral trade, financial and migration flows, have a long tradition in the economic literature. Starting with the analysis by Tinbergen (1962), the gravity model plays a key role in this literature. Hence, numerous studies have provided important information about the impact of potential enhancers and barriers on international factor mobility such as membership in the GATT/WTO (e.g., Rose, 2004), colonial linkages (e.g., Head et al., 2010), and geographical distance (e.g., Portes and Rey, 2005). This paper is motivated by the impression that the relationship between language and cultural differences and factor movements is underinvestigated in the economic literature. The standard measures of language and cultural dissimilarities employed in gravity models (such as indicators for sharing a common language, geographic distance, and colonial ties) are insufficient in picking up the entire influence of these impact factors on bilateral factor movements. Moreover, so far unconsidered communication channels, i.e., English as the lingua franca in international business relations, appear to have a major impact on bilateral flows. In this paper, I show that language and cultural differences have a significant negative impact on bilateral factor mobility, even after controlling for a host of standard gravity variables. While these effects are economically important and highly significant for international trade flows, only language differences affect financial flows. This finding might be interpreted as a weak support for the diversification motive in international investment transactions, which indicates that the stock markets of culturally unfamiliar countries reveal a lower correlation to the stock markets of foreign investors. In the preferred specification, I control for a large number of covariates commonly used in gravity equations and account for exporter- and importer-year fixed effects. The results show that an increase in the linguistic distance by one standard deviation decreases bilateral exports by 7.4% and both cross-border asset stocks and international banking claims by about 12%. Although the findings on the isolated effect of a country’s linguistic distance toward CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 103

English (LDE) suggest highly significant and economically important effects across all specifications and datasets, they should be interpret with some caution. To deliver unbiased estimates of the true effect, the employed random effects model rests on strong identification assumptions. Other empirical methods such as the approach suggested by Baier and Bergstrand (2009) and the analysis of measures of LDE that additionally vary over time may shed some further light on the relation between international factor mobility and English proficiency. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 104 Tables

Table 4.1: Closest and Furthest Language Pairs with Respect to the Levenshtein Distance Closest Furthest Language Distance Language Distance Distance to English Afrikaans 62.08 Vietnamese 104.06 Dutch 63.22 Turkmen 103.84 Norwegian 64.12 Hakka (China) 103.10 Swedish 64.40 Cambodian 103.00 Frisian 69.49 Finnish 102.27

Language pairs worldwide Croatian Laotien 28.36 106.35 Slovenian Korean Jamaican Creole Fijian 30.88 106.75 Vincentian Creole Shona (Zimbabwe) Egyptian Arabic Vietnamese 31.44 107.33 Iraqi Arabic Western Punjabi (Pakistan) Notes: – The table shows the five closest and furthest languages toward English and the three closest and furthest language-pairs worldwide according to the normalized and divided Levenshtein distance. – Only languages spoken within samples are listed. – Geographic origin of language in parentheses. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 105 ∗∗∗ ∗∗∗ ∗∗∗ ∗ ∗ ∗∗ ∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – – 257 712 023) 409 092) 488 027) 263 099) 604 038) 337 047) 804 095) 110 076) 185 082) 214 084) 260 086) 136 054) 167 056) ...... 0 1 0 0 0 0 0 0 0 0 0 1 0 − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ – – 047) (0 799 096)077 (0 016) 049 017) 250 712 0 023) (0 430 091) (0 466 027) (0 236 097) (0 616 038) (0 337 ...... 1 0 0 1 0 0 0 0 0 − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – 026) (0 289 098) (0 605 038) (0 345 047) (0 808 095) (0 175 053)205 (0 055) (0 274 712 0 022) (0 409 092) (0 498 ...... 1 0 0 1 0 0 0 0 0 (0 (0 − − − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ––– ––– ––– ––– 279 712 0 022) (0 441 092) (0 499 026) (0 298 098) (0 615 038) (0 355 047) (0 788 095) (0 050 017) (0 ...... 0 0 0 1 0 0 1 0 (0 − − 10% level. – Cluster-robust standard errors are reported in † ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ––– 274 712 0 022) (0 444 092) (0 486 027) (0 261 099) (0 614 038) (0 342 047) (0 821 095) (0 118 075) (0 211 081) (0 245 083) (0 293 085) (0 ...... 0 1 0 0 0 0 0 0 0 1 0 (0 (0 (0 (0 − − − − − 5% level; ∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 047) (0 837 095) (0 078 016) (0 268 712 0 022) (0 445 092) (0 466 027) (0 227 097) (0 619 038) (0 336 ...... 1 0 0 0 1 0 0 0 0 Linguistic Distance Genetic Distance Baseline Model (0 (0 (0 (0 (0 (0 (0 (0 1% level; − − Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE ∗∗ 0.1% level; ∗∗∗ Effect of Linguistic and Genetic Distance on Bilateral Exports parentheses. – The dependent variable is the logarithm of bilateral export flows. Notes: – Significant at: Linguistic Distance Spline 2 – Linguistic Distance Spline 3 – Linguistic Distance Spline 4 – Linguistic Distance Spline 5 – Genetic Distance Spline 2 – – – Genetic Distance Spline 3 – – – 2 Shared Currency Linguistic Distance Linguistic Distance (Ref. Same Lang.) Genetic Distance – – Genetic Distance (Ref. Same Pop.) Exporter-Year FEImporter-Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ln Geo. Distance R Observations 529,562 529,562 529,562 529,562 529,562 529,562 Shared Border Shared Legal Colonial Relationship RTA Both GATT/WTO Table 4.2: CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 106

Table 4.3: Effect of the Bilateral LDE Indicator on Bilateral Exports Both LDE > LD One LDE > LD Coef/StdE Coef/StdE Coef/StdE Coef/StdE ln Geo. Distance −1.250∗∗∗ −1.257∗∗∗ −1.252∗∗∗ −1.258∗∗∗ (0.023) (0.023) (0.023) (0.023) Shared Border 0.426∗∗∗ 0.408∗∗∗ 0.428∗∗∗ 0.409∗∗∗ (0.091) (0.092) (0.091) (0.092) Shared Legal 0.464∗∗∗ 0.486∗∗∗ 0.461∗∗∗ 0.484∗∗∗ (0.027) (0.027) (0.027) (0.027) Colonial Relationship 1.241∗∗∗ 1.264∗∗∗ 1.242∗∗∗ 1.265∗∗∗ (0.097) (0.099) (0.098) (0.099) RTA 0.615∗∗∗ 0.604∗∗∗ 0.613∗∗∗ 0.604∗∗∗ (0.038) (0.038) (0.038) (0.038) Both GATT/WTO 0.335∗∗∗ 0.337∗∗∗ 0.335∗∗∗ 0.336∗∗∗ (0.047) (0.047) (0.047) (0.047) Shared Currency 0.797∗∗∗ 0.804∗∗∗ 0.796∗∗∗ 0.804∗∗∗ (0.096) (0.095) (0.096) (0.095) Linguistic Distance −0.073∗∗∗ – −0.074∗∗∗ – (0.016) (0.016) Linguistic Distance (Ref. Same Lang.) Linguistic Distance Spline 2 – −0.117 – −0.121 (0.077) (0.076) Linguistic Distance Spline 3 – −0.184∗ – −0.189∗ (0.082) (0.082) Linguistic Distance Spline 4 – −0.209∗ – −0.211∗ (0.086) (0.085) Linguistic Distance Spline 5 – −0.252∗∗ – −0.244∗∗ (0.088) (0.088) Genetic Distance −0.047∗∗ – −0.046∗∗ – (0.017) (0.017) Genetic Distance (Ref. Same Pop.) Genetic Distance Spline 2 – −0.135∗ – −0.135∗ (0.054) (0.054) Genetic Distance Spline 3 – −0.167∗∗ – −0.167∗∗ (0.056) (0.056) LDE(1) 0.046 0.020 –– (0.036) (0.045) LDE(2) –– 0.059∗ 0.033 (0.027) (0.034) Exporter-Year FE Yes Yes Yes Yes Importer-Year FE Yes Yes Yes Yes R2 0.712 0.712 0.712 0.712 Observations 529,562 529,562 529,562 529,562 Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral export ﬂows. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 107

Table 4.4: Effect of the Linguistic Distance toward English on Bilateral Exports

Baseline RE-Model RE-Model incl. LDE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

ln Pop, origin – 0.950∗∗∗ 0.926∗∗∗ 0.851∗∗∗ 0.850∗∗∗ 0.850∗∗∗ (0.010) (0.010) (0.012) (0.012) (0.012) ln Pop, dest – 0.821∗∗∗ 0.816∗∗∗ 0.839∗∗∗ 0.838∗∗∗ 0.838∗∗∗ (0.010) (0.010) (0.012) (0.012) (0.012) ln GDP/pc, origin – 0.945∗∗∗ 0.911∗∗∗ 0.892∗∗∗ 0.892∗∗∗ 0.892∗∗∗ (0.013) (0.014) (0.015) (0.015) (0.015) ln GDP/pc, dest – 0.658∗∗∗ 0.650∗∗∗ 0.646∗∗∗ 0.647∗∗∗ 0.647∗∗∗ (0.011) (0.012) (0.013) (0.013) (0.013) ln Geo. Distance −1.271∗∗∗ −1.189∗∗∗ −1.150∗∗∗ −1.264∗∗∗ −1.262∗∗∗ −1.264∗∗∗ (0.023) (0.022) (0.027) (0.027) (0.027) (0.027) Shared Border 0.414∗∗∗ 0.890∗∗∗ 0.985∗∗∗ 0.733∗∗∗ 0.726∗∗∗ 0.730∗∗∗ (0.092) (0.092) (0.092) (0.092) (0.092) (0.092) Shared Legal 0.477∗∗∗ 0.347∗∗∗ 0.477∗∗∗ 0.525∗∗∗ 0.522∗∗∗ 0.518∗∗∗ (0.027) (0.036) (0.035) (0.035) (0.035) (0.035) Colonial Relationship 1.183∗∗∗ 1.761∗∗∗ 1.514∗∗∗ 1.537∗∗∗ 1.544∗∗∗ 1.543∗∗∗ (0.096) (0.112) (0.115) (0.116) (0.116) (0.116) RTA 0.595∗∗∗ 0.393∗∗∗ 0.393∗∗∗ 0.392∗∗∗ 0.392∗∗∗ 0.392∗∗∗ (0.038) (0.024) (0.024) (0.024) (0.024) (0.024) Both GATT/WTO 0.292∗∗∗ 0.142∗∗∗ 0.155∗∗∗ 0.154∗∗∗ 0.154∗∗∗ 0.154∗∗∗ (0.049) (0.018) (0.019) (0.019) (0.019) (0.019) Shared Currency 0.747∗∗∗ 0.523∗∗∗ 0.550∗∗∗ 0.541∗∗∗ 0.541∗∗∗ 0.541∗∗∗ (0.103) (0.081) (0.081) (0.081) (0.081) (0.081) Linguistic Distance −0.071∗∗∗ −0.122∗∗∗ −0.180∗∗∗ −0.175∗∗∗ −0.165∗∗∗ −0.170∗∗∗ (0.016) (0.018) (0.019) (0.019) (0.020) (0.019) Genetic Distance −0.055∗∗ −0.095∗∗∗ −0.019 −0.091∗∗∗ −0.089∗∗∗ −0.088∗∗∗ (0.018) (0.017) (0.017) (0.018) (0.018) (0.018) LDE, origin – −0.192∗∗∗ −0.219∗∗∗ −0.242∗∗∗ −0.248∗∗∗ −0.258∗∗∗ (0.017) (0.017) (0.022) (0.022) (0.023) LDE, dest – −0.242∗∗∗ −0.277∗∗∗ −0.227∗∗∗ −0.233∗∗∗ −0.243∗∗∗ (0.016) (0.017) (0.022) (0.022) (0.022) LDE(1) –––– 0.086† – (0.045) LDE(2) ––––– 0.087∗ (0.034) Exporter-Year FE Yes No No No No No Importer-Year FE Yes No No No No No Year FE No Yes Yes Yes Yes Yes Continent FE No No Yes No No No World Region FE No No No Yes Yes Yes

R2 0.718 0.587 0.593 0.613 0.613 0.613 Observations 495,286 495,286 495,286 495,286 495,286 495,286 Notes: – Significant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral export flows. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 108 ∗∗∗ ∗∗∗ † † ∗∗ ∗∗∗ ∗∗∗ ∗ – – 054 767 070) 142 185) 348 086) 419 179) 489 116) 630 155) 086 187) 314 190) 278 207) 379 225) 337 112) 060 157) ...... 0 0 0 0 0 0 0 0 0 0 1 − − − − − − − ∗∗∗ ∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ – – 156)128 (0 040) 069 094) 024 766 0 070) (0 218 0 187) (0 335 086) (0 371 178) (0 496 117) (0 619 ...... 1 0 0 0 0 0 0 − − ∗∗∗ ∗∗∗ ∗ ∗∗∗ ∗∗∗ ∗∗∗ – 080) (0 445 173) (0 486 116) (0 615 154) (0 367 111)120 (0 149) (0 079 766 0 066) (0 145 0 182) (0 410 ...... 1 0 0 0 0 0 0 (0 (0 − − − ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ––– ––– ––– ––– 085 765 0 065) (0 210 0 185) (0 424 080) (0 468 177) (0 478 116) (0 588 152) (0 050 092) (0 ...... 0 0 1 0 0 0 (0 − 10% level. – Cluster-robust standard errors are reported in † ∗∗∗ ∗∗∗ ∗ † ∗ ∗∗∗ ∗∗∗ ∗ ––– 030 766 0 067) (0 218 0 188) (0 350 087) (0 425 183) (0 486 117) (0 613 153) (0 173 186) (0 395 189) (0 372 205) (0 446 220) (0 ...... 0 0 0 0 0 0 0 0 1 (0 (0 (0 (0 − − − − − 5% level; ∗ ∗∗∗ ∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ 156) (0 125 040) (0 007 766 0 067) (0 225 0 187) (0 335 086) (0 375 179) (0 496 117) (0 620 ...... 1 0 0 0 0 0 0 0 Linguistic Distance Genetic Distance Baseline Model (0 (0 (0 (0 (0 (0 (0 1% level; − − Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE ∗∗ 0.1% level; ∗∗∗ Effect of Linguistic and Genetic Distance on Cross-Border Asset Stocks parentheses. – The dependent variable is the logarithm of bilateral cross-border asset stocks. Notes: – Significant at: Linguistic Distance Spline 2 – Linguistic Distance Spline 3 – Linguistic Distance Spline 4 – Linguistic Distance Spline 5 – Genetic Distance Spline 2 – – – Genetic Distance Spline 3 – – – 2 Linguistic Distance Linguistic Distance (Ref. Same Lang.) Genetic Distance – – Genetic Distance (Ref. Same Pop.) Origin-Year FEDestination-Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ln Geo. Distance R Observations 8,596 8,596 8,596 8,596 8,596 8,596 Shared Border Shared Legal Colonial Relationship RTA Shared Currency Table 4.5: CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 109 ∗∗ ∗∗∗ ∗∗ ∗ ∗∗∗ ∗∗ ∗∗∗ – – 006 816 078) 015 208) 169 058) 048 116) 091 112) 168 165) 437 147) 567 158) 505 178) 386 177) 100 095) 055 145) ...... 0 1 0 0 0 0 0 0 0 0 1 0 − − − − − − − − − ∗ ∗∗∗ ∗∗∗ ∗∗∗ – – 127 033) 107 098) 009 815 0 071) (0 003 204) (0 144 059) (0 031 116) (0 105 112) (0 171 163) (0 ...... 1 0 0 1 0 0 0 0 − − − − − ∗∗∗ ∗∗∗ ∗∗∗ † – 057) (0 122 115) (0 123 110) (0 149 157) (0 175 093) (0 031 141) (0 032 814 0 078) (0 034 204) (0 217 ...... 1 0 0 1 0 0 0 0 (0 (0 − − − − − ∗∗∗ ∗∗∗ ∗∗∗ ––– ––– ––– ––– 037 814 0 070) (0 005 200) (0 215 056) (0 135 114) (0 128 110) (0 160 153) (0 080 099) (0 ...... 0 1 0 1 0 0 (0 − − − 10% level. – Cluster-robust standard errors are reported in † ∗∗ ∗∗∗ ∗∗ ∗ ∗∗∗ ∗∗ ∗∗∗ ––– 992 815 0 076) (0 017 0 207) (0 165 057) (0 052 116) (0 087 112) (0 169 163) (0 458 145) (0 604 155) (0 534 173) (0 404 172) (0 ...... 0 1 0 0 0 0 0 0 0 (0 (0 (0 (0 − − − − − − − 5% level; ∗ ∗ ∗∗∗ ∗∗∗ ∗∗∗ 033) (0 987 815 0 077) (0 013 0 206) (0 147 059) (0 032 117) (0 099 112) (0 165 162) (0 123 ...... 0 0 0 0 1 0 0 0 Linguistic Distance Genetic Distance Baseline Model (0 (0 (0 (0 (0 (0 (0 1% level; − − − − Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE ∗∗ 0.1% level; ∗∗∗ Effect of Linguistic and Genetic Distance on Bilateral Banking Claims parentheses. – The dependent variable is the logarithm of bilateral consolidated international banking claims. Notes: – Significant at: Linguistic Distance Spline 2 – Linguistic Distance Spline 3 – Linguistic Distance Spline 4 – Linguistic Distance Spline 5 – Genetic Distance Spline 2 – – – Genetic Distance Spline 3 – – – 2 Linguistic Distance (Ref. Same Lang.) Genetic Distance – – Genetic Distance (Ref. Same Pop.) Origin-Year FEDestination-Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ln Geo. Distance R Observations 34,838 34,838 34,838 34,838 34,838 34,838 Shared Border Shared Legal Colonial Relationship RTA Shared Currency Linguistic Distance Table 4.6: CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 110

Table 4.7: Effect of the Bilateral LDE Indicator on Cross-Border Asset Stocks Both LDE > LD One LDE > LD Coef/StdE Coef/StdE Coef/StdE Coef/StdE ln Geo. Distance −1.023∗∗∗ −1.053∗∗∗ −1.024∗∗∗ −1.050∗∗∗ (0.070) (0.070) (0.070) (0.070) Shared Border 0.201 0.140 0.218 0.152 (0.187) (0.185) (0.187) (0.185) Shared Legal 0.329∗∗∗ 0.346∗∗∗ 0.335∗∗∗ 0.363∗∗∗ (0.087) (0.088) (0.088) (0.088) Colonial Relationship 0.385∗ 0.420∗ 0.372∗ 0.414∗ (0.182) (0.179) (0.179) (0.178) RTA 0.497∗∗∗ 0.489∗∗∗ 0.496∗∗∗ 0.488∗∗∗ (0.117) (0.116) (0.117) (0.116) Shared Currency 0.617∗∗∗ 0.630∗∗∗ 0.619∗∗∗ 0.636∗∗∗ (0.156) (0.155) (0.156) (0.154) Linguistic Distance −0.121∗∗ – −0.127∗∗ – (0.041) (0.041) Linguistic Distance (Ref. Same Lang.) Linguistic Distance Spline 2 – −0.096 – −0.049 (0.194) (0.188) Linguistic Distance Spline 3 – −0.316† – −0.307 (0.189) (0.189) Linguistic Distance Spline 4 – −0.278 – −0.291 (0.207) (0.207) Linguistic Distance Spline 5 – −0.376† – −0.436† (0.226) (0.229) Genetic Distance 0.084 – 0.069 – (0.097) (0.095) Genetic Distance (Ref. Same Pop.) Genetic Distance Spline 2 – −0.335∗∗ – −0.344∗∗ (0.112) (0.112) Genetic Distance Spline 3 – −0.058 – −0.059 (0.158) (0.157) LDE(1) 0.107 0.023 –– (0.124) (0.144) LDE(2) –– 0.003 −0.107 (0.082) (0.094) Origin-Year FE Yes Yes Yes Yes Destination-Year FE Yes Yes Yes Yes R2 0.766 0.767 0.766 0.767 Observations 8,596 8,596 8,596 8,596 Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral cross-border asset stocks. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 111

Table 4.8: Effect of the Bilateral LDE Indicator on Bilateral Banking Claims Both LDE > LD One LDE > LD Coef/StdE Coef/StdE Coef/StdE Coef/StdE ln Geo. Distance −1.008∗∗∗ −1.006∗∗∗ −1.009∗∗∗ −1.004∗∗∗ (0.071) (0.078) (0.071) (0.078) Shared Border 0.009 −0.017 0.004 −0.012 (0.206) (0.210) (0.204) (0.208) Shared Legal 0.149∗ 0.167∗∗ 0.155∗∗ 0.173∗∗ (0.059) (0.060) (0.059) (0.058) Colonial Relationship 1.021∗∗∗ 1.049∗∗∗ 1.017∗∗∗ 1.044∗∗∗ (0.117) (0.117) (0.117) (0.116) RTA −0.105 −0.091 −0.102 −0.090 (0.112) (0.112) (0.112) (0.112) Shared Currency −0.171 −0.168 −0.171 −0.166 (0.163) (0.165) (0.162) (0.164) Linguistic Distance −0.130∗∗∗ – −0.131∗∗∗ – (0.033) (0.033) Linguistic Distance (Ref. Same Lang.) Linguistic Distance Spline 2 – −0.439∗∗ – −0.433∗∗ (0.149) (0.147) Linguistic Distance Spline 3 – −0.566∗∗∗ – −0.575∗∗∗ (0.159) (0.158) Linguistic Distance Spline 4 – −0.503∗∗ – −0.516∗∗ (0.179) (0.179) Linguistic Distance Spline 5 – −0.384∗ – −0.411∗ (0.180) (0.185) Genetic Distance 0.100 – 0.102 – (0.098) (0.098) Genetic Distance (Ref. Same Pop.) Genetic Distance Spline 2 – −0.100 – −0.099 (0.094) (0.094) Genetic Distance Spline 3 – 0.056 – 0.059 (0.145) (0.145) LDE(1) −0.052 0.009 –– (0.088) (0.100) LDE(2) –– −0.081 −0.031 (0.056) (0.068) Origin-Year FE Yes Yes Yes Yes Destination-Year FE Yes Yes Yes Yes R2 0.815 0.816 0.815 0.816 Observations 34,838 34,838 34,838 34,838 Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral consolidated international banking claims. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 112

Table 4.9: Effect of the Linguistic Distance toward English on Cross-Border Asset Stocks

Baseline RE-Model RE-Model incl. LDE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

ln Pop, origin – 0.397∗∗∗ 0.414∗∗∗ 0.388∗∗∗ 0.386∗∗∗ 0.386∗∗∗ (0.025) (0.024) (0.029) (0.029) (0.029) ln Pop, dest – 0.704∗∗∗ 0.714∗∗∗ 0.690∗∗∗ 0.689∗∗∗ 0.689∗∗∗ (0.026) (0.026) (0.030) (0.031) (0.030) ln GDP/pc, origin – 1.597∗∗∗ 1.763∗∗∗ 1.639∗∗∗ 1.644∗∗∗ 1.649∗∗∗ (0.052) (0.059) (0.091) (0.091) (0.091) ln GDP/pc, dest – 1.160∗∗∗ 1.157∗∗∗ 1.229∗∗∗ 1.231∗∗∗ 1.239∗∗∗ (0.038) (0.042) (0.052) (0.052) (0.052) ln Geo. Distance −1.049∗∗∗ −0.629∗∗∗ −0.698∗∗∗ −0.915∗∗∗ −0.913∗∗∗ −0.922∗∗∗ (0.072) (0.057) (0.068) (0.069) (0.069) (0.069) Shared Border 0.173 0.631∗∗ 0.639∗∗ 0.279 0.245 0.254 (0.188) (0.201) (0.201) (0.214) (0.213) (0.212) Shared Legal 0.331∗∗∗ 0.379∗∗∗ 0.332∗∗ 0.349∗∗∗ 0.335∗∗ 0.320∗∗ (0.087) (0.110) (0.110) (0.106) (0.107) (0.107) Colonial Relationship 0.394∗ 0.583∗∗ 0.609∗∗ 0.777∗∗∗ 0.805∗∗∗ 0.809∗∗∗ (0.181) (0.209) (0.211) (0.214) (0.215) (0.215) RTA 0.485∗∗∗ 0.160† 0.144† 0.187∗ 0.188∗ 0.186∗ (0.124) (0.082) (0.082) (0.085) (0.085) (0.084) Shared Currency 0.694∗∗∗ 2.559∗∗∗ 2.634∗∗∗ 1.956∗∗∗ 1.944∗∗∗ 1.939∗∗∗ (0.160) (0.142) (0.146) (0.162) (0.163) (0.162) Linguistic Distance −0.132∗∗ −0.116∗∗ −0.155∗∗ −0.091† −0.078 −0.083† (0.042) (0.044) (0.048) (0.049) (0.050) (0.049) Genetic Distance 0.081 0.227∗∗∗ 0.154∗ −0.006 0.013 0.000 (0.092) (0.061) (0.061) (0.071) (0.073) (0.071) LDE, origin – −0.369∗∗∗ −0.522∗∗∗ −0.788∗∗∗ −0.804∗∗∗ −0.834∗∗∗ (0.039) (0.047) (0.062) (0.063) (0.066) LDE, dest – −0.378∗∗∗ −0.285∗∗∗ −0.329∗∗∗ −0.340∗∗∗ −0.364∗∗∗ (0.044) (0.050) (0.068) (0.069) (0.071) LDE(1) –––– 0.211 – (0.146) LDE(2) ––––– 0.217∗ (0.100) Origin-Year FE Yes No No No No No Destination-Year FE Yes No No No No No Year FE No Yes Yes Yes Yes Yes Continent FE No No Yes No No No World Region FE No No No Yes Yes Yes

R2 0.767 0.580 0.601 0.640 0.641 0.641 Observations 8,277 8,277 8,277 8,277 8,277 8,277 Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral cross-border asset stocks. CHAPTER 4. LANGUAGE AND CULTURAL BARRIERS 113

Table 4.10: Effect of the Linguistic Distance toward English on Bilateral Banking Claims

Baseline RE-Model RE-Model incl. LDE

Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE Coef/StdE

ln Pop, origin – 0.578∗∗∗ 0.603∗∗∗ 0.587∗∗∗ 0.583∗∗∗ 0.588∗∗∗ (0.027) (0.031) (0.032) (0.032) (0.032) ln Pop, dest – 0.569∗∗∗ 0.578∗∗∗ 0.573∗∗∗ 0.573∗∗∗ 0.573∗∗∗ (0.019) (0.019) (0.019) (0.019) (0.019) ln GDP/pc, origin – 1.224∗∗∗ 1.181∗∗∗ 0.927∗∗∗ 0.928∗∗∗ 0.927∗∗∗ (0.063) (0.069) (0.097) (0.098) (0.097) ln GDP/pc, dest – 0.766∗∗∗ 0.759∗∗∗ 0.753∗∗∗ 0.753∗∗∗ 0.753∗∗∗ (0.034) (0.037) (0.039) (0.039) (0.039) ln Geo. Distance −1.043∗∗∗ −0.568∗∗∗ −0.586∗∗∗ −0.784∗∗∗ −0.785∗∗∗ −0.783∗∗∗ (0.066) (0.049) (0.065) (0.064) (0.064) (0.064) Shared Border −0.036 0.837∗∗∗ 0.828∗∗∗ 0.247 0.221 0.251 (0.200) (0.212) (0.219) (0.206) (0.209) (0.205) Shared Legal 0.149∗ 0.261∗∗∗ 0.311∗∗∗ 0.265∗∗∗ 0.252∗∗∗ 0.269∗∗∗ (0.059) (0.074) (0.076) (0.072) (0.073) (0.072) Colonial Relationship 1.066∗∗∗ 1.032∗∗∗ 0.918∗∗∗ 1.156∗∗∗ 1.174∗∗∗ 1.151∗∗∗ (0.106) (0.128) (0.137) (0.131) (0.133) (0.131) RTA −0.150 −0.031 −0.037 −0.039 −0.039 −0.038 (0.111) (0.060) (0.061) (0.061) (0.061) (0.061) Shared Currency −0.143 1.272∗∗∗ 1.280∗∗∗ 1.058∗∗∗ 1.056∗∗∗ 1.059∗∗∗ (0.163) (0.155) (0.155) (0.151) (0.151) (0.151) Linguistic Distance −0.115∗∗∗ −0.162∗∗∗ −0.156∗∗∗ −0.102∗ −0.094∗ −0.104∗ (0.033) (0.041) (0.043) (0.041) (0.042) (0.041) Genetic Distance 0.077 −0.065 0.017 −0.115∗ −0.110∗ −0.115∗ (0.091) (0.043) (0.046) (0.054) (0.054) (0.054) LDE, origin – −0.028 −0.121∗∗∗ −0.227∗∗∗ −0.238∗∗∗ −0.218∗∗∗ (0.028) (0.034) (0.042) (0.043) (0.045) LDE, dest – −0.427∗∗∗ −0.394∗∗∗ −0.454∗∗∗ −0.462∗∗∗ −0.448∗∗∗ (0.047) (0.052) (0.067) (0.068) (0.068) LDE(1) –––– 0.122 – (0.114) LDE(2) ––––– −0.036 (0.066) Origin-Year FE Yes No No No No No Destination-Year FE Yes No No No No No Year FE No Yes Yes Yes Yes Yes Continent FE No No Yes No No No World Region FE No No No Yes Yes Yes

R2 0.818 0.561 0.571 0.630 0.630 0.630 Observations 33,984 33,984 33,984 33,984 33,984 33,984 Notes: – Signiﬁcant at: ∗∗∗0.1% level; ∗∗1% level; ∗5% level; †10% level. – Cluster-robust standard errors are reported in parentheses. – The dependent variable is the logarithm of bilateral consolidated international banking claims. 114

Chapter 5

The Role of Language Skills in the German Labor Market∗

5.1 Introduction

The societal and economic integration of immigrants relies heavily on their investment in host country-specific human capital, such as language skills (Chiswick, 1978). Since the future value of an investment in human capital is generally unknown, capital market imperfections may induce a potentially severe underinvestment. Friedman (1962) was the first to describe the underinvestment in human capital as a result of capital market failure in the context of higher education. Unfortunately, we know very little about the private and social costs and returns to host country language skills. Against this background, this paper studies the private returns to language skills in the German labor market. It appears likely that immigrants underinvest in language training because they are credit-constrained and/or unwilling to take financial risks associated with the investment. There may be considerable efficiency losses resulting from an underinvestment in language training if private returns to language skills are high. Knowledge about the extent to which language skills affect labor market outcomes (such as employment probabilities and wages) may improve our understanding of the underlying mechanisms of a successful integration of immigrants and illustrate the need and scope for government intervention (such as the provision of student loans or free language courses). Germany is an excellent example for the analysis of the effects of language skills on labor market outcomes. During the 1960s and early 1970s immigration to Germany

∗Co-authored with Ingo E. Isphording (IZA) and Mathias Sinning (Australian National University, RWI, IZA). A revised version of this chapter is available from the authors. The authors are grateful to the participants of the ESPE 2014 and the International German Socio-Economic Panel User Conference 2014 for helpful comments and suggestions. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 115

was characterized by the active recruitment of so-called “guest workers”, mainly from Southern European countries (Zimmermann, 1995). Although these guest workers were notionally temporary, many of them decided to stay in Germany permanently (Bauer et al., 2005). With the beginning of the recession following the oil crisis in 1973, the active recruitment of guest workers was stopped and immigration to Germany was restrained. Family reunification and asylum became the major channels of legal immigration during the period from 1973 to 1988 (Bauer et al., 2005). The dissolution of socialism in 1989 lead to an inflow of immigrants from Central and Eastern Europe, many of them ethnic Germans who received German citizenship immediately. Germany further experienced a strong inflow of refugees and asylum seekers during the civil war in Yugoslavia in the early 1990s (Zimmermann, 1995). Restrictive administrative regulations were responsible for a decline in the number of ethnic Germans arriving in Germany during the 1990s (Münz and Ohliger, 1998). Despite gradual international labor mobility resulting from the European freedom of movement, immigration to Germany continued to decline during the 2000s, suggesting a strong impact of the remaining cultural and linguistic barriers between EU member states on international migration (Belot and Ederveen, 2012). In 2008 and 2009, net migration to Germany turned negative but bounced back in 2010 as a result of an increase in the number of new arrivals from euro crisis countries (mainly Greece and Spain). Compared to the extensive literature on the private returns to education, we know relatively little about the returns to language skills. Several studies provide empirical evidence on the positive association between language skills and wages of immigrants (Aldashev et al., 2009; Carliner, 1981; Chiswick, 1991; Kossoudji, 1988; McManus et al., 1983; Robinson, 1988; Tainer, 1988). Unfortunately, the identification of a causal effect of language skills on labor market outcomes is challenging because language skills and labor market outcomes are both determined by unobserved individual ability. Moreover, empirical studies often rely on self-reported language skill measures that are prone to measurement error. Due to these problems, several studies have employed instrumental variables to identify the causal effect of language skills on labor market outcomes (Angrist and Lavy, 1997; Bleakley and Chin, 2004, 2010; Chiswick and Miller, 1995; Dustmann and van Soest, 2002; Isphording and Sinning, 2012). To address endogeneity issues arising from unobserved heterogeneity and measurement error, we construct an instrument based on the relationship between immigrants’ duration of residence in their host country and their language skills, taking into account heterogeneity in the linguistic distance among immigrants from different countries of origin. Our approach is related to the empirical strategy of Bleakley and Chin (2004), which we extend in several directions. Most importantly, we use differences in the exposure to the host country CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 116

language resulting from differences in the linguistic origin of an immigrant to identify our parameter of interest. The usage of an innovative measure of linguistic distance allows us to identify the returns to language skills even in the absence of a comparison group of immigrants speaking the host country language as mother tongue.1 The linguistic distance, i.e., the dissimilarity between languages in terms of pronunciation, grammar, script, etc., plays a critical role in the acquisition of additional languages. The linguistic distance determines the efficiency of learning new languages and therefore constitutes and important barrier to international migration, which affects both the individual decision to migrate (Adsera and Pytlikova, 2012) and the integration of immigrants in their host country (Chiswick and Miller, 1999; Isphording and Otten, 2013). We find that the effect of language skills on employment probabilities is insignificant, which is in line with the economic literature on residential segregation of immigrants. In particular, it seems likely that geographic clustering allows immigrants to find jobs even without knowledge of the host country language. At the same time, geographic clustering may reduce incentives for immigrants to invest in country-specific human capital that is required for finding a higher-wage job in the larger labor market of the host country (Borjas, 2000). It is therefore not surprising that we find a significantly positive effect of language skills on wages. However, this effect diminishes when we control for occupation, indicating that the returns to language skills are a result of the sorting of immigrants across occupations. We further demonstrate that simple OLS regressions systematically underestimate the positive effects of language skills on wages. This result is consistent with the empirical findings of Bleakley and Chin (2004) who conclude that the downward bias of the OLS estimate caused by measurement error is much stronger than the upward bias caused by unobserved ability. The remainder of the paper is structured as follows. Section 5.2 provides a description of the data, including the linguistic distance measure employed to construct our instrument. Our empirical strategy is explained in Section 5.3. Section 5.4 discusses the findings obtained from our empirical analysis and Section 5.5 concludes.

5.2 Data

Our empirical analysis is based on data from the German Socio-Economic Panel (SOEP), a long-run and large scale micro data set covering nearly 11,000 households and more than 20,000 persons per wave. The SOEP started in 1984 and consists of native- and foreign-born persons residing in West and (since 1990) East Germany. Since language

1Our sample only contains a very small subsample of German-speaking immigrants from Austria and Switzerland (176 observations), which we do not consider as a convincing comparison group. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 117

skills were surveyed every second year from 1993 to 2007 and each year after 2007, our sample consists of 12 waves surveyed over the period 1993 to 2011.2 We restrict the sample to foreign-born individuals who arrived in Germany after 1948 and focus on (employed or registered unemployed) labor force participants aged between 25 and 62 years who are not in military service. After deleting observations with missing information on one of the variables used in our analysis, our sample of labor force participants consists of 11,789 (10,282 employed and 1,507 unemployed) person-year observations. We consider two labor market outcomes in our analysis: a binary variable indicating whether the observed person is employed or registered unemployed, and (the logarithm of) hourly wages if the observed person is employed.3 We further consider two self-reported measures of oral and written fluency in German, which are placed on a 4-point scale, ranging from 0 (“Not at all/Fairly bad”) to 3 (“Very Good”). Our set of control variables includes the number of years of education, actual work experience (full- and part-time), years since migration, and indicator variables for women, immigration cohorts (1969-73, 1974-89, 1990-2011), and (in our wage regression) occupations. Since our estimation strategy relies on differences in language acquisition profiles that arise from the linguistic distance between German and mother tongue, we combine data from the SOEP with information on the linguistic distance between German and the predominant language of a migrants’ country of origin.4 The influence of the linguistic distance on the efficiency of learning new languages has been thoroughly analyzed in linguistic case studies (see Llach (2010) for an overview). Differences between words and sounds are harder to judge if two languages exhibit larger differences in pronunciation. Consequently, languages are more costly to acquire (in terms of personal effort) if they exhibit fewer commonalities with regard to pronunciation. A number of economic studies aims at quantifying the notion of linguistic distance to analyze the effect of the linguistic background beyond small scale case studies. Chiswick and Miller (1999) and following work introduced a test-score-based measure of linguistic distance, which is restricted to Anglophone countries due to data limitations. A more general approach has been used by Isphording and Otten (2013) to explain heterogeneity

2The data used in this paper was extracted using the Add-On package PanelWhiz for Stata®. Panel- Whiz (http://www.panelwhiz.eu) was written by Dr. John P. Haisken-DeNew ([email protected]). See Haisken-DeNew and Hahn (2006) for details. The PanelWhiz generated DO file to retrieve the data used here is available from the authors upon request. Any data or computational errors in this paper are our own. 3Hourly wages are defined as personal monthly gross wages divided by monthly working hours. We define monthly working hours as weekly working hours × 4.3 and replace contractual working hours by actual working hours if the number of actual working hours exceeds the number of contractual working hours. 4While assigning a mother tongue to monolingual countries of origin is straightforward, we use the most prevalent language in multilingual countries of origin to determine the linguistic distance. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 118 in language skills in Germany, Spain, and the US. This approach is based on linguistic research, the Automatic Similarity Judgment Program (ASJP) of the German Max Planck Institute of Evolutionary Anthropology. While the original purpose of this approach has been the analysis of the geographic distribution and historical development of languages, it allows us to obtain a descriptive, continuous, and cardinally interpretable measure of the degree of dissimilarity between languages. The ASJP provides a simple way of deriving a measure of linguistic distance that is computed by comparing the pronunciation of words with the same meaning in different languages. The words are taken from the ‘Swadesh list’ (Swadesh, 1952), a list of deductively derived words, including words of basic communication, and words describing objects of everyday life. This list of words is believed to be culturally independent and can be found in almost all languages. The 40-item sublist used in our empirical analysis is presented in the upper panel of Table 5.1. The ASJP approach relies on a comparison of pronunciations of words. The words of the Swadesh list are expressed in a phonetic script. Then the Levenshtein distance, a simple measure of dissimilarity between strings, is computed for each word pair sharing the same meaning in two languages. Finally, a normalized average is taken across all word pairs to derive a continuous measure. The linguistic distance measure can be interpreted as a percentage measure of similarity between languages. By construction, values beyond 100% can occur if languages do not even possess similarities that are expected to exist by chance. Bakker et al. (2009) provide a more detailed description of the approach. The lower panel of Table 5.1 lists some examples of differences between German and English. The five closest and five furthest languages with regard to their linguistic distance from German are presented in Table 5.2. The closest languages are members of the Germanic language family. The furthest languages include Asian and African languages. Table 5.3 presents the sample means and standard deviations of the variables employed in our analysis by oral language fluency for the two samples used to study the respective effects of language skills on employment (Panel A) and wages (Panel B). The numbers reveal that both employment rates and wages are positively related to language skills. We further observe an increase in average years since migration across the language skill distribution, indicating that immigrants improve their language skills over the duration of residence in Germany. There is also a positive association between language skills and education, but no distinctive experience-language skill profile. Finally, we observe a relatively low linguistic distance among immigrants who report to speak German “Very good”, providing evidence for a negative relationship between linguistic distance and German language fluency, which is consistent with previous findings of Isphording and Otten (2013). CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 119 5.3 Empirical Strategy

The aim of our empirical analysis is to determine the effects of language skills on labor market outcomes. Unfortunately, endogeneity problems complicate the identification of causal effects. We address these problems by employing an instrumental variable strategy based on differences in assimilation profiles between immigrants with different language backgrounds. We may formalize the endogeneity problems by relating the labor market outcome Yi of individual i to the number of years since migration, YSMi, a language skill measure, Li, and a set of control variables, Xi:

0 Yi = β0 + β1YSMi + β2Li + Xiβ3 + λc + εi, (5.1) where λc captures source country ﬁxed eﬀects and ε is a standard error term. We are mainly interested in obtaining an unbiased estimate of the parameter β2. Uncorrected

OLS regressions of equation (5.1) will yield a biased estimate of β2 for two reasons. First, it appears likely that self-reported language skill measures are prone to measurement error, which may cause a downward bias of the OLS parameter of language skills in form of an attenuation bias (Charette and Meng, 1994; de Coulon and Wolff, 2007; Dustmann and van Soest, 2001; Dustmann and van Soest, 2002). Dustmann and van Soest (2002) note that measurement error in subjective language indicators may either reflect white noise or be related to a systematic unwillingness or inability of individuals to report the own language ability. Second, it appears likely that the acquisition of language skills is influenced by the unobserved ability of an individual, causing an upward bias of the uncorrected OLS parameter, which consists of a composite effect of ability and language skills. Following the notation of Dustmann and van Soest (2002), we may write equation (5.1) as:

0 Yi = β0 + β1YSMi + β2Li + Xiβ3 + λc + θi + νi, (5.2) where ε is replaced by a composite error term including heterogeneity in unobserved components θ and a random error term ν. Moreover, since our ability measure Li suﬀers from measurement error, we only observe the subjective indicator Lei:

Lei = Li + ηi. (5.3)

Combining equations (5.2) and (5.3) yields

0 Yi = β0 + β1YSMi + β2Lei + Xiβ3 + (θi − β2ηi + νi). (5.4) CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 120

The composite error term (θi − β2ηi + νi) now consists of an individual fixed effect, θi, a potential measurement error, β2ηi, and a random error term, νi. A simple OLS regression of equation (5.4) will only produce an unbiased estimate of β2 if Lei is not only uncorrelated with νi but also with θi and β2ηi, which is very unlikely. Consequently, we have to consider the possibility that the OLS estimate of β2 is biased and that the direction of the bias is a priori unclear. We propose an instrumental variable (IV) approach that allows us to deal with both sources of endogeneity. Specifically, we exploit differences in exposure to the host country language as a source of exogenous variation. Although exposure is measured by years since migration, the time of residence in the host country cannot be used as an instrument because this variable captures additional factors of assimilation, such as acculturation and acquisition of host country-specific human capital apart from language skills. Instead, we propose to differentiate between these non-linguistic factors and the linguistic effect of exposure. Technically, we use the interaction between the duration of residence in

Germany, YSMi, and the linguistic distance between the language of the source country and German, ∆i, to partial out the non-linguistic factors of assimilation. Using this interaction term as an instrument for language skills, the ﬁrst stage equation of the IV approach is given by:

0 Lei = γ0 + γ1YSMi + γ2YSMi · ∆i + Xiγ3 + λc + ωi. (5.5)

Before turning to the instrument, it is important to mention that our model includes YSMi as a control variable and picks up source country-specific variation (including variation in linguistic distance) by the inclusion of source country fixed effects. The variation we seek to exploit to identify the effects of language skills on labor market outcomes comes from the interaction between duration of residence in Germany and linguistic distance. In order to understand the intuition behind our instrument, it is useful to consider the simple case in which ∆i is a dummy variable indicating whether the mother tongue of individual i is close to German or not. In this case, the underlying identification assumption is that differences in assimilation profiles between immigrants facing high and low language barriers affect labor market outcomes exclusively through language skills. The comparison of immigrants from countries with high and low language barriers allows us to difference out non-linguistic factors of their assimilation profiles. This approach is consistent with Bleakley and Chin (2004), who study the effect of language skills on wages of child migrants and interact age at arrival with a source country language indicator. Figure 5.1 provides a graphical explanation for the construction of our instrument. The graphs show 5-year moving averages for the lower and the upper quartile of linguistic CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 121

distance. Figure 5.1(a) reveals that immigrants from linguistically close countries exhibit considerably higher German skills and show a steeper learning curve particularly during the first 10 years of residence in Germany compared to immigrants from linguistically more distant countries of origin. After 10-15 years of residence, the slope of the learning curve of the first group decreases and German skills converge. Figures 5.1(b) and 5.1(c) show that the convergence in German skills between the two groups is translated into a convergence in employment status and wages, respectively. Our instrument seeks to exploit this convergence and assumes that differences in the assimilation profiles are entirely attributable to German skills. Since immigrants with different language backgrounds face different costs of acquiring German, we use our linguistic distance measure instead of a dummy variable indicating language background to construct our instrument. The following section presents the results of employment and wage regressions, using oral and written German proficiency as language measures. To obtain a sufficiently large sample, we use a pooled sample consisting of 12 waves surveyed between 1993 and 2011. All regressions include time fixed effects and all standard errors were adjusted to take repeated observations of the same individual into account.

5.4 Results

Table 5.4 summarizes the results of the employment regressions. The estimates presented in Column (1) provide evidence for a significantly positive effect of our instrument on language acquisition, reflecting that individuals from distant language backgrounds face an initial disadvantage but catch up over the settlement process.5 The uncorrected OLS regressions (Column (2)) suggest that both oral (Panel A) and written (Panel B) language proficiency increase the probability of being employed by about 4-6 percentage points. In contrast, the estimates obtained from the IV approach (Column (3)) reveal that the causal effects of oral and written language skills on employment are not significant. Controlling for individual years of education (Columns (4) to (6)) does not alter this picture. The results presented in Table 5.4 are not consistent with those of Aldashev et al. (2009) who find a positive association between German language usage at home and the probability of being employed. However, the lack of a significant effect of language skills on employment is in line with the notion that immigrants residing in ethnic and linguistic enclaves have access to segments of the labor market that do not require fluency in German. At the same time, it appears likely that immigrants may earn higher wages in the larger German labor market if they are able to speak German. Tables 5.5 and

5Comparing the F-test of excluded instruments (following Angrist and Pischke, 2009) with the critical values tabulated by Stock and Yogo (2005) does not indicate any weak instrument issues. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 122

5.6, which summarize the respective estimates of the wage regressions based on oral and written language proficiency, confirm this hypothesis. Specifically, the base results in Columns (1) to (3) of Tables 5.5 and 5.6 reveal a significant wage premium resulting from language proficiency, which exceeds the low, though significant, partial correlation from uncorrected OLS regressions considerably. Although an interpretation of the magnitude of the estimated coefficients is challenging due to the ordinal character of the subjective language skill measure, the estimated relationship highlights the importance of language skills in the wage formation process of immigrants.6 The parsimonious base specification implies a very broad interpretation of the causal effect of language skills, which may affect wages through various channels, although not all of them are empirically identifiable. We consider education and occupation as two potential channels through which language skills may affect wages. Columns (4) to (6) of Tables 5.5 and 5.6 include the estimates obtained from a model in which the base specification is augmented with the number of years of education. Including education as an additional explanatory variable affects the interpretation of our results. Since educational attainment may be a function of language skills itself, the inclusion of education into our model interferes with language skills but allows us to assess the relevance of this channel. The inclusion of education leads to a marginal reduction in the point estimate of the returns to language skills, suggesting that the effect of language skills on wages does not depend on education. The estimates in Column (4) indicate that educational attainment has an effect on the formation of language skills but, since our sample mainly consists of immigrants who arrived in adulthood, does not depend on host country language skills and therefore does not act as a mediator between language skills and wages. Columns (7) to (9) of Tables 5.5 and 5.6 summarize the results obtained from a model in which we add nine occupational categories to the base specification. Tainer (1988) and Kossoudji (1988) stress that heterogeneity in language skill requirements across jobs may create a link between returns to language skills and occupations. In fact, we observe a heterogenous distribution of language skills across occupations. For example, only 1.3% of individuals currently employed as “managers” (ISCO sub-code 1) report to have no or fairly bad oral German proficiency, compared to 13% of individuals employed in “elementary occupations” (ISCO sub-code 9). Similar to the model in which we control for education, occupations may represent a function of language skills. Following Chiswick and Miller (2010), the overall effect of language skills of the base model may be interpreted as a composite effect including a direct wage effect resulting from skill premia and an indirect effect mediated through occupational sorting into better paid occupations. We find

6Similar to Table 5.4, F-tests of excluded instruments exceeding the critical Stock/Yogo-values provide evidence in support of the identifying instrument. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 123 that partialling out the influence of occupational sorting by controlling for occupational categories reduces the point estimate of the coefficient of language skills by about 20%, highlighting the importance of occupational sorting as a mediator between language skills and wages. Interestingly, the deviations of partial correlations from the estimated effects in employment and wage regressions have opposite signs. While results obtained from the employment regressions indicate an overestimation in uncorrected OLS regressions, the partial correlations between wages and language skills seem to understate the ‘true’ causal effect of language skills on wages. Against the background of ambiguous biases discussed in Section 5.3, employment regressions seem to suffer from a potential ability bias arising from a correlation between language skills and unobserved ability. By contrast, the attenuation bias induced by noisy subjective indicators seems to outweigh the ability bias in our wage regressions. A potential explanation for this result is the predominance of occupation-wide payment regulations resulting from wage agreements, which reduce the importance of unobserved ability in the wage formation process.

5.5 Conclusions

Labor market effects of language skills in terms of wage returns and employment probabilities are highly relevant in the human capital formation process of immigrants. Large private returns to language skills provide strong incentives to invest in host country-specific human capital. Against this background, we investigate the effects of language skills on labor market outcomes of immigrants in Germany. To address endogeneity problems arising from unobserved heterogeneity and measurement error, we employ an instrumental variables strategy that exploits differences in language skill acquisition patterns caused by differences in linguistic origin as a source of exogenous variation. Specifically, we use an interaction term between a linguistic distance measure and the duration of residence in Germany as an instrument for subjective measures of oral and written language proficiency. Our findings reveal that the effect of language skills on employment probabilities is insignificant. In contrast, we observe a significantly positive effect of language skills on wages, which decreases when we control for occupation, suggesting that the returns to language skills are a result of the sorting of immigrants across occupations. Our findings further reveal that uncorrected OLS regressions overestimate the effects of language skills on employment but underestimate the effects on wages. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 124 Tables and Figures

Table 5.1: Swadesh 40-Item List with Computational Examples I You We One Two Person Fish Dog Louse Tree Leaf Skin Blood Bone Horn Ear Eye Nose Tooth Tongue Knee Hand Breast Liver Drink See Hear Die Come Sun Star Water Stone Fire Path Mountain Night Full New Name

Word English German Distance fish fiS fiS 0 breast brest brust 1 hand hEnd hant 2 tree tri baum 4 Mountain maunt3n bErk 7 Notes: – Adapted from Brown (2008).

Table 5.2: Distance from German – Closest and Furthest Languages Closest Furthest Language Distance Language Distance Luxembourgish 42.12 Mandarin 102.88 Dutch 51.50 Yogyakarta 103.00 Westvlaams 57.86 Yoruba 103.58 Swedish 66.56 Palistinian Arabic 103.72 Danish 66.96 Korean 104.30 Notes: – Own calculations using programs for calculating ASJP distance matrices (Version 2.1), see http:// email.eva. mpg.de/ ~wichmann/ software.htm. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 125

Table 5.3: Descriptive Statistics by Oral German Ability Speaks German Speaks German Speaks German Speaks German bad/not at all not bad good very good Total (0) (1) (2) (3)

Panel A: Employment Sample Employed Yes/No 0.859 0.738 0.804 0.864 0.910 (0.348) (0.440) (0.397) (0.343) (0.286) Years since migration 22.670 18.442 20.049 22.426 25.281 (10.891) (9.958) (10.385) (10.378) (11.190) Female 0.444 0.389 0.403 0.416 0.508 (0.497) (0.488) (0.491) (0.493) (0.500) Work experience 18.891 19.352 19.348 19.283 18.145 (10.849) (10.962) (11.040) (10.543) (10.973) Years of education 10.938 9.515 9.974 10.802 11.920 (2.548) (2.213) (2.106) (2.329) (2.674) Linguistic distance 90.626 95.007 95.230 93.444 84.269 (17.057) (6.761) (4.607) (9.016) (25.311) Number of observations 11,789 944 2,777 4,125 3,943

Panel B: Wage Sample Log hourly wages 2.425 2.224 2.309 2.426 2.516 (0.523) (0.456) (0.487) (0.479) (0.568) Years since migration 22.621 18.801 19.574 22.242 25.189 (10.766) (9.587) (10.020) (10.218) (11.172) Female 0.441 0.383 0.398 0.416 0.497 (0.497) (0.486) (0.490) (0.493) (0.500) Work experience 19.041 18.933 19.668 19.503 18.292 (10.608) (10.600) (10.882) (10.239) (10.758) Years of education 11.046 9.615 10.005 10.836 12.017 (2.551) (2.278) (2.064) (2.309) (2.671) Linguistic distance 90.127 94.473 94.964 93.144 84.051 (17.787) (7.597) (4.665) (9.583) (25.748) Number of observations 9,790 676 2,189 3,459 3,466

Notes: – Standard deviations reported in parentheses. Weighted numbers based on weights provided by the SOEP. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 126

(a) Average German Ability (b) Average Employment Rate 1 2.5 2 .9 1.5 .8 Employment Rate Oral German Proficiency 1 .7 .5 0 5 10 15 20 25 0 5 10 15 20 25 Years since Migration Years since Migration

2.1 Linguistic Distance, Lower quartile

2 Linguistic Distance, Upper quartile

0 5 10 15 20 25 Years since Migration

Figure 5.1: German Ability, Employment Rate, and Hourly Wages by Years since Migration (5-Year Moving Average) CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 127

Table 5.4: Results of Employment Regressions Base Controlling for Education German Employment German Employment Ability Status Ability Status OLS OLS 2SLS OLS OLS 2SLS (1) (2) (3) (4) (5) (6)

Panel A: Results for Oral Ability Endogenous regressor: Oral German proﬁciency – 0.056∗∗∗ −0.028 – 0.051∗∗∗ −0.040 (scale of 0 to 3, 3=best) (0.006) (0.048) (0.006) (0.052) Identifying instrument: Years since migration × 0.040∗∗∗ –– 0.036∗∗∗ –– Linguistic distance/100 (0.005) (0.005) Controls: Years since migration 0.019∗∗∗ −0.001 0.003 0.021∗∗∗ −0.001 0.003 (0.005) (0.001) (0.003) (0.005) (0.001) (0.003) Female −0.108∗∗∗ −0.013 −0.021∗ −0.070∗∗ −0.009 −0.014 (0.031) (0.010) (0.011) (0.030) (0.010) (0.010) Work experience −0.021∗∗∗ 0.000 −0.001 −0.018∗∗∗ 0.001 −0.001 (0.002) (0.001) (0.001) (0.001) (0.001) (0.001) Years of education – – – 0.076∗∗∗ 0.009∗∗∗ 0.016∗∗∗ (0.006) (0.002) (0.005) Country-of-birth dummies Yes Yes Yes Yes Yes Yes Immigration cohort dummies Yes Yes Yes Yes Yes Yes Year dummies Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.288 0.057 0.316 0.060 F-Test of excluded instruments 70.79 45.60 Number of observations 11,789 11,789 11,789 11,789 11,789 11,789

Panel B: Results for Written Ability Endogenous regressor: Written German proﬁciency – 0.045∗∗∗ −0.030 – 0.040∗∗∗ −0.044 (scale of 0 to 3, 3=best) (0.005) (0.051) (0.005) (0.058) Identifying instrument: Years since migration × 0.039∗∗∗ –– 0.033∗∗∗ –– Linguistic distance/100 (0.006) (0.007) Controls: Years since migration 0.028∗∗∗ −0.001 0.003 0.031∗∗∗ −0.001 0.004 (0.006) (0.001) (0.003) (0.007) (0.001) (0.003) Female −0.124∗∗∗ −0.013 −0.021∗ −0.067∗∗ −0.010 −0.014 (0.036) (0.010) (0.011) (0.034) (0.010) (0.011) Work experience −0.030∗∗∗ 0.001 −0.002 −0.025∗∗∗ 0.001 −0.001 (0.002) (0.001) (0.002) (0.002) (0.001) (0.002) Years of education – – – 0.116∗∗∗ 0.008∗∗∗ 0.018∗∗ (0.007) (0.002) (0.007) Country-of-birth dummies Yes Yes Yes Yes Yes Yes Immigration cohort dummies Yes Yes Yes Yes Yes Yes Year dummies Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.363 0.054 0.408 0.057 F-Test of excluded instruments 45.76 21.32 Number of observations 11,789 11,789 11,789 11,789 11,789 11,789

Notes: – ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. – Standard errors reported in parentheses are clustered at the individual level. – Linguistic distance assigned by country of birth. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 128 ∗∗∗ ∗∗∗ ∗ 377 220) 013 011) 214 030) 012 004) ...... 0 0 0 0 − − ∗∗∗ ∗∗∗ ∗∗∗ ∗ –– 313 034 007) (0 003 002) (0 250 017) (0 006 001) (0 ...... 0 0 0 0 − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – ––– 018 002) (0 325 0 006) 033 020 006) (0 112 032) (0 ...... 0 0 0 − − ∗∗ ∗∗∗ ∗∗∗ 238 025) (0 014 004) (0 024 016) 440 211) (0 015011) 0 (0 ...... 0 0 0 0 0 − − ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ –– 002) (0 265 016) (0 006 001) (0 054 004) (0 280 0 036 007) (0 005 ...... 0 0 0 0 0 − ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ – 006) (0 019 036 006) (0 076 031) (0 018 002) (0 073 006) (0 318 0 ...... 0 0 0 0 (0 − − ∗∗ ∗∗∗ ∗∗∗ 479 195) (0 017 0 011) (0 246 029) (0 013 004) (0 ...... 0 0 0 0 − − ∗∗∗ ∗∗∗ ∗ ∗∗∗ –– 064 008) (0 004 002) (0 290 017) (0 004 001) (0 229 0 ...... Base Controlling for Education Controlling for Occupation 0 0 0 0 (0 Results of Wage Regressions – Oral Ability − . – Standard errors reported in parentheses are clustered at the individual level. – Linguistic distance assigned 01 . ∗∗ ∗∗∗ ∗∗ ∗∗∗ 0 (1) (2) (3) (4) (5) (6) (7) (8) (9) 002) (0 291 0 006)014 006) (0 (0 043 114 032) (0 022 OLS OLS 2SLS OLS OLS 2SLS OLS OLS 2SLS ...... p < 0 0 0 0 0 Ability Hourly Wages Ability Hourly Wages Ability Hourly Wages German Log German Log German Log (0 (0 (0 (0 − − ∗∗∗ Table 5.5: , 05 . 0 × p < ∗∗ , 1 . 0 p < ∗ Oral German proﬁciency – (scale of 0 to 3, 3=best) Years since migration Linguistic distance/100 Years since migration Female Work experience Years of education – – – Occupational dummiesCountry-of-birth dummiesImmigration cohort dummiesYear dummies Yes Yes No Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Adjusted R-squared F-Test of excluded instrumentsNumber of observations 9,790 9,790 9,790 56.91 9,790 9,790 9,790 9,790 35.39 9,790 9,790 29.72 Endogenous regressor: Controls: Identifying instrument: Notes: – by country of birth. CHAPTER 5. LANGUAGE SKILLS IN THE GERMAN LABOR MARKET 129 ∗∗∗ ∗∗ 423 271) 019 016) 193 044) 016 007) ...... 0 0 0 0 − − ∗∗∗ ∗∗∗ ∗∗∗ –– 314 034 007) (0 003 002) (0 249 017) (0 006 001) (0 ...... 0 0 0 0 − ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ – ––– 027 002) (0 414 0 008) 030 032 008) (0 148 037) (0 ...... 0 0 0 − − ∗ ∗∗∗ ∗∗∗ 231 029) (0 018 006) (0 005 028) 458 242) (0 021015) 0 (0 ...... 0 0 0 0 0 − − ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ –– 002) (0 265 016) (0 007 001) (0 053 004) (0 281 0 036 007) (0 005 ...... 0 0 0 0 0 − ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ – 008) (0 031 034 008) (0 087 036) (0 026 002) (0 112 007) (0 414 0 ...... 0 0 0 0 0 (0 − − ∗∗ ∗ ∗∗∗ ∗∗∗ 466 200) (0 022 013) (0 233 035) (0 018 006) (0 ...... 0 0 0 0 − − ∗∗∗ ∗∗∗ ∗∗∗ –– 068 007) (0 003 002) (0 287 016) (0 005 001) (0 234 0 ...... Base Controlling for Education Controlling for Occupation 0 0 0 0 (0 − Results of Wage Regressions – Written Ability . – Standard errors reported in parentheses are clustered at the individual level. – Linguistic distance assigned 01 . ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ 0 (1) (2) (3) (4) (5) (6) (7) (8) (9) 002) (0 371 0 007)023 007) (0 (0 044 146 038) (0 032 OLS OLS 2SLS OLS OLS 2SLS OLS OLS 2SLS ...... p < 0 0 0 0 0 Ability Hourly Wages Ability Hourly Wages Ability Hourly Wages German Log German Log German Log (0 (0 (0 (0 − − ∗∗∗ , 05 Table 5.6: . 0 × p < ∗∗ , 1 . 0 p < ∗ Written German proﬁciency – (scale of 0 to 3, 3=best) Years since migration Linguistic distance/100 Years since migration Female Work experience Years of education – – – Occupational dummiesCountry-of-birth dummiesImmigration cohort dummiesYear dummies Yes Yes No Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Adjusted R-squared F-Test of excluded instrumentsNumber of observations 9,790 9,790 9,790 41.53 9,790 9,790 9,790 9,790 20.52 9,790 9,790 15.30 Endogenous regressor: Controls: Identifying instrument: Notes: – by country of birth. 130

Chapter 6

The Role of Source- and Host-Country Characteristics in Female Immigrant Labor Supply∗

6.1 Introduction

The first decade of the 21st century has seen large waves of migration to the EU Member States from both within the EU and from outside it. In 2008, 3.8 million people migrated to and between the EU-27 Member States. Moreover, the share of immigrants that migrate from countries with substantially different cultures and traditions toward the European origin population increases. From the 47.3 million immigrants living in EU Member States, about two-thirds were born outside the European Union, almost equally divided between America, Asia, Africa, and countries in Europe outside the EU-27 (European Commission, 2011).1 As many European countries face the problem of an aging population, which is expected to put downward-pressure on labor supply in the years to come, immigration is seen as a means of filling in current and future labor market needs and thus ensure economic sustainability and growth. As a result, the active recruitment of high-skilled immigrants on the one hand, and the integration of recent immigrants into the host-countries’ labor

∗Co-authored with Julia Bredtmann (Ruhr University Bochum, RWI). A preliminary version of this chapter is available as MPRA Paper No. 44544 and a revised version is available from the authors. The authors are grateful to Ronald Bachmann, Thomas K. Bauer, Ingo E. Isphording, Christoph M. Schmidt, Mathias Sinning, and seminar participants at Victoria University of Wellington, Motu Economic and Public Policy Research, the University of Otago, the Melbourne Institute of Applied Economic and Social Research, Aarhus University, the University of Salzburg, the NORFACE Migration Conference 2013, the SOLE 2013, the ESPE 2013, the EALE 2013, the AEA 2014, and the IEA World Congress 2014 for helpful comments and suggestions. 1For an overview of the history of immigration into Europe, see Bauer et al. (2000). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 131

markets on the other hand, have become important policy goals within Europe (European Commission, 2010b). However, although issues concerning the labor market integration of immigrants are high on the political agenda in many European countries, immigrants still show a significantly lower labor market attachment than the native population (European Commission, 2011). As a result, an intense political debate is taking place in Europe around migration issues with a focus on the costs and benefits of cultural diversity.2 The aspect of a low labor market attachment of immigrants is especially relevant for immigrant women. In 2008, the labor market participation of foreign-born women living within the EU-27 was nine percentage points below that for native-born women (69% as opposed to 78%). The lower overall participation rate of foreign-born women, however, is mainly due to the significantly lower activity rate of women originating from non-EU countries (67%), whereas women born in another EU country do hardly differ from natives (76%) (European Commission, 2011). The determinants of these differences in labor market participation across immigrants’ home-country groups remain an open question. Previous studies for immigrants in the U.S. suggest that differences in labor market behavior across immigrant women’s source countries can, at least partly, be explained by differences in female labor force participation rates (FLFPR) between these countries (Antecol, 2000; Blau and Kahn, 2011; Blau et al., 2011; Fernández and Fogli, 2009). The authors argue that differences in FLFPR across immigrants’ source-country groups reflect differences in the preferences and beliefs regarding women’s roles in family and society between these countries, and that these cultural differences ultimately affect the labor market behavior of immigrant women in their host country. Furthermore, they reveal that these cultural effects persist in the long run (Blau et al., 2011) and influence the labor supply behavior of second- and higher-generation women (Antecol, 2000; Fernández and Fogli, 2009). As Figure 6.1 shows, there is indeed a considerable variation in FLFPR across the world. In particular, pronounced differences between the developed countries, which experienced a steady increase in FLFP since the 1970s, and the less developed countries, which experienced no such trend so far, can be observed. Against this background, it is important to note that European immigrants increasingly come from countries that are characterized by relatively low FLFPR as compared to the European average, such as the Middle Eastern and Northern African countries. These cross-country differences in FLFPR

2Amongst others, German Chancellor Angela Merkel and British Prime Minister David Cameron recently called Europe’s approach to multiculturalism into question, thereby triggering a public controversy over the cultural integration of immigrants. While Angela Merkel said that the attempts to build a multicultural society in Germany have “failed, utterly failed” (BBC, 2010), David Cameron stated that the “doctrine of state multiculturalism” has failed and will no longer be state policy (BBC, 2011). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 132 may therefore help explain the heterogeneity in LFP of immigrant women in Europe. But not only do we observe large variations in FLFPR across immigrants’ source countries, we also observe large differences in FLFPR across European countries, ranging from 51.0 percent for Italy to 76.6 percent for Sweden in 2011. While previous evidence for the U.S. suggests that home-country FLFP influences immigrant women’s labor supply behavior in the host country, so far little is known about the role of host-country FLFP in immigrant women’s behavior. In particular, it is of interest whether immigrant women assimilate to the labor market behavior of native women, or whether their labor supply decisions are not affected by the FLFP in their host country at all. The aim of this paper is study the impact of source- and host-country characteristics on female immigrant labor supply. In our empirical analysis, we employ data from five rounds of the European Social Survey (ESS) covering immigrants in 26 European countries surveyed between 2002 and 2011. These data are augmented with an extensive set of aggregated source- and destination-country characteristics. We find that women who migrate from countries with relatively high levels of female labor supply have a higher probability of participating in the labor force in their respective host country. This positive effect of the FLFP in the (parents’) source country on women’s labor supply in their host country holds for second-generation immigrants as well. We are further able to show that most of this effect remains when controlling for the human capital of a woman’s partner, the past labor supply of her parents, and a variety of source-country characteristics that might be correlated with FLFPR. These results suggest that the culture and norms of the source country play an important role for immigrant women’s labor supply decisions. Moreover, we find evidence for an impact of host-country FLFPR on female immigrant labor supply, suggesting that immigrant women assimilate to the work behavior of natives. The remainder of the paper is organized as follows. The next section provides a brief overview of the literature on the role of culture in economic behavior and presents the results of former studies analyzing the labor supply of female immigrants. Section 6.3 describes the data used, provides some descriptive statistics and explains the identification strategy of our analysis. We present and discuss the main estimation results in Section 6.4, while the results of several robustness checks are presented in Section 6.5. The final section summarizes the results and discusses their implications.

6.2 Background

The present study contributes to the evolving literature on the impact of culture on social and economic behavior. In this strand of literature, differences in culture are broadly CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 133 interpreted as systematic variations in preferences and beliefs across time, space, or social groups (Fernández, 2011). The main difficulty in identifying the role of culture in economic behavior is to isolate it from those of the economic and institutional environment in which economic decisions are taken. A possible solution to this problem is brought about by what Fernández (2011) refers to as the epidemiological approach. The main idea of this approach is to identify the effect of culture through the variation in economic outcomes of individuals who share the same economic and institutional environment, but whose social beliefs are potentially different. One way to apply this approach is to focus on the economic behavior of immigrants. When individuals emigrate, they take some aspects of their culture with them and transmit them intergenerationally, while they live in the economic and formal institutional environment of the host country. Studying the economic behavior of immigrants from different countries of origin in their host country may therefore be a useful strategy to isolate culture from strictly economic and institutional effects. In this paper, we study the effect of culture on the labor supply of first- and second- generation female immigrants in Europe. In doing so, our study builds on research that has examined the effect of home-country characteristics on U.S. immigrant women’s labor supply.3 An early attempt to identify the effect of culture on immigrant labor supply is the study by Reimers (1985), who uses ethnic dummy variables to examine whether cultural factors play a direct role in married women’s LFP in the U.S. As Reimers’ dummy-variable approach does not allow for a quantification of these cultural effects, more recent studies address this limitation by using quantitative variables as proxies for culture. In particular, they use past values of FLFPR in the immigrant’s country of origin as a cultural proxy. As Fernández and Fogli (2009) point out, the main idea for using this aggregate variable is that it reflects the market work decisions of women in the source country, which (in addition to each woman’s individual characteristics) depend on the economic and institutional environment as well as the preferences and beliefs within the country. While the economic and (formal) institutional conditions of the country of origin should no longer be relevant for emigrated women, the preferences and beliefs embodied in this variable may still matter. Hence, if this aggregate variable has explanatory power for the variation in the labor market behavior of immigrant women, even after controlling for their individual economic attributes, only the cultural component of this variable can be responsible for this correlation. The first study to analyze the effect of source-country FLFPR on the work outcomes of female immigrants is the study by Antecol (2000), who finds the source-country FLFPR to

3The role of source-country variables, in different contexts, has been examined in several studies. For example, Borjas (1987) on the native/immigrant wage differential, Blau (1992) on the fertility behavior among first-generation immigrant women, and Antecol (2001) on the role of home-country variables in explaining variation in the gender wage gap across home-country groups within the U.S. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 134

be positively correlated with the LFP of first-generation immigrant women in the U.S. These findings, though weaker, even hold for second- and higher-generation immigrants. However, as Fernández and Fogli (2009) point out, these results might be driven by unobserved heterogeneity, as the analysis does not control for important individual characteristics such as years of education or parental background. In their study on the work and fertility behavior of U.S.-born daughters of immigrants to the U.S., Fernández and Fogli (2009) use various measures of average parental education and average education of the immigrant group to control for human capital factors. They find that the labor supply and fertility behavior of second-generation female immigrants is positively associated with both FLFPR and fertility rates in their parents’ country of origin. The authors also show that the husband’s culture, as proxied by the FLFPR in the country of ancestry of his parents, has a large impact on his wife’s labor supply. The effect of the immigrant women’s own labor supply prior to migrating and the FLFPR in their source country is investigated by Blau and Kahn (2011) to provide evidence on the role of human capital and culture in affecting immigrants’ labor supply and wages in the U.S. Their results provide further evidence that women from source countries with relatively high levels of FLFP have higher working hours in the U.S. Moreover, they reveal that most of this effect remains even when controlling for the immigrant’s own pre-migration labor supply, which itself strongly affects immigrants’ labor supply in the U.S. In a similar study, Blau et al. (2011) show that source-country FLFPR is also positively associated with immigrant women’s labor supply assimilation profiles, with those coming from high female labor supply countries eventually assimilating fully to native labor supply levels. The results of these studies suggest an important role for source-country culture in affecting immigrant women’s labor supply. However, the effect of culture on immigrants’ behavior may weaken as immigrants assimilate to the culture of their host country. This argument is based on Fernández’ notion that nothing in the conception of culture considers it as static or slow changing. In fact, culture might change over time and the speed of cultural change depends on how quickly social beliefs and preferences alter over time, which in turn depends on the individual’s environment (Fernández, 2011). A salient example of a cultural change is seen in the evolution of social attitudes and beliefs toward women’s market work, which serves as one possible explanation for the dramatic change in FLFP over time. In order to explain the sharp increase in FLFPR, Fogli and Veldkamp (2011) as well as Fernández (2013) develop a model of cultural change that is brought about by a process of endogenous intergenerational learning. In their model, women are assumed to learn about the long-term payoffs of working by observing (noisy) private and public signals and then make a work decision. When very few women CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 135

participate in the labor market, the noisiness of the public signal is high and learning is very slow. As information accumulates in some regions, the signal improves and beliefs about work become more positive. As a result, the proportion of women who work in that region increases.4 While it is not the aim of this paper to provide an empirical test of these theories, their main implications can be easily applied to female immigrant labor supply decisions. By observing other working women in the host country, female immigrants might change their attitudes and beliefs regarding women’s role in the workplace and gradually adapt to the behavior of native women. The higher thereby, all else equal, the proportion of working women in the host country, the more positive the beliefs about work and the higher the probability that an immigrant women decides to participate in the labor market. Assessing the relationship between host-country FLFP and the labor supply of female immigrants might therefore provide some insights into whether immigrant women change their attitudes and beliefs and assimilate to the labor market behavior of natives. While – since the seminal work of Chiswick (1978) – a sizable body of literature has evolved that examines immigrant-native assimilation patterns within a given destination country, studies that analyze immigrants in different resident countries to provide evidence on the role of host-country characteristics in immigrant behavior are scarce. The only study that aims at assessing the effect of host-country FLFP on female immigrant labor supply is Kok et al. (2011) for the Netherlands. However, as their study is based on immigrants within a single country, their identification of the host-country effect does not rely on differences in FLFPR between immigrants’ countries of residence, but on the difference in levels and speed of adjustment between different cohorts of immigrants. In particular, they use the increase in FLFPR over successive birth cohorts of native women as a proxy for Dutch culture. The authors’ results suggest that both differences in home-country female participation and the trend in native female participation, as a measure for host-country culture, have an impact on the participation of immigrant women. The authors conclude from these results that host-country participation is at least as important as home-country participation in affecting immigrants’ labor supply decisions. Although a positive relationship between host-country FLFP and immigrant women’s labor supply might be indicative of immigrant women adapting to the culture of their host country and therefore to the work behavior of natives, other explanations are also possible. As a given woman’s decision to participate in the labor market does not only depend

4The main difference between the two models lies in the assumption regarding the driving force behind female labor supply dynamics. While Fernández (2013) assumes that women start with biased, pessimistic beliefs about working women which become more positive as participation rises, Fogli and Veldkamp (2011) assume that women start with unbiased beliefs, but face uncertainty about the effects of maternal employment on their children, which falls as information accumulates. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 136 on her preferences and beliefs, but also on a whole series of economic and institutional factors that may differ across countries, FLFP at the aggregate level will not only reflect a country’s cultural environment, but its economic and institutional conditions as well. However, although we are not able to identify the source of assimilation, the effect of the LFPR of native women in a given country on the work behavior of immigrants is still indicative as to whether immigrants adapt to the labor market behavior of natives. In the present paper, we make a number of contributions to the existing literature. First, we contribute to the literature on the role of source-country culture on female immigrant labor supply. While previous literature has exclusively focused on the U.S.5, we analyze the labor market behavior of immigrants in 26 European countries, thereby providing first evidence on this topic for Europe. Second, we take advantage of the use of cross-country data as compared to single- country data to analyze immigrant labor supply behavior. Observing immigrants in different destination countries enables us to provide evidence on the relationship between host-country FLFP and immigrants LFP, thereby shedding light on assimilation patterns of immigrants to the work behavior of natives. Effectively, we are able to disentangle the effects of source- and host-country FLFP on immigrant women’s labor supply. In contrast to earlier work, our research design allows us to control for a variety of source- and host- country characteristics beyond FLFPR. While controlling for a large set of macroeconomic indicators ensures that we estimate the true effect of source- and host-country FLFP on immigrant women’s labor supply, assessing the effect of these economic and institutional conditions on immigrant behavior is of considerable interest in itself. Lastly, we conduct our analysis separately for first- and second-generation immigrants. As Fernández and Fogli (2009) outline, the effect of source-country culture on economic actions should be weaker for second-generation immigrants than for first-generation immigrants, as cultural transmission is restricted mostly to parents and ethnic social networks rather than operating in society at large (e.g., schools, media, etc.). On the other hand, we expect the effect of host-country FLFP to be stronger for second-generation immigrants than for first-generation immigrants, as second-generation immigrants grow up with the culture of their host country. Hence, analyzing the differing effects of source- and host-country characteristics on the labor supply of first- and second-generation immigrants sheds further light on cultural and economic assimilation patterns.

5With exception of the paper by Kok et al. (2011) for the Netherlands. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 137 6.3 Method, Data, and Descriptive Statistics

6.3.1 Method

In our empirical analysis, we start with estimating the following model:

J K 0 X s X h 0 0 lfpijk = Φ(xiβ + δjcj + γkck + pjkλ + tiϑ + ijk), (6.1) j=2 k=2 where lfpijk is a binary indicator that takes value 1 if immigrant woman i from source country j in host country k participates in the labor market at the time of observation. In xi, we include a set of individual characteristics and household characteristics as outlined P s P h below. δjcj and γkck are full sets of dummy variables for the immigrant’s source and host country, respectively, while pjk is a vector of country-pair variables describing the economic and cultural relationship between an immigrant’s source and host country. ti is a set of dummy variables for the year of observation and ijk is the error term. Hence, we start our analysis of immigrant women’s labor supply by using country dummies rather than the quantitative source- and host-country variables as our cultural proxies. This has the benefit of not requiring the relationship between culture and lfpijk to be linear in the cultural proxy. Furthermore, it allows to fully capture the effects of source-country culture and host-country characteristics on immigrant women’s labor supply. However, the main drawback of including the woman’s country of ancestry and her residing country as proxy variables is that such an approach is not explicit as to why different groups of immigrants, as defined by their source and host country, differ in their labor market behavior. P s The next logical step therefore is to replace the source-country dummies – δjcj – by a vector of source-country characteristics – sj:

K 0 0 X h 0 0 lfpijk = Φ(xiβ + sjθ + γkck + pjkλ + tiϑ + ijk). (6.2) k=2 Model 2 is similar to the so-called epidemiological approach used, amongst others, by Antecol (2000), Fernández et al. (2004) and Fernández (2007). This approach enables us to measure the effect of source-country FLFP on immigrant women’s labor supply in their host country, while holding the host-country characteristics fixed, i.e., by still including a P h set of dummies for the immigrant’s host country – γkck. In doing so, we are able to test whether the positive correlation between source-country FLFP and immigrant women’s labor supply in the U.S. holds for immigrants into Europe as well. The identification of this cultural effect on the labor supply decisions of female immigrants rests on the assumption that there are no unobserved factors that influence an immigrant woman’s CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 138

labor supply in her host country and are correlated with the FLFPR in her source country, once the other covariates are controlled for. One of the main contributions of our paper is that we are not only able to assess the eﬀect of source-country characteristics on female immigrant labor supply, but are also able to shed some light on the role of host-country characteristics in the labor market behavior of female immigrants in these countries. In doing so, we estimate the following model:

J 0 X s 0 0 0 lfpijk = Φ(xiβ + δjcj + hkπ + pjkλ + tiϑ + ijk). (6.3) j=2 This model differs from Model 1 only by including a vector of host-country characteristics P h – hk – instead of the host country dummies – γkck. This approach enables us to measure the effect of host-country FLFP on immigrant women’s labor supply, while holding the source-country characteristics fixed, i.e., by still including a set of dummies P s for the immigrant’s source country – δjcj. Model 3 therefore allows us to test whether immigrant women assimilate to the labor market behavior of native women in their host country. The identification of the host-country FLFPR effect rests on the assumption that, given the other covariates, immigrant women’s labor force participation decisions are not related to any unobserved factors that are correlated with the FLFPR in the immigrants’ host country. In order to consistently estimate the parameters of equations (6.1) to (6.3), we specify the probability that a certain individual participates in the labor market by the use of a 6 binary probit model, implying the assumption that ijk follows a normal distribution. We estimate marginal effects in all models. To address the problem of intraclass correlation in standard errors of immigrants within source- and host-country groups, respectively, we cluster standard errors at the source-country level (Model 2) and host-country level (Model 3), respectively.7 We further use host-country population weights in all regressions, which ensure that each country is represented in proportion to its actual population size.

6.3.2 The European Social Survey

Our basic data source at the individual level is the European Social Survey (ESS), a multi-country biennial cross-sectional survey funded jointly by the European Commission, the European Science Foundation and academic funding bodies in each participating

6Logit and linear probability models yielded similar results. 7While estimating clustered standard errors is the standard solution to address the problem of within- group error correlation, the standard errors obtained by this method might still be downward biased if the number of clusters is very small (see, e.g., Angrist and Pischke, 2009). We therefore check the robustness of our results (see Section 6.5.5) by estimating a linear regression model with standard errors obtained by a bias-reduced linearization method as proposed by Bell and McCaffrey (2002). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 139 country.8 The central aim of the ESS is to gather data about people’s social values, cultural norms and behavioral patterns within Europe. The first round of the ESS was fielded in 2002/2003. Up to now, five waves are available, covering a total of 33 nations. The survey consists of two elements – a basic interview questionnaire conducted in every round and a supplementary questionnaire devoted to specific topics, which changes over time. In particular, the ESS contains information on the country of birth of both the respondent and the parents, which allows us to precisely identify the source country of both first- and second-generation immigrants. We define first-generation immigrants as individuals born outside their resident country. Respondents are classified as second- generation immigrants if one or both parents are born outside the host country. We use the cumulative ESS data, which pools the common information from the first to the fifth ESS round, including a total of 31 countries and roughly 243,000 individuals. We exclude host countries not belonging to the European Union (except for Switzerland and Norway)9 as well as those for which the number of surveyed female immigrants is particularly small (lower than 15 individuals). The latter restriction is also applied to the source countries, i.e., we eliminate source countries with fewer than 15 observations.10 We consider women aged 26 to 59 years only, in order to avoid variations in FLFP due to differences in education leaving ages and statutory retirement ages across countries. Our final sample consists of 8,251 immigrants in 26 countries11, of which roughly 63% are first-generation and 37% are second-generation immigrants.12 These immigrants come from 59 different source countries, while the number of distinct source countries is much higher for first-generation than for second-generation immigrants (58 as opposed to 30).13

8The ESS uses a methodologically rigorous multinational design that guarantees representativeness. Extensive documentation of the data is available at http://ess.nsd.uib.no/. 9In particular, we exclude immigrants in Croatia, Israel, Russia, Turkey, and the Ukraine. In doing so, we assure that the countries in our sample fundamentally underly the same institutions and regulations, and thus comprise a more homogeneous sample. 10Increasing the threshold to 20 or 25 individuals per host and source country, respectively, yielded similar results. 11The host countries included in our sample are Austria, Belgium, Bulgaria, Cyprus, the Czech Republic, Denmark, Estonia, Finland, France, Great Britain, Germany, Greece, Hungary, Ireland, Italy, Lithuania, Luxembourg, the Netherlands, Norway, Portugal, Poland, Slovenia, Slovakia, Spain, Sweden, and Switzerland. We do not observe a sufficient number of first-generation immigrants in Bulgaria and Poland, and of second-generation immigrants in Cyprus, Italy, and Portugal, which reduces the generation-specific samples to 24 and 23 countries, respectively. A robustness analysis including only the intersection of both country samples yields similar results. 12The low share of second-generation immigrants in our data can be explained by the fact that information on the parents’ country of birth is only included from round 2 of the ESS onwards. 13For a list of the source countries included in our sample, see Table 6.A3 in the Appendix. Note that we had to aggregate some source countries in case political transformations led to a separation or unification of these countries over time. These aggregate countries are Czechoslovakia, the USSR, and Yugoslavia. The macroeconomic indicators for these countries are calculated as a population-weighted average of the single-country values. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 140

Our outcome of interest is an individual’s labor market status at the time of the

interview (lfpijk). In particular, lfpijk is a binary indicator that takes value 1 if immigrant woman i from source country j in host country k stated that her main activity within the past 7 days was either being employed or being unemployed while actively looking for a job, and 0 otherwise. The ESS data contain detailed information on the respondent’s socio-demographic characteristics as well as the composition of her household. Based on this information, we generate the following variables which serve as controls in all our regressions: age (3 categories), highest level of education (primary, secondary, or tertiary education), partner living within the household, number of children, youngest child is 0-2 years and 3-5 years, respectively, and population density (thinly, medium, or densely populated). For both first- and second-generation immigrants, we further include some immigration- specific variables. For first-generation immigrants, we include indicators for the immigrant’s years since migration (5 categories) and for whether she immigrated after age 18.14 The inclusion of the latter variable allows us to control for whether a woman obtained her (primary and secondary) education in her host or in her source country, with the former presumably being less affected by home-country characteristics and more similar to natives when they reach adulthood than those migrating as adults. Moreover, we include a dummy variable indicating whether an immigrant woman speaks the host country’s language. This information is obtained from a question included in the ESS that asks respondents to name up to two languages they speak most often at home. The variable takes value 1 if one of these two languages is also one of the official languages of the immigrant’s country of residence. While the aforementioned variables are specific to first-generation immigrants, we also include an additional variable for second-generation immigrants, indicating whether both parents or only one of them were born outside the resident country. Although the ESS is not designed as a household survey, but is in effect an individual survey, it contains information on a respondent’s partner. Controlling for partner characteristics in women’s labor supply decisions is meaningful for two reasons. First, for those living with a partner some kind of joint decision-making process with respect to labor supply and household production has to be assumed.15 Independent of which kind of model is assumed to underlie a couple’s decision-making process, women are predicted to be less likely to participate the higher their partner’s earnings potential. Second, there is evidence of assortative mating in the marriage market, i.e., more educated (and hence

14As controlling for age, years since migration, and age at migration in a linear form is not possible due to perfect correlation of these variables, we decided to include both age and years since migration in categories which allows us to further add a dummy variable indicating the age at migration. 15The economic theory of joint labor supply decisions within the household was initiated by Becker (1965) and developed, amongst others, by Gronau (1977), Manser and Brown (1980) and McElroy and Horney (1981). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 141

higher income) men tend to be married to more educated women (see, e.g., Pencavel, 1998). The husband’s higher income will decrease the incentives for his wife to engage in market work and, in this way, mask the strength of the effect of source-country culture on women’s labor supply decisions. We attempt to capture the impact of both assortative mating and joint decision making within the household by controlling for the partner’s highest level of education and his working hours. However, as these variables are endogenous to a women’s LFP decision, we will not include them in our basic regression model. As Blau and Kahn (2011) show, a strong predictor of an immigrant women’s labor supply in the host country is their own labor supply in the source country prior to migrating. If immigrant women from high FLFP countries have more work experience prior to their arrival in the host country, then the observed effect of source-country FLFP may be due to the relatively high levels of job-related human capital that they accumulated before migration. Hence, without taking the immigrants’ pre-migration labor supply into account, we cannot be sure whether a positive relationship between high source-country FLFP and immigrant women’s labor supply in their host country provides evidence for an effect of broader culture on immigrants’ labor market behavior, or whether it simply reflects the immigrant woman’s own pre-migration behavior. While the ESS data do not allow to control for an immigrant’s own labor supply in the source country, they provide information on the human capital and labor supply of the immigrant’s parents. In particular, each respondent is asked about (i) his mother’s and father’s highest level of education and (ii) their labor market status at the time the respondent was 14 years old. As the empirical literature on intergenerational mobility has consistently documented a high persistence between parents’ and children’s economic outcomes16, we use these indicators as a proxy for the immigrant’s own labor supply prior to migration. In doing so, we are able to test whether the effect of source-country FLFP on immigrant labor supply persists even if the immigrant’s pre-migration human capital is controlled for. Table 6.1 shows the descriptive statistics of the individual and household characteristics outlined above separately for the sample of first- and second-generation female immigrants (columns 2 and 3). For comparison, column 4 further shows the respective values for native women. With respect to our dependent variable, women’s probability of participating in

the labor market (lfpijk), distinct differences between the three samples appear. At the time of the interview, 69% of the native women, as compared to 65% of the first-generation and 71% of the second-generation immigrant women indicate to actively participate in the labor market. Hence, while the LFP of first-generation immigrant women is indeed considerably lower than that of native women, the LFP of second-generation immigrant

16For a recent overview of studies on intergenerational mobility, see Black and Devereux (2011). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 142 women even exceeds the LFP of natives.17 This result might support our notion that recent waves of immigrants into Europe increasingly come from countries that are characterized by low FLFPR, and therefore show a lower labor market attachment than former immigrant women. However, it is also necessary to take into account the changing reasons for migration. During the 1950s and 1960s, many European countries, such as Germany, Great Britain, and France, encouraged labor immigration in order to fill gaps in the national labor market, while in the later decades migration for family reunion and the seeking of political asylum became more important (European Commission, 2011). Table 6.1 further shows that first-generation immigrant women are slightly younger (41 years on average) than second-generation and native women (43 years on average) and have a higher number of children (0.73 as opposed to 0.63 for second-generation immigrants and 0.59 for native women). Regarding the educational attainment of the three groups, no clear pattern emerges. While the share of women with a tertiary degree is highest among first-generation immigrants, they also have the highest share of women with a primary degree. This might again reflect that the reasons for migration are quite diverse. With respect to the immigrant-specific variables, the results show that more than 40% of the first-generation immigrant women live in their destination country for more than 20 years, and the majority of these women migrated after the age of 18 (83%). We further see that 30% of the second-generation women have both a mother and a father who were born outside the residence country, while the rest are daughters of interethnic marriages. Whereas the personal characteristics of the partners and fathers do not differ substantially across the three groups of women, we observe large differences regarding the employment status and the educational attainment of the mothers of these women. In particular, mothers of first-generation immigrant women are much less likely to have been employed when their daughter was 14 years old than mothers of second-generation and native women (48% as opposed to 58% and 55%), though being better educated than the latter. This observation highlights the importance of testing the robustness of our results to controlling for parental characteristics. If the latter are not controlled for, a positive correlation between source-country FLFP and the labor supply of immigrant women might purely arise from the fact that the mothers of immigrants from high-LFP countries are more likely to have been employed than those from low-FLFP countries. In this case, it is rather the actual behavior of the mother than the preferences and beliefs held within the source country that ultimately determine the labor supply of immigrant women in Europe.

17Note that the mean values for the three groups are not statistically diﬀerent from each other. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 143

6.3.3 Aggregated Data

For the analysis of source- and host-country effects, we augment our individual data with an extensive time-series, cross-country database of aggregated source- and host-country characteristics.18 While for first-generation immigrants source-country characteristics refer to the immigrant’s country of birth, the source-country characteristics for second-generation immigrants refer to the country of birth of the father or the mother of the immigrant, depending on who of the two was born in a foreign country. In case both parents were born outside the host country and emigrated from different countries, we use the mother’s birthplace to assign the country-of-ancestry indicators to second-generation women, as we assume the intergenerational transmission of beliefs and values regarding women’s role in society to be stronger between mothers and daughters than between fathers and daughters (cf. Casey and Dustmann, 2010).19 For both first- and second-generation immigrants, the host-country indicators were assigned to immigrants based on their country of destination and the year of observation (2002 to 2011). With respect to the source-country characteristics, however, the optimal point in time to take these indicators from is not obvious. For first-generation immigrants, one possibility is to measure the source-country variables at the time these immigrants left that country. These values reflect the norms and values the immigrants grew up with and carry to their host country. A second possibility is to use the current values of the source-country indicators. These values reflect the norms and values currently held by the immigrant’s counterparts, i.e., the individuals living in the immigrant’s country of ancestry at time of observation. For second-generation immigrants, the same reasoning applies. On the one hand, it could be argued that the values of the source-country variables measured at the time the immigrant’s parents left their home country would best reflect the culture of the country of ancestry. On the other hand, one could argue that the norms and values that parents and society transmit to second-generation immigrants might be best reflected by a comparison of what the counterparts of these women are currently doing in the country of ancestry. The main problem with assigning second-generation immigrants the source-country characteristics based on the year their parents left the country is that this information is not included in our data. Moreover, even if we were able to observe the parents’ year of arrival, data limitations would not permit us to use years prior to 1960, since values for most of our macroeconomic indicators are not available prior to that year. Hence,

18See Table 6.A4 in the Appendix for a detailed description of the macroeconomic data. 19In our sample, 5.4% of the second-generation female immigrants have parents who are born in different source countries. As a robustness check, we have also run our regressions using the country characteristics of the father’s birthplace for these women. The results of our regressions remain unaffected. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 144 we decided to assign both first- and second-generation immigrants the source-country characteristics based on the year of observation (2002 to 2011).20 Following this approach has several advantages. First, we can make sure that the macroeconomic indicators are available for the majority of the source countries in our sample. In doing so, we avoid the problem of a non-random selection of source countries into our sample, which would otherwise arise from the fact that long-ranging time-series data for our macroeconomic indicators of interest are only available for a limited number of countries. Second, using current values of the macroeconomic indicators for both first- and second-generation immigrants has the advantage of treating first- and second-generation immigrants similarly, which makes a comparison of the behavior of the two groups more meaningful. Lastly, the use of current values of the source-country characteristics takes into account that, if not emigrated, immigrant women would have gradually changed their preferences and beliefs in the same way as those still living in the source country, and does therefore not assume culture to be constant over time. However, in order to assure that our results are not driven by the choice of observation time, we further perform a sensitivity analysis in which we assign first-generation immigrants the source-country indicators based on their year of migration (see Section 6.5.4).

The variables of our main interest are F LF P Rj and F LF P Rk, the female labor force participation rates of the immigrant’s source and host country, respectively. These variables cover the rate of the economically active population for women in a given age group, which are available in 5-year-intervals ranging from “25 to 29” to “55 to 59”. We use age-specific participation rates instead of a single measure over all age groups in order to avoid the FLFPR to vary with the age structure among the population, thereby blurring differences in women’s economic activity between the countries. The differentiation by age group is especially important for the host-country FLFPR, as the demographic composition of immigrants differs largely across European countries. While the host-country FLFPR may reflect the economic, institutional, and cultural environment of the immigrant, only the cultural component is reflected in the source- country FLFPR. The estimated effect of the latter on female immigrant labor supply therefore provides insights into the role of culture as opposed to institutions and purely economic factors in explaining the diversity of labor market outcomes between immigrants. However, as Fernández (2011) claims, parents are not the only transmitters of culture. The relationships and institutions of the local environment also impact individual behavior. By observing other working women in the host country, female immigrants might change their attitudes and beliefs regarding women’s role in the workplace and gradually adapt to

20In doing so, we follow Antecol (2000), Fernández and Fogli (2009), and Kok et al. (2011), while Blau and Kahn (2011) and Blau et al. (2011) use past values of the source-country characteristics for their analysis of the labor market behavior of first-generation immigrants. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 145 the behavior of native women. The estimated effect of the host-country FLFPR on the labor supply of female immigrants therefore reveals whether immigrant women compare themselves with native women and gradually assimilate to their labor market behavior.21 On both the source- and the host-country level, we control for a variety of additional economic and institutional indicators that might have an impact on individual labor supply decisions. On both levels, we include the country’s total fertility rate and its GDP per capita, the latter being an important push and pull factor of immigration, respectively. On the source-country level, we further include a variable denoting the average years of schooling of the source-country population in the immigrant’s age group.22 As shown by Borjas (1992, 1995), the level of ethnic human capital (as measured by average wages or education of the immigrant group) may help to explain individual outcomes such as education or earnings due to ethnic externalities in the human capital process. As Fernández and Fogli (2009) state, one way to think about these human capital externalities is that the human capital embodied in an individual’s ethnic network matters. Including the years of schooling in the source country in our analysis can therefore serve as a proxy for average (parental) human capital and for the human capital embodied in the woman’s ethnic network. We further include some additional indicators on the host-country and the country-pair level. A major concern when examining the labor market behavior of immigrants across host countries is the selection of immigrants into these countries. Although cross-country migration decisions are clearly non-random, our primary concern here is whether selective migration could spuriously generate an effect of host-country FLFPR on immigrant women’s labor supply in their host country. It can be argued that female immigrants with high preferences for women’s market work, who intend to participate in the labor market in their host country, will migrate to countries that offer the best opportunities to do so. If this is the case, a positive correlation between immigrant women’s probability of participating in the labor market and the FLFPR in their host country may not provide clear evidence for an assimilation of immigrants towards the FLFP of natives, but might rather reflect a selection of high pre-migration labor supply women into high FLFP host countries. In order to address this problem, we attempt to control for the immigrant’s migration

21A remaining concern when analyzing the relationship between immigrant women’s labor supply and the FLFPR in their host-country is that the participation decisions of immigrant women are already embodied in the latter. Hence, even in the absence of any assimilation patterns, there might exist a positive, though very small, correlation between the LFP of immigrant women and the FLFPR in their host country. In order to eliminate this possible correlation, we alternatively used the predicted LFPR of native women (calculated from the ESS data) as our measure for the host-country FLFPR. The results are robust to using this alternative measure. 22As for the FLFPR, the age groups range from “25 to 29” to “55 to 59” in 5-year-intervals. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 146 decision as well as possible. In addition to including indicators for the host country’s GDP and fertility rate, we control for the country’s unemployment rate, arguing that women with high preferences for market work, whose migration decision is economically motivated, will migrate to countries with good employment opportunities and therefore low unemployment rates. We further capture the selection of immigrants into host countries by controlling for the total share of migrants as well as the share of migrants from the women’s source country among the host country’s population. While the former variable captures the host country’s cultural diversity in general, the latter variable controls for the fact that immigrants from countries with less traditional gender roles may choose to move to less traditional countries, and similarly, those from countries with more traditional gender roles may choose to move to more traditional countries. However, although this variable serves as a proxy for the immigrant’s migration decision, it might also reflect the immigrants’ composition of the neighborhood. As Fernández and Fogli (2009) argue, an individual’s neighborhood may play an important role in transmitting and preserving a set of beliefs or preferences, independently of the human capital embodied in an individual’s ethnic network. A neighborhood that has a relatively high proportion of individuals from the same source country may help preserve that country’s culture by punishing behavior that is different from the norm and thereby keep the culture of the source country alive. Although the share of immigrants of the same ancestry is only a raw proxy of the immigrant’s neighborhood, it might still provide some insights into the role of neighborhoods in cultural transmission. Finally, we add some variables capturing the relationship between the immigrant’s country of birth and her country of residence. First, we control for whether the two countries share or have ever shared a colonial relationship. This is to acknowledge the fact that countries that had the same colonial history often established similar institutional settings, which not only facilitates migration flows, but also reduces the barriers of immigrants to enter the host country’s labor market. Moreover, we include indicators for the geographical, linguistic, and genetic distance between the immigrant’s source country and her host country, which serve as proxies for the individual costs of migration. The geographical distance is defined as the geodesic distance between the capitals of the source and the host country in 1,000 kilometers. The linguistic distance measures the phonetic similarity between all of the world’s languages. The basic idea is to compare pairs of words having the same meaning in two different languages according to their pronunciation. The average similarity across a specific set of words is then taken as a measure for the linguistic distance between the CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 147 languages (Bakker et al., 2009).23 Lastly, genetic distance is measured as the difference in allele frequencies. Alleles are the specific manifestation of a gene, which might differ between individuals. The genetic distance measure as defined by Cavalli-Sforza et al. (1994) is related to the inverse probability that groups of alleles are the same for two populations. Hence, the lower the common frequency of alleles in two populations, the longer these populations have been separated.24 Genetic distance therefore serves as a proxy for the cultural distance between two countries, which might have an impact on the immigrants’ migration decision. While geographical, linguistic, and genetic distance, to a certain extent, all measure the direct and indirect costs of migration, their effect on the labor supply behavior of immigrants is theoretically ambiguous. As Chiswick (1999) states, immigrants who come from a greater distance are likely to have higher labor market returns to migration than those coming shorter distances, all else equal. Stated differently, immigrants who migrated though facing high costs of migration are a positively selected sample of all immigrants and may therefore have a higher chance of participating in the host-country’s labor market. A second important issue that has to be considered when analyzing the labor supply of immigrants across different host countries is that immigrants might face restrictions in their access to the host country’s labor market. Specifically, immigrants from non-EU countries might not be allowed to work in their host country in the first years after arrival. In order to distinguish immigrants that are permitted to work from day one in their host country from those who might face restrictions to do so, we include a dummy variable that indicates whether immigrants underly the “right of free movement of workers” at the time of observation. The right of free movement of workers is a fundamental principle enshrined in Article 45 of the Treaty on the Functioning of the European Union, which generally permits workers to search for employment, to be employed, and to reside in any Member State of the European Union (European Commission, 2010a).25

23This measure was first applied to economics by Isphording and Otten (2011), who analyze the effect of linguistic distance on the language fluency of immigrants in Germany. 24Changes in genes, hence the emergence of new alleles, happen randomly at an almost constant time. As evolutionary pressure might direct this random change into certain directions, the genetic distance measure focuses on neutral genes, which are not prone to evolutionary pressure. By focusing on neutral changes, the genetic distance measure therefore does not explain differences in labor supply due to superior skills or ability. 25While the right of free movement of workers generally applies to all immigrants migrating within the European Union, there is a clause about a transition period before workers from the new Member States can be employed on equal, non-discriminatory terms in the old Member States. The old Member States have the right to impose such transitional period for 2 years, then to decide whether to extend it for additional 3 years, and then, if there is serious proof that labor from new Member States would be disruptive to the market in the old Member States, the period can be extended for the last time for 2 more years. Furthermore, citizens of the Member States of the European Economic Area and Switzerland have the same right of freedom of movement and these countries are treated as old Member States inside the EEA (European Commission, 2003, 2005). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 148

While the aforementioned variable mainly captures the different rights of EU and non-EU immigrants, the labor market access of the latter might still vary across the European countries. Not only may third-country immigrants be prohibited to work in the country of residence in the first years after arrival, they may further have limited access to the full labor market, education system or employment services of the host country. In order to address this issue, we make use of the Migrant Integration Policy Index (MIPEX)26, which measures policies integrating migrants in 25 EU Member States as and 3 non-EU countries (i.e., Canada, Norway, and Switzerland). It considers over 140 policy indicators grouped into 6 broad policy areas, one of which is the “labor market mobility” of immigrants. “Labor market mobility” measures if migrant workers are eligible for the same opportunities as EU nationals to work in most sectors. In particular, it takes into account whether migrant workers can expect help from labor market integration measures to adjust to the language and professional demands of the labor market. Moreover, it measures how secure migrant workers are in their employment, whether they can renew most types of work permits and remain living in the country and look for work if they lose their job. The index varies between 0 and 100, with higher values meaning that migrants have more rights in the corresponding policy area. Table 6.2 shows the descriptive statistics of the aggregated source- and host-country variables as well as the bilateral variables separately for the sample of first- and second- generation immigrants. In order to best represent the country characteristics relevant for the immigrants included in our sample, the values have been calculated as host-country population weighted averages over all observations within each sample. The country characteristics in the top of Table 6.2 are measured at the time of observation, while the bottom of Table 6.2 shows the source-country variables for first-generation immigrants measured at the time these immigrants left the country.27 With respect to our variable of main interest, F LF P R, Table 6.2 indicates that as compared to the European average, first-generation immigrants come from a source country that has on average a 13 percentage points lower FLFPR and second-generation immigrants come from a source country that has on average a 14 percentage points lower FLFPR at the time of the interview. At the same time, hardly any difference in the average LFPR of males between the immigrants’ source and host countries appear. These results support

26MIPEX is led by the British Council and Migration Policy Group (MPG) and is freely accessible at: http://www.mipex.eu/. 27Note that the variables describing the relationship between the source and the host country are time invariant, except for the share of migrants from the same source country in the immigrant’s host country. Technically, the “right of free movement”-variable is time variant as well, as the countries underlying this fundamental principle change over time. However, as this variable serves as a proxy for the immigrants’ restrictions in their access to the host country’s labor market, a calculation of past values for this variable is of little meaning. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 149

our hypothesis that the low labor market activity of (first-generation) immigrant women in Europe might be explained by the more traditional views about gender roles held in their source countries. However, the fact that second-generation immigrant women are even more likely to participate in the labor market than native women, although their parents come from high-traditional source countries as well, also lends support to our argument that immigrant women might change their preferences and beliefs and assimilate to the labor market behavior of natives. Regarding the other country characteristics, the results reveal that first-generation immigrant women come from source countries with a higher total fertility rate at the time of observation, while there is no difference in average source- and host-country fertility rates for second-generation immigrants. As expected, GDP per capita is much higher among the immigrants’ host countries than among the immigrants’ source countries, while the difference between source- and host-country GDP is higher for first- than for second-generation immigrants. Further differences between first- and second-generation immigrants appear with respect to the relationship between the immigrants’ source and host country. Both the geographic, the genetic, and the linguistic distance between the source and the host country have increased considerably over migration cohorts, while the role of colonial ties in the immigrants’ choice of destination country has decreased. Lastly, a comparison of the source-country characteristics for the sample of first- generation immigrants calculated at different points of time, i.e., the year of observation (2002 to 2011) and the year the immigrant left her country (1982 to 2011), reveals a large variation in the macroeconomic indicators over time. While FLFPR and years of schooling have increased over time (by 6 percentage points and 1.5 years, respectively), fertility rates have decreased over the observation period (by 0.5 children per women). These findings highlight the importance of conducting a sensitivity analysis in which we assign first-generation immigrants the source-country characteristics based on the year of migration.

6.4 Basic Results

6.4.1 Source- and Host-Country Fixed Eﬀects

The estimation results of Model 1, containing both source- and host-country fixed effects, are shown in Table 6.3. The results for the individual and household controls are in line with previous evidence on female (immigrant) labor supply. For both first- and second-generation immigrants, LFP is significantly lower among older women (46 to 59) as compared to middle-aged women (36 to 45 years). A further strong predictor of the CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 150 labor supply of immigrant women is their level of education, with those having completed tertiary education being significantly more likely and those with only a primary school degree being significantly less likely to participate in the labor market than those with a secondary school degree. While first-generation female immigrants living together with a partner show a lower LFP probability as compared to single women, cohabitation is uncorrelated with the labor supply of second-generation immigrants. Although we do not know whether the partner is also an immigrant and the two migrated together, the strong negative correlation for first-generation immigrants might reflect that those women who migrated together with their partner are less likely to have migrated for their own economic interests and are therefore less likely to participate in the labor market than single women. Both the number of children living in the household and the presence of small children (aged 0 to 2) is negatively correlated with female immigrant labor supply. The degree of urbanization of the immigrants’ place of residence is hardly correlated with their labor supply decision. For first-generation female immigrants, labor supply is significantly lower for those who just arrived in their host country (less than 6 years ago) than for those who live in the country for more than 20 years. Those who migrated as adults (age 18 and over), however, do not differ from those who migrated as children. Moreover, speaking the host country’s language at home is positively correlated with the likelihood of participating in the labor market. Lastly, second-generation immigrants whose father and mother were both born outside the residence country do not differ from those with a single migrant parent with respect to their labor market behavior. The bottom of Table 6.3 shows the results of the variables that describe the relationship between the immigrants’ country of origin and their host country. With controlling for both the immigrant’s source country and her country of residence, hardly any of these variables show explanatory power in female immigrant labor supply. For first- and second-generation immigrants, both the source-country dummies and the host-country dummies are jointly highly significant, reflecting a considerable variation in LFP, both between immigrant women from different countries of origin and between immigrant women across the European countries. In order to assess the relative importance of an immigrant’s cultural background, as measured by the source-country fixed effects, and her cultural, institutional, and economic environment, as measured by the host- country fixed effects, we re-estimate our model by OLS and calculate the semipartial R2 of the source- and host-country dummies, respectively. The semipartial R2 represents the proportion of variance of lfpijk accounted for by the source- and host-country dummies, respectively, after all other covariates are controlled for. The respective results are displayed in Table 6.A1 in the Appendix. For first-generation immigrants, the results show that CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 151

17,4% of the overall variance of lfpijk can be explained by our covariates, including the source- and host-country fixed effects. Of this explained variance, 21.2% are accounted for by the source-country fixed effects and 7.0% are accounted for by the host-country fixed effects. Hence, the LFP decisions of first-generation female immigrants are more strongly determined by their cultural background than by the cultural, institutional, and economic conditions in their host country. For second-generation immigrants, the difference in the explanatory power of the source- and host-country fixed effects is less prounced. While all covariates account for 11,7% of the overall variation in lfpijk, 11.8% of this explained variance can be attributed to the source-country fixed effects and 10.3% can be attributed to the host-country fixed effects. This result supports our expectation that second-generation immigrants are less affected by source-country conditions and more affected by host-country conditions as compared to first-generation immigrants. However, it also reveals that although second-generation immigrant women grow up in the environment of their host country, their labor market behavior is still strongly determined by their country of origin.

6.4.2 Source-Country FLFPR

In order to gain insights into the driving forces behind the differences in labor supply between women from different countries of origin, we re-estimate the above specification by now replacing the source-country dummies with the respective source-country characteristics (Model 2). The estimation results for this model are shown in Table 6.4.

The estimated marginal effect of our variable of main interest, F LF P Rj, shows a strong positive correlation between the FLFPR in the immigrant’s country of origin and her probability of participating in the host country’s labor market. This result holds for both first- and second-generation immigrants, while the magnitude of the estimated effect is higher for the latter. On average, a 1-percentage-point increase in the source country’s FLFPR is associated with a 0.2 percentage-points increase in the LFP probability of first- generation immigrant women and a 0.4 percentage-points increase in the LFP probability of second-generation immigrant women. However, as the source-country FLFPR is differently distributed for first- and second-generation immigrants, a comparison of the size of the estimated marginal effects is only meaningful to a limited extent. In order to illustrate and compare the magnitude of the source-country FLFPR effects for first- and second-generation immigrants, we can compare the LFP probability of women from a country with a relatively high FLFPR, at the 75th percentile of our sample, with women from a country with a relatively low FLFPR, at the 25th percentile. Regarding first-generation immigrants, the 25th percentile of the FLFPR in our sample is 48.0, which CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 152

roughly equals the FLFPR of Sri Lanka in 2011, and the 75th percentile is 80.0 (~Ukraine, 2011). The results suggest that an increase in the source-country’s FLFPR from the 25th to the 75th percentile increases the LFP of first-generation female immigrants by approximately 7.0 percentage points. For second-generation immigrants, an increase in the source-country FLFPR from the 25th percentile (50.7, which roughly equals the Philippines in 2011) to the 75th percentile (81.5, which roughly equals Canada in 2011) increases the likelihood of participating in the labor market by approximately 11.8 percentage points.28 The illustration of the magnitude of the effect of source-country FLFPR on female immigrant labor supply reveals two things: First, the effect is by far not negligible and second, the magnitude of the effect is indeed higher for second-generation immigrants than for first-generation immigrants. The latter result contradicts the argument of Blau (1992), who points out that cultural factors should be more apparent among first-generation immigrants, because second-generation immigrants have had time to adapt to the prevailing tastes and economic conditions of the host country. However, it should be kept in mind that our analysis does not take into account any cohort effects. If early cohorts of immigrants have a stronger source-country identity than later cohorts of immigrants, and the intergenerational transmission of this identity is strong29, then the children of former immigrant cohorts might be more affected by source-country culture than recent immigrants into the country. Hence, the finding of a relatively stronger source-country FLFPR for second-generation immigrant women does not necessarily reflect that the effect of source-country culture becomes stronger as time spent in the host country increases. In order to gain insights into whether the influence of source-country culture changes as time spent in the host country increases, we re-estimate Model 2 for first-generation immigrants by now additionally including an interaction term between source-country FLFPR and the dummy variables for the immigrant’s years since migration. The marginal

effect of F LF P Rj at each category of the years-since-migration variable is displayed in Figure 6.2. The results show that within the first five years after migration, source-country FLFPR is uncorrelated with women’s probability of participating in the labor market.30 The positive correlation between source-country FLFPR and immigrant labor supply becomes only significant from year six onwards, and then slightly decreases with time spent

28Note that the high variation in FLFPR across source countries partly accrues from the fact that we use age-group-specific instead of total FLFPR in our analysis. The above-mentioned country-year combinations chosen to illustrate the magnitude of the source-country FLFPR all refer to the FLFPR of the population aged 30 to 34. 29Using longitudinal data for Germany, which contain information on the ethnic identity of both first-generation immigrants and their children, Casey and Dustmann (2010) find a strong link between parents’ and children’s home-country identities. Moreover, they find the intergenerational transmission of the home-country identity to be strongest between mothers and their daughters. 30As only 1.4% of the women in our sample indicate that they have migrated within the last year, the insignificance of the effect of FLFPR for this subgroup is likely to be due to the small sample size. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 153

in the host-country. However, the category-specific effects are not significantly different from each other. Again, this result does not support the assumption that the effect of source-country culture decreases with time since migration. The results further show a strong negative correlation between source-country GDP per capita and the labor supply of first- and second-generation immigrants. This result seems counterintuitive at first sight, as one would expect that the higher the GDP in the country of origin, the greater the resemblance between that country’s economic structure and that of the European countries, and therefore the higher the preparedness of immigrants for the European labor market.31 However, this line of argumentation does not take into account the aspect of immigrant selection. The economic theory of migration identifies two major determinants of immigrants’ selection: the costs of migration (Chiswick, 1999) and the income inequality between source and host countries (Borjas, 1987; Roy, 1951). While the latter model predicts the educational selection of immigrants to be more positive (negative) the higher the return to skills in the destination (source) country as compared to the source (destination) country, empirical evidence on this relationship is not conclusive (e.g., Borjas, 1987; Orrenius and Zavodny, 2005). A possible explanation for ambiguous findings regarding this relationship is brought about by Belot and Hatton (2012), who show that a positive selection of immigrants from high-inequality countries can only be observed once the poverty constraint of immigrants migrating from poor countries is controlled for. In particular, the authors find that immigrants from poor countries are strongly positively selected from among the source country’s population. This result is consistent with Chiswick’s argumentation that immigrants are the more positively selected the higher their migration costs. Though having high incentives to move, immigrants from poorer countries are less likely to move as they face high (relative) migration costs, which results in the fact that only the most able will succeed. This relationship between migration costs and the selectivity of immigrants is able to explain the negative correlation between source-country GDP and immigrant women’s probability of participating in the labor market. All else equal, immigrants from low-GDP countries are expected to be a more positively selected sample of the source-country population than immigrants from high-GDP countries, and thus outperform the latter in the host-country’s labor market. For first-generation immigrants, we further find a positive and significant correlation between the average years of schooling of the source country’s population and immigrant women’s probability of participating in the host country’s labor market. This suggests that although controlling for the immigrant’s own education, the level of human capital in her source country matters for her labor market behavior. The fact that this correlation does

31This argument is put forward by Blau et al. (2011) for immigrants to the U.S. labor market. However, the authors also find a strong negative correlation between source-country GDP and the labor supply assimilation profiles of first-generation immigrant women. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 154 only hold for first-generation immigrants suggests that source-country education rather captures some unobservable human capital of the immigrant herself, such as the quality of education obtained or her labor market experience before migrating, than reflecting ethnic externalities in the human capital process. Lastly, the results for the source-country characteristics reveal a positive correlation between source-country total fertility and the labor market participation of second-generation immigrants. This result contradicts Fernández and Fogli (2009), who find a negative correlation between the fertility rate in the source country and the labor supply of second-generation immigrants in the U.S. The authors argue that the fertility rate in the immigrant’s source country captures the beliefs regarding the appropriate role of women in society as well as some independent cultural preferences for family size, which leads to a negative effect of this cultural proxy on immigrant women’s labor supply. However, to assess the true effect of these cultural proxies, the FLFPR and the fertility rate in the source country, on the labor supply of female immigrants, the correlation between the two variables has to be taken into account. For the source-countries included in our sample, fertility and FLFPR are strongly positively correlated once GDP per capita and average years of schooling are controlled for.32 Hence, women from countries with less traditional gender roles are more likely to come from high-fertility countries, and women from countries with strong traditional gender roles are more likely to come from low-fertility countries. Hence, the positive correlation between source-country fertility and immigrant women’s labor supply in their host country is most likely to reflect an indirect effect of FLFPR on immigrant women’s labor supply. The results for the variables describing the relationship between the immigrants’ source and host country show that women who migrate between countries that share or have ever shared a colonial relationship show a higher probability of participating in the labor market. This is in line with the assumption that countries with a colonial history often established similar institutional settings, which reduces the barriers of immigrants to enter the host country’s labor market. As expected, we further find a significantly higher LFP probability for women who migrate from countries whose citizens underlie the right of free movement of workers in the host country. The other relationship variables, however, show hardly any explanatory power in female immigrants’ labor supply decisions. Lastly, we see that the effects of the individual and household controls on female immigrant labor supply are robust to the substitution of the source-country dummies by the respective source-country characteristics. This indicates that our estimates do not suffer from unobserved source-country characteristics that are correlated with the

32An OLS regression of source-country FLFPR on the total fertility rate, GDP per capita, years of schooling, and time dummies yielded a coeﬃcient for the fertility rate of 2.95 with a standard error of 0.34. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 155

individual determinants of labor supply.

6.4.3 Host-Country FLFPR

While our finding of a significant positive relationship between the labor supply of immigrant women in Europe and the FLFPR in their source country supports the results of earlier studies for the U.S., little is known about the role of host-country characteristics in immigrants’ labor supply. In order to gain insights into whether immigrant women’s labor supply is affected by the FLFPR in their host country, we re-estimate Model 1 by now replacing the host-country dummies with the respective host-country characteristics (Model 3). The estimation results for this model are shown in Table 6.5.

For both first- and second-generation immigrants, the estimated effect of F LF P Rk is significantly positive, indicating a positive relationship between the FLFPR in the immigrant’s host country and her probability of participating in the labor market. On average, a 1-percentage-point increase in the host country’s FLFPR increases the likelihood of participating in the labor market by 0.7 percentage points for first-generation immigrant women and 0.9 percentage points for second-generation immigrant women. If we think of the host country’s FLFPR as reflecting the LFP decisions of all women living in the immigrants’ host country, which, amongst other factors, depend on the economic, institutional, and cultural environment within the country, the fact that this aggregate variable has explanatory power in immigrant women’s labor supply decisions suggests that immigrant women, at least to a certain extent, adapt to the labor market behavior of native women. The source of this assimilation effect, however, is ambiguous. One possible explanation for the positive correlation between host-country FLFPR and the labor supply of female immigrants is brought about by the model of cultural change developed by Fogli and Veldkamp (2011) and Fernández (2013). By observing other working women in their environment, immigrant women might change their preferences and beliefs regarding women’s roles and gradually adapt to the labor market behavior of native women. A second possible explanation is the influence of institutional circumstances on immigrant women’s labor supply decisions. A positive correlation between the host-country’s FLFPR and immigrant women’s labor supply might indicate that the LFP decisions of immigrant women are subject to the same institutional conditions as those of native women. Regulations affecting the work incentives for women, such as the tax treatment of single persons and second earners, respectively, as well as measures to facilitate the reconciliation of work and family, such as the provision of paid parental leave and the supply of public daycare, are possible candidates to affect the labor supply decisions of CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 156 native and immigrant women as well. Moreover, the correlation between host-country FLFPR and female immigrant labor supply might be due to differences in economic conditions across the European countries. For example, differences in employment prospects or wage levels might lead to different incentives for women to participate in the labor market. Lastly, it cannot be ruled out that selective migration spuriously generates an effect of host-country FLFPR on immigrant women’s labor supply in their host country. If less traditional women select themselves into high-FLFPR countries, as these countries offer the best opportunities for women’s market work, a positive correlation between immigrant women’s probability of participating in the labor market and the FLFPR in their host country may simply reflect this selection process. However, as we control for a variety of host-country characteristics beyond FLFPR, as well as for several variables capturing the relationship between the immigrant’s source and host country, selective migration alone can hardly explain the strong effect of host-country FLFPR on immigrant women’s labor supply. The relative magnitude of the host-country FLFPR effect can again be best illustrated by the use of interquartile ranges. For first-generation immigrants, the 25th percentile of the host-country FLFPR in our sample is 74.7 (~United Kingdom, 2011), while the 75th percentile is 82.7 (~Switzerland, 2011). The results suggest that an increase in the host country’s FLFPR from the 25th to the 75th percentile increases the LFP of first-generation female immigrants by approximately 5.1 percentage points. For second- generation immigrants, the 25th percentile of the host-country FLFPR is 76.3 (~Greece, 2011) and the 75th percentile is 83.8 (~Spain, 2011). An increase in the host-country FLFPR from the 25th to the 75th percentile is associated with a 6.7 percentage-point increase in the probability of participating in the labor market.33 These results suggest that the magnitude of the effect of host-country FLFPR on female immigrant labor supply is higher for second-generation immigrants than for first-generation immigrants. This finding meets our expectation that the labor supply decisions of first-generation immigrants, who grew up under a different cultural and institutional environment, are less strongly affected by the economic, institutional, and cultural conditions of their host country than those of second-generation immigrants. However, we do not find evidence that the effect of host-country FLFPR on the labor supply of first-generation immigrants increases with time since migration (see Figure 6.3). The illustration of the magnitude of the FLFPR effects further reveals that for both first- and second-generation immigrants, the relative size of the effect of host-country FLFPR on female immigrant labor supply is smaller than the corresponding effect of source-country FLFPR (Model 2). This result again highlights

33Again, the country-year examples refer to the FLFPR of the population aged 30 to 34. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 157 the importance of source-country culture in shaping immigrant women’s labor supply decisions. Regarding the other host-country characteristics, we find that none of the macroeconomic indicators shows additional explanatory power for the variation in the LFP of first-generation immigrant women. For second-generation immigrants, we find a positive correlation between host-country fertility and a negative correlation between host-country GDP per capita and women’s likelihood of participating in the labor market. While the low LFP of women in high-GDP countries might reflect an indirect effect of the country’s generosity of welfare provision on women’s incentives to work, the positive relationship between host-country fertility and immigrant labor supply is hard to explain. We further find the genetic distance between the source and the host country to be negatively correlated, and the linguistic distance between the two countries to be positively correlated with the LFP of second-generation female immigrants. These results seem contradictory at first sight, as both variables should capture the costs of migration of the immigrants’ parents. Hence, if showing any explanatory power in the labor supply decisions of second-generation immigrants, one would expect these variables to be positively correlated with women’s LFP probability, reflecting that parents who migrate though facing high migration costs are a positively selected sample of all immigrants. However, while both the linguistic and the genetic distance capture the selection of the immigrants’ parents, the latter might further have a direct impact on the labor market outcomes of the second generation. One can imagine that the higher the genetic distance between the host country’s and the source-country’s population, i.e., the higher the dissimilarities between the two populations with respect to their physical appearance, their behavior, and their cultural habits, the higher the barriers for immigrants to integrate into the host country’s society, an effect that might even continue through the second generation. Lastly, our results show that once the host country’s total migrant stock is controlled for, the LFP of first-generation women increases with the share of immigrants from the same source country. This result might be explained by network effects, indicating that individuals who migrate to a country with a high proportion of people from the same ancestry will find it easier to gain information about the host country’s labor market and therefore be more likely to find a job shortly after arrival. While the above results highlight the importance of source-country culture and host- country conditions on the labor supply of first- and second-generation immigrants in Europe, we now check the sensitivity of these results to several robustness analyses. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 158 6.5 Sensitivity Analyses

6.5.1 Control for Partner Characteristics

As outlined above, for women living in couple households, labor supply decisions might be related to the characteristics of their partner for reasons related to assortative mating and joint labor supply decision-making within the household. In order to test whether our results are robust to controlling for the characteristics of a womans’s partner, we re-estimate Models 2 and 3 by now including the partner’s working hours and his highest level of education as additional control variables. The estimation results are displayed in Table 6.6. For second-generation female immigrants, husband’s working hours are positively correlated with their probability of participating in the labor market, which might be indicative of assortative mating with respect to similar preferences for market work. For first-generation immigrants, none such relationship is found. A possible explanation of this result is given by the family migration model, which was proposed by Baker and Benjamin (1997) and empirically tested, amongst others, by Basilio et al. (2009). The model predicts that immigrant women will initially take dead-end jobs to finance their husbands’ human capital investments and eventually drop out of the labor market or reduce their labor supply as their husbands’ labor market outcomes improve. The existence of such a substitutionary relationship between husband’s and wife’s labor supply might blur the positive correlation in the partner’s labor supply found for second-generation immigrants. However, as we do not know whether an immigrant woman’s partner is also an immigrant and the two migrated together, this interpretation is somewhat speculative. Neither for first- nor for second-generation immigrants, we find any correlation between the husbands’ highest level of education and their wives’ labor supply. This might be a result of the opposing effects of assortative mating and joint labor supply decision-making within the household. The higher the husband’s education (and income), the lower his wife’s incentives to work, but the higher the probability that his wife is well educated as well and will participate in the labor market. The most important finding, however, is that the effects of source-country FLFPR and host-country FLFPR on female immigrant labor supply are hardly affected by the inclusion of spousal characteristics. Both effects are similar in statistical significance and magnitude. The correlations of the other country and country-pair variables with immigrant women’s labor supply remain constant as well. An exception are the effects of host-country fertility and GDP per capita on the LFP of second-generation immigrants, which are smaller in magnitude and not significant any more. While the above analysis shows that our results are robust to controlling for the human CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 159 capital of a woman’s partner, the partner’s cultural background is also likely to play a role in her LFP decision. Fernández and Fogli (2009) show that a husband’s culture, as measured by the LFPR in his father’s country of birth, is an important determinant of his wife’s employment decision. More generally, Fernández et al. (2004) as well as Johnston et al. (2012) find evidence that an important factor explaining whether a man’s wife works is whether his own mother worked when he was growing up. The authors argue that a mother’s decision to work or not is influenced by her beliefs about women’s roles, which then have been transmitted to her son and influenced any household decision affecting his wife’s work outcome. Unfortunately, the ESS data do neither contain information on a partner’s cultural background (i.e., his immigration status and his country of origin), nor do they include information on his parent’s employment outcomes, making it impossible to control for any kind of assortative mating with respect to perceptions about gender roles. In particular, a woman who would like to work is presumably more likely to marry a man who would be in agreement with these choices. Given that the FLFPR in the source country serves as proxy for an individual’s beliefs regarding women’s role in society, we would assume that women from high FLFPR countries will be more likely to marry men from high FLFPR countries. Hence, we have to keep in mind that part of the effect of our cultural proxy might not capture a direct impact on an immigrant women’s decision to participate in the labor market. Rather, it might reflect an indirect effect of a woman’s mating decision, which is influenced by her beliefs regarding gender roles and ultimately effects her decision about market work.

6.5.2 Control for Parents’ Human Capital and Employment

As outlined above, evidence suggests that individual beliefs, preferences, and attitudes are transmitted from parents to children, and that this intergenerational transmission shapes the child’s economic outcomes (see, e.g., Fernández and Fogli, 2009; Fernández et al., 2004; Guiso et al., 2006). In particular, Johnston et al. (2012) find a strong correlation between mothers’ and children’s gender role attitudes and that a mother’s attitudes are strongly predictive of her daughter’s labor supply. However, the authors also show that even when controlling for the mother’s attitudes toward gender roles, her full-time employment status when her daughter was 5 years old has additional explanatory power in her daughter’s labor supply, suggesting that both parental attitudes and the parents’ actual behavior predict their children’s future labor supply decisions. In this respect, it is of interest to test whether the positive effect of source-country culture on immigrant labor supply still holds after controlling for the labor supply of the immigrant’s parents. Controlling for parental economic outcomes has the further advantage of disentangling CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 160 the effect of source-country culture from that of the immigrants’ own labor supply before migrating. For first-generation immigrants, work experience prior to their arrival in the host country might be positively correlated with the source country’s FLFPR. If this is true, the estimated effect of the latter does not only reflect the role of source-country culture, but partly contains the effect of the level of job-related human capital accumulated before migration. Having information on the human capital and labor supply of the immigrant’s parents can help to solve this problem, as parental economic behavior in the source country may serve as a proxy for the daughter’s labor supply before migrating. The estimation results of Models 2 and 3 including controls for the parents’ highest level of education and their labor market status when their daughter was 14 years are displayed in Table 6.7. For both first- and second-generation immigrants, we find that women whose mothers and fathers were employed when they were young are more likely to participate in the host-country’s labor market than those whose parents were not employed at this time.34 This result shows that the parents’ past employment behavior is a strong predictor of their daughter’s labor supply even if the daughter’s cultural background is controlled for. With respect to the parents’ education, we find women whose fathers have a tertiary degree to be more likely to participate in the labor market than those whose fathers have a secondary degree, while this relationship is not found for mothers and their daughters. Apart from that, the results show no clear relationship between the labor supply of immigrant women and their parents’ education.35 Our results further show that the estimated effects of the host- and source-country characteristics are robust to the inclusion of the controls for parental education and employment. In particular, the effects of host- and source-country FLFPR remain positive and significant. The latter result suggests that source-country culture plays an important role in the labor supply decisions of first- and second-generation immigrants even if the intergenerational transmission of human capital is controlled for.

6.5.3 Ratio of FLFPR to MLFPR

A possible concern when attempting to assess the effect of source-country culture on female immigrant labor supply is that such an approach might suffer from an omitted variable bias. If there exist any unobserved economic conditions in the source country (beyond the macroeconomic indicators we controlled for) that affect an immigrant woman’s labor supply decisions, and if these factors are further correlated with the source-country FLFPR,

34The respective marginal effects are positive across all specifications but only statistically significant for Model 2. 35We also estimated Models 2 and 3 including only the father’s characteristics and only the mother’s characteristics, respectively, in order to account for the fact that the parents’ educational degrees might be highly correlated. The results of these models are similar to those displayed in Table 6.7. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 161

then the estimated effect of our cultural proxy will be biased, as it contains the effect of these unobserved conditions as well. Although it is hard to think of any macroeconomic conditions that fulfill both conditions, we attempt to rule out the possibility of the existence of an omitted variable bias by checking the robustness of our cultural proxy. Following Blau and Kahn (2011) and Blau et al. (2011), we use the LFPR of women relative to men’s (i.e., FLFPR/MLFPR) instead of FLFPR as our cultural proxy. This relative measure is appropriate in that it captures the gender division of labor explicitly. If there exist any unobserved macroeconomic conditions correlated with a country’s FLFPR, these factors must differently affect the LFPR of men and women in order to still bias our estimates. A further advantage of using the ratio of FLFPR to MLFPR is that it implicitly adjusts for problems in measuring the labor force, particularly at different levels of economic development, at least to the extent that such problems affect men’s and women’s measured participation rates similarly (Blau et al., 2011). We apply the same robustness check to Model 3, with the idea that the effect of FLFPR/MLFPR shows us the role of the labor market behavior of native women in immigrant women’s labor supply net of any host-country conditions that affect male LFP as well. The estimation results of Model 2 and 3 using FLFPR/MLFPR as our explanatory variable of interest are displayed in Table 6.8. We find that the ratio of the female to the male LFPR in the immigrants’ source country is significantly positively related to the LFP of first- and second-generation immigrant women. Moreover, the effects of the other source-country characteristics on female immigrant labor supply remain similar in significance and magnitude. These results indicate that the correlation between source- country FLFPR and immigrant women’s labor supply is not due to unobserved economic conditions that are correlated with the labor market activity in the immigrants’ source country. Our results further show a strong positive correlation between the ratio of the female to the male LFPR in the immigrants’ host country and the probability of first- and second-generation women to participate in the labor market. As FLFPR and MLFPR, respectively, represent the aggregated LFP decisions of women and men living in the immigrants’ host country, which depend on a variety of individual and country-related characteristics, the ratio of the two variables can be thought of as representing only those factors that are relevant to the LFP decisions of women, but not of men. A positive correlation between this aggregate variable and immigrant women’s labor supply therefore provides some further evidence that the LFP decisions of immigrant women are affected by similar country-specific conditions as those of native women, and thus immigrant women assimilate to the labor market behavior of natives. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 162

6.5.4 Source-Country Characteristics at Year of Migration

While in the above analyses of the role of source-country characteristics in immigrant women’s labor supply the aggregated source-country variables refer to the year of observation, we know check the robustness of our results by assigning first-generation immigrants source-country values based on the year the immigrants left their source country, as was done by Bisin et al. (2011), Blau and Kahn (2011), and Blau et al. (2011). That way, these values reflect the norms and values the immigrants grew up with and carry to their host country. We calculate the year the immigrant left the home country by using information on the year of observation and the immigrant’s years since arrival in the host country.36 Since the latter is not a continuous variable but is subdivided in predefined categories, we set years since migration equal to the mid-point of each interval and to the lower bound of the top interval (i.e., 20 years). Thus, our source-country data for first-generation immigrants now span the years 1982 to 2011. The estimation results of Model 2 using past instead of current values of the source- country characteristics for first-generation immigrants are displayed in Table 6.9. Again, we find a significant positive correlation between source-country FLFPR and immigrant women’s probability of participating in the labor market. The magnitude of this effect is similar to the effect of FLFPR measured at time of observation (see Table 6.4). Hence, using past instead of current values of our cultural proxy does not alter the results substantially. This result is consistent with the finding of Fernández and Fogli (2009), who show that both fertility rates and FLFPR are strongly correlated over time, such that the choice over which point of time to take these values from is of minor relevance. Apparently, this argument does not hold true for our education variable, as the positive correlation between the average years of schooling of the source country’s population and immigrant women’s labor supply becomes insignificant once past values of the former variable are used.

6.5.5 Bias-Reduced Linearization of Standard Errors

As emphasized by Moulton (1986, 1990), ignoring within-group dependence in standard errors that appears whenever estimating the eﬀects of aggregate explanatory variables on individual-speciﬁc response variables can underestimate true standard errors. The usual solution to address this problem is to calculate cluster-robust standard errors that permit heteroskedasticity and within-cluster error correlation. However, a practical limitation of

36As previous studies, we thereby implicitly assume that the year the immigrant left her home country equals the year she arrived in the host country. I.e., we assume that immigrants directly move from their source country to their destination country and thus ignore the possible case of repeat migration. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 163

inference with cluster-robust standard errors is that the asymptotic justification assumes that the number of clusters goes to infinity. With a small number of clusters the cluster- robust standard errors can still be downward biased, a problem that has been documented, amongst others, in Bell and McCaffrey (2002), Bertrand et al. (2004) and Cameron et al. (2008). While there is still no consensus in the literature on when the number of clusters is considered to be small37, we check the robustness of our results by using an alternative approach to clustering to correct for a correlation in the regression disturbances within source and host countries, respectively. In particular, we use a bias-reduced linearization (BRL) method as proposed by Bell and McCaffrey (2002). Using Monte-Carlo simulations, the authors show that even if the number of clusters is small (20 clusters), BRL produces unbiased variance estimates in the event that errors are i.i.d., and it greatly reduces bias otherwise. As their approach is only applicable to linear regression models, we re-estimate equations (6.2) and (6.3) by OLS.38 In order to assess the difference between the cluster- robust and the BRL standard errors, we first estimate OLS regressions with standard errors clustered at the source- and host-country level, respectively (top of Table 6.10). In a second step, the standard errors of the respective models are estimated by the BRL method (bottom of Table 6.10). Regarding Model 2, we indeed find the BRL standard errors to exceed the cluster- robust standard errors for all explanatory variables. On average, BRL estimates are 35 percent larger than the respective cluster-robust estimates. For our variable of interest, the FLFPR in the immigrants’ source country, the deviation of the BRL standard errors is somewhat smaller (25% for first-generation immgrants and 32% for second-generation immigrants), and the respective coefficients are still significantly different from zero. The same applies to the coefficients for the FLFPR in the immigrants’ host country (Model 3), whose BRL standard errors only slightly exceed those obtained by OLS (22% and 3% for first- and second-generation immgrants, respectively). However, the results for Model 3 also show that for some coefficients, BRL standard errors are even smaller than the respective cluster-robust standard errors. Bell and McCaffrey (2002) explain this result by the fact that linearization methods, as other non-parametric variance estimators, can produce estimators with high variance under certain conditions. The authors therefore conclude that BRL methods will reduce, but not completely solve the inference problem for multi-stage samples with small cluster sizes. Hence, while the above results show that

37If indicated at all, the critical number of clusters to assure the unbiasedness of cluster-robust standard errors ranges between 20 and 50. 38While in the context of linear regression models several bias corrections have been proposed in the literature, comparable solutions for non-linear regression models are scarce. For an overview of methods attempting to address the problem of both intraclass and serial correlation in standard errors, see Angrist and Pischke (2009). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 164

our conclusions are consistent with the inference based on an alternative approach to clustering, we will follow previous literature (e.g., Bisin et al., 2011; Blau and Kahn, 2011; Fernández and Fogli, 2009) and cluster standard errors at the source- and host-country level, respectively, in our main regressions.

6.6 Conclusion

In the present paper, we focus on an important aspect of migration and integration policy: the labor supply of first- and second-generation female immigrants. In particular, we investigate the extent to which home- and host-country characteristics affect immigrant women’s labor supply in Europe. Our contributions to the literature are manifold. While previous literature on the role of source-country culture in female immigrant labor market behavior has exclusively focused on the U.S., we complement the existing literature by providing first evidence on this relationship for Europe. The use of cross-country data further allows us to investigate the role of host-country characteristics in immigrant women’s labor supply decisions, a topic that has so far been neglected by previous research. Lastly, we conduct our analysis separately for first- and second-generation immigrants to shed further light on cultural and economic assimilation patterns. Using data from the European Social Survey 2002-2011 covering immigrants in 26 European countries, we find that the labor supply of both first- and second-generation immigrants is positively associated with the FLFPR in their (parents’) source country. This result supports previous evidence for immigrants in the U.S. and suggests that immigrant women’s labor supply is affected by preferences and beliefs regarding women’s roles in society in her source country. The effect of this cultural proxy on the labor supply of immigrant women is robust to controlling for spousal characteristics, parental characteristics, and a variety of source-country characteristics. Moreover, we find evidence for a strong positive correlation between the FLFPR in the immigrant’s host country and immigrant women’s decision to participate in the labor market. This result suggests that immigrant women adapt to the culture, institutions, and economic conditions in their host country and that way assimilate to the work behavior of natives. Again, this result is robust to various sensitivity analyses. Our results have important policy implications. As the native-born working-age population declines in many European countries, issues on the financing and the fiscal sustainability of the welfare state capture increasing attention. As a result, the active recruitment of high-skilled immigrants as well as the integration of recent immigrants into the host countries’ labor markets have become important policy goals within Europe (European Commission, 2010b). The latter aspect is especially relevant for immigrant CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 165 women, whose formal labor market participation is still on a considerably low level. For the effective design of such policies, however, knowledge about whether and to what extent immigrant women’s labor supply is shaped by their cultural background on the one hand, and the cultural, economic, and institutional conditions in the host country on the other hand, is of great interest. Our finding that the labor supply of immigrant women is strongly related to the FLFPR in their host country reveals that host-country conditions indeed matter for immigrant women’s decision to participate in the labor market. This suggests that integration and labor market policies that aim at increasing the labor market attachment of immigrants can indeed be a successful tool in stimulating the labor supply of immigrant women in Europe. However, our results also suggest that the success of such policies is likely to vary depending on the immigrants’ cultural background. In addition to the conditions of their host country, the preferences and beliefs held in their source country strongly determine the LFP of female immigrants. This suggests that integration policies alone might be of limited effectiveness in achieving the envisaged goal. Rather, the balance between tailored integration policies on the one hand, and selective immigration policies on the other hand, might be a successful tool in increasing the labor market attachment of immigrants in Europe. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 166 Figures

Figure 6.1: Female Labor Force Participation Rate (2011, Age 15-64) Source: ILO. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 167 .015 .01 .005 0 Effect on Pr (Labor force participation) -.005 Within last year 6-10 years ago More than 20 years ago 1-5 years ago 11-20 years ago Years since migration

95% confidence intervals are shown. Average marginal effect of source-country FLFPR is 0.0021 (StdE. = 0.0007; z-value = 2.82).

Figure 6.2: Effect of Source-Country FLFPR by Years since Migration .03 .02 .01 0 -.01 Effect on Pr (Labor force participation) -.02

Within last year 6-10 years ago More than 20 years ago 1-5 years ago 11-20 years ago Years since migration

95% confidence intervals are shown. Average marginal effect of host-country FLFPR is 0.0061 (StdE. = 0.0015; z-value = 4.20).

Figure 6.3: Effect of Host-Country FLFPR by Years since Migration CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 168 Tables

Table 6.1: Descriptive Statistics – Individual Variables 1st-Generation 2nd-Generation Native Immigrants Immigrants Women

Mean StdD Mean StdD Mean StdD Participates in the labor market 0.647 0.478 0.705 0.456 0.688 0.463 Age 40.748 9.343 42.783 9.380 42.924 9.498 Highest level of education Primary education 0.347 0.476 0.286 0.452 0.339 0.473 Secondary education 0.286 0.452 0.386 0.487 0.358 0.479 Tertiary education 0.362 0.480 0.325 0.468 0.301 0.459 Other education 0.005 0.073 0.003 0.055 0.002 0.046 Partner in household 0.746 0.435 0.698 0.459 0.735 0.441 No. of children in household 0.732 0.977 0.626 0.940 0.586 0.899 Youngest child 0-2 0.115 0.319 0.093 0.290 0.086 0.280 Youngest child 3-5 0.115 0.319 0.091 0.288 0.085 0.279 Population density Densely populated 0.410 0.492 0.358 0.479 0.292 0.455 Medium populated 0.356 0.479 0.346 0.476 0.351 0.477 Thinly populated 0.234 0.424 0.296 0.457 0.357 0.479 Years since migration Less than 1 year 0.022 0.146 –––– 1 to 5 years 0.157 0.364 –––– 6 to 10 years 0.176 0.381 –––– 11 to 20 years 0.237 0.425 –––– More than 20 years 0.408 0.491 –––– Migrated after age 18 0.828 0.377 –––– Speaks host-country language 0.841 0.366 –––– Both parents migrants – – 0.299 0.458 –– Partner characteristics a Working hours 34.980 19.077 34.920 19.031 35.663 19.353 Education Primary education 0.312 0.463 0.268 0.443 0.331 0.471 Secondary education 0.325 0.469 0.371 0.483 0.365 0.482 Tertiary education 0.344 0.475 0.348 0.476 0.290 0.454 Other education 0.019 0.136 0.014 0.116 0.014 0.116 Parents characteristics a Father employed at age 14 0.912 0.283 0.922 0.268 0.935 0.247 Father’s Education Primary education 0.559 0.497 0.544 0.498 0.594 0.491 Secondary education 0.204 0.403 0.259 0.438 0.255 0.436 Tertiary education 0.221 0.415 0.186 0.389 0.140 0.347 Other education 0.015 0.123 0.011 0.104 0.010 0.102 Mother employed at age 14 0.481 0.500 0.577 0.494 0.547 0.498 Mother’s Education Primary education 0.661 0.474 0.671 0.470 0.697 0.460 Secondary education 0.177 0.381 0.211 0.408 0.217 0.412 Tertiary education 0.147 0.354 0.110 0.313 0.076 0.265 Other education 0.015 0.123 0.009 0.093 0.010 0.099 Observations 5,187 3,064 53,090 Notes: – aPartner and parents characteristics are calculated for a reduced sample size. Partner characteristics are shown for households with partner only. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 169

Table 6.2: Descriptive Statistics – Aggregated Variables 1st-Generation 2nd-Generation Immigrants Immigrants

Source Country Host Country Source Country Host Country Mean/StdD Mean/StdD Mean/StdD Mean/StdD Measured at time of observation Source-/host-country characteristics FLFP rate (in %) 63.716 76.783 64.154 77.628 (21.822) (9.577) (21.537) (10.353) MLFP rate (in %) 90.038 91.300 88.706 90.126 (8.153) (6.853) (9.267) (8.776) FLFPR/MLFPR 70.622 83.957 72.103 85.929 (23.090) (7.861) (22.284) (6.975) Total fertility rate 1.940 1.607 1.689 1.686 (0.740) (0.263) (0.403) (0.280) GDP per capita (in USD 1,000) 14.302 35.002 20.362 34.191 (15.205) (8.906) (15.878) (10.196) Average years of schooling 9.538 – 10.393 – (2.721) (2.277) Unemployment rate (in %) – 8.214 – 7.886 (3.570) (2.909) Total migrant stock (% of population) – 11.601 – 10.895 (3.742) (3.928) MIPEX: Labor market mobility – 66.219 –– (15.883) Relationship between source and host country Source-country migrant stock (% of population) 1.049 1.225 (1.784) (2.111) Colonial ties 0.287 0.366 (0.452) (0.482) Geographic distance (in 1,000 km) 3.026 1.412 (3.320) (1.941) Genetic distance 0.327 0.186 (0.512) (0.341) Linguistic distance 79.923 77.129 (30.692) (30.365) Right of free movement of workers 0.325 – (0.469)

Measured at time of migration Source-country characteristics FLFP rate (in %) 58.289 ––– (23.215) Total fertility rate 2.439 ––– (1.271) GDP per capita (in USD 1,000) 10.829 ––– (11.898) Average years of schooling 7.960 ––– (3.208) Relationship between source and host country Source-country migrant stock (% of population) 1.030 – (2.077) Observations 5,187 5,187 3,064 3,064 Note: – Time of observation refers to the years 2002 to 2011, while time of migration spans the years 1982 to 2011. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 170

Table 6.3: Model 1 – Source- and Host-Country Fixed Effects 1st-Generation 2nd-Generation Immigrants Immigrants

ME StdE ME StdE Age group (Ref.: Age 36-45) Age 26-35 −0.0346 (0.0313) −0.0099 (0.0362) Age 46-59 −0.1300† (0.0322) −0.1363† (0.0347) Highest level of education (Ref.: Secd. education) Primary education −0.1028† (0.0299) −0.0853∗∗ (0.0366) Tertiary education 0.0678∗∗ (0.0281) 0.1030† (0.0306) Partner in household −0.1288† (0.0249) 0.0371 (0.0306) No. of children in household −0.0842† (0.0150) −0.0859† (0.0181) Youngest child 0-2 −0.1867† (0.0442) −0.1862∗∗∗ (0.0616) Youngest child 3-5 −0.0089 (0.0393) −0.0710 (0.0543) Population density (Ref.: Medium populated) Densely populated 0.0225 (0.0255) 0.0513∗ (0.0302) Thinly populated 0.0162 (0.0290) −0.0046 (0.0321) Years since migration (Ref.: > 20 years) Less than 1 year −0.1701∗ (0.0962) –– 1 to 5 years −0.0980∗∗ (0.0453) –– 6 to 10 years −0.0313 (0.0402) –– 11 to 20 years 0.0443 (0.0305) –– Migrated after age 18 −0.0171 (0.0384) –– Speaks host-country language 0.1198† (0.0355) –– Both parents migrants – – 0.0104 (0.0305) Relationship between source and host country Source-country migrant stock (% of population) 0.0155 (0.0106) −0.0150 (0.0125) Colonial ties 0.0018 (0.0480) 0.0557 (0.0556) Geographic distance (in 1,000km) 0.0311 (0.0259) 0.0288 (0.0507) Genetic distance 0.1150 (0.1458) −0.4084 (0.2918) Linguistic distance 0.0003 (0.0008) 0.0008 (0.0009) Right of free movement of workers 0.1137∗ (0.0580) –– Host-country FE yes yes Source-country FE yes yes Year dummies yes yes Log likelihood -2278.8 -1447.3 Pseudo R2 0.144 0.103 Observations 5,187 3,064 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Robust standard errors in parentheses. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 171

Table 6.4: Model 2 – Source-Country Characteristics 1st-Generation 2nd-Generation Immigrants Immigrants

ME StdE ME StdE Age group (Ref.: Age 36-45) Age 26-35 −0.0299 (0.0412) −0.0042 (0.0344) Age 46-59 −0.0835∗∗∗ (0.0272) −0.0976† (0.0255) Highest level of education (Ref.: Secd. education) Primary education −0.0921† (0.0272) −0.0776 (0.0477) Tertiary education 0.0652∗∗ (0.0292) 0.1015∗∗∗ (0.0380) Partner in household −0.1168† (0.0267) 0.0367∗ (0.0201) No. of children in household −0.0862† (0.0136) −0.0906† (0.0140) Youngest child 0-2 −0.1772† (0.0448) −0.1790∗∗∗ (0.0611) Youngest child 3-5 −0.0060 (0.0431) −0.0708 (0.0646) Population density (Ref.: Medium populated) Densely populated 0.0254 (0.0278) 0.0454 (0.0302) Thinly populated 0.0168 (0.0236) −0.0069 (0.0249) Years since migration (Ref.: > 20 years) Less than 1 year −0.1858∗∗ (0.0898) –– 1 to 5 years −0.0988∗∗∗ (0.0381) –– 6 to 10 years −0.0410 (0.0333) –– 11 to 20 years 0.0357 (0.0264) –– Migrated after age 18 −0.0049 (0.0353) –– Speaks host-country language 0.1096† (0.0256) –– Both parents migrants – – 0.0243 (0.0346) Source-country characteristics FLFP rate (in %) 0.0022∗∗∗ (0.0008) 0.0039∗∗∗ (0.0014) Total fertility rate 0.0340 (0.0250) 0.1020∗∗ (0.0429) GDP per capita (in USD 1,000) −0.0047† (0.0013) −0.0022∗∗ (0.0010) Average years of schooling 0.0201∗∗∗ (0.0071) −0.0056 (0.0109) Relationship between source and host country Source-country migrant stock (% of population) 0.0083 (0.0083) −0.0080 (0.0061) Colonial ties 0.0511∗ (0.0289) 0.0630∗∗∗ (0.0197) Geographic distance (in 1,000km) 0.0069 (0.0054) −0.0003 (0.0126) Genetic distance 0.0054 (0.0386) −0.0767 (0.0476) Linguistic distance 0.0005 (0.0005) 0.0004 (0.0004) Right of free movement of workers 0.1543† (0.0361) –– Host-country FE yes yes Year dummies yes yes Log likelihood -2327.5 -1451.3 Pseudo R2 0.125 0.101 Observations 5,187 3,064 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the source-country level. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 172

Table 6.5: Model 3 – Host-Country Characteristics 1st-Generation 2nd-Generation Immigrants Immigrants

ME StdE ME StdE Age group (Ref.: Age 36-45) Age 26-35 −0.0289 (0.0546) 0.0169 (0.0312) Age 46-59 −0.0702∗ (0.0372) −0.0460∗∗ (0.0211) Highest level of education (Ref.: Secd. education) Primary education −0.0992† (0.0236) −0.0765∗∗ (0.0352) Tertiary education 0.0602† (0.0178) 0.0963† (0.0196) Partner in household −0.1291† (0.0286) 0.0328∗ (0.0195) No. of children in household −0.0909† (0.0242) −0.0955† (0.0173) Youngest child 0-2 −0.1780† (0.0376) −0.1662∗ (0.0881) Youngest child 3-5 −0.0078 (0.0329) −0.0611 (0.0581) Population density (Ref.: Medium populated) Densely populated 0.0175 (0.0177) 0.0555∗∗∗ (0.0190) Thinly populated 0.0191 (0.0216) 0.0019 (0.0216) Years since migration (Ref.: > 20 years) Less than 1 year −0.1586 (0.1177) –– 1 to 5 years −0.0717∗ (0.0371) –– 6 to 10 years −0.0155 (0.0285) –– 11 to 20 years 0.0391∗∗∗ (0.0144) –– Migrated after age 18 −0.0225 (0.0606) –– Speaks host-country language 0.1057† (0.0294) –– Both parents migrants – – 0.0158 (0.0202) Host-country characteristics FLFP rate (in %) 0.0065† (0.0012) 0.0094† (0.0006) Total fertility rate −0.0357 (0.0638) 0.1006∗∗ (0.0420) GDP per capita (in USD 1,000) −0.0002 (0.0032) −0.0048∗∗ (0.0023) Unemployment rate (in %) 0.0010 (0.0047) −0.0027 (0.0042) Total migrant stock (% of population) −0.0059 (0.0056) 0.0052 (0.0046) MIPEX: Labor market mobility 0.0005 (0.0013) –– Relationship between source and host country Source-country migrant stock (% of population) 0.0127∗∗ (0.0059) −0.0037 (0.0049) Colonial ties 0.0055 (0.0269) 0.0285 (0.0272) Geographic distance (in 1,000km) 0.0543∗ (0.0291) 0.0120 (0.0235) Genetic distance −0.0074 (0.0608) −0.1084∗∗∗ (0.0410) Linguistic distance −0.0001 (0.0006) 0.0011∗∗ (0.0004) Right of free movement of workers 0.1210∗∗∗ (0.0439) –– Source-country FE yes yes Year dummies yes yes Log likelihood -2284.9 -1419.4 Pseudo R2 0.141 0.121 Observations 5,187 3,064 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the host-country level. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 173

Table 6.6: Models 2 & 3 – Controlling for Partner Characteristics

Model 2 Model 3

1st-Generation 2nd-Generation 1st-Generation 2nd-Generation Immigrants Immigrants Immigrants Immigrants ME/StdE ME/StdE ME/StdE ME/StdE

Partner characteristics Working hours 0.0006 0.0015∗∗ 0.0005 0.0013∗∗∗ (0.0006) (0.0006) (0.0003) (0.0004) Education (Ref.: Secd. education) Primary education 0.0174 0.0078 0.0339 0.0027 (0.0338) (0.0366) (0.0227) (0.0253) Tertiary education −0.0036 −0.0346 −0.0087 −0.0411 (0.0247) (0.0363) (0.0118) (0.0251) Other education −0.0036 0.0534 0.0604 0.1174∗ (0.0852) (0.1183) (0.0711) (0.0637) Source-/host-country characteristics FLFP rate (in %) 0.0019∗∗ 0.0039∗∗∗ 0.0065† 0.0088† (0.0008) (0.0014) (0.0011) (0.0007) Total fertility rate 0.0417 0.1096∗∗ −0.0516 0.0803 (0.0264) (0.0460) (0.0665) (0.0514) GDP per capita (in USD 1,000) −0.0048† −0.0021∗∗ 0.0007 −0.0034 (0.0014) (0.0010) (0.0031) (0.0027) Average years of schooling 0.0195∗∗∗ −0.0070 –– (0.0074) (0.0121) Unemployment rate (in %) – – 0.0024 0.0009 (0.0046) (0.0055) Total migrant stock (% of population) – – −0.0072 0.0031 (0.0058) (0.0054) MIPEX: Labor market mobility – – 0.0003 – (0.0013) Relationship between source and host country Source-country migrant stock (% of population) 0.0078 −0.0065 0.0120∗ 0.0005 (0.0088) (0.0070) (0.0065) (0.0060) Colonial ties 0.0626∗∗ 0.0660∗∗∗ 0.0108 0.0202 (0.0305) (0.0230) (0.0250) (0.0307) Geographic distance 0.0066 0.0021 0.0580∗ 0.0255 (0.0060) (0.0156) (0.0296) (0.0259) Genetic distance −0.0084 −0.0919 0.0284 −0.1114∗∗ (0.0384) (0.0585) (0.0662) (0.0498) Linguistic distance 0.0003 0.0003 −0.0004 0.0012∗∗∗ (0.0006) (0.0004) (0.0006) (0.0004) Right of free movement of workers 0.1525† – 0.1014∗∗ – (0.0385) (0.0454) Individual controls yes yes yes yes Host-country FE yes yes no no Source-country FE no no yes yes Year dummies yes yes yes yes

Log likelihood -2189.2 -1356.2 -2140.2 -1326.9 Pseudo R2 0.121 0.111 0.141 0.130 Observations 4,805 2,819 4,805 2,819 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the source- country level (Model 2) and host-country level (Model 3), respectively. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 174

Table 6.7: Models 2 & 3 – Controlling for Parents Characteristics

Model 2 Model 3

1st-Generation 2nd-Generation 1st-Generation 2nd-Generation Immigrants Immigrants Immigrants Immigrants ME/StdE ME/StdE ME/StdE ME/StdE

Parents characteristics Father employed at age 14 0.0455 0.0501 0.0680∗∗ 0.0688∗ (0.0444) (0.0436) (0.0313) (0.0409) Father’s education (Ref.: Secd. education) Primary education 0.0246 0.0775∗∗∗ 0.0273 0.0674∗∗∗ (0.0316) (0.0269) (0.0252) (0.0228) Tertiary education 0.0702∗∗ 0.1157† 0.0781∗∗∗ 0.1043† (0.0338) (0.0341) (0.0279) (0.0284) Other education −0.1206 −0.0060 −0.1543 0.0618 (0.1409) (0.1594) (0.2092) (0.1307) Mother employed at age 14 0.0273 0.0640∗∗ 0.0355† 0.0539∗∗ (0.0206) (0.0305) (0.0087) (0.0268) Mother’s education (Ref.: Secd. education) Primary education 0.0626∗∗ −0.0503 0.0647∗∗ −0.0632∗∗ (0.0307) (0.0425) (0.0280) (0.0301) Tertiary education −0.0251 −0.0914 −0.0301 −0.1123∗∗∗ (0.0419) (0.0596) (0.0324) (0.0362) Other education 0.1641∗∗ −0.0229 0.2044∗∗ −0.0636 (0.0716) (0.2102) (0.0876) (0.1762) Source-/host-country characteristics FLFP rate (in %) 0.0025† 0.0042∗∗ 0.0063† 0.0104† (0.0008) (0.0017) (0.0012) (0.0007) Total fertility rate 0.0244 0.1489∗∗∗ −0.0658 0.1103∗∗∗ (0.0241) (0.0565) (0.0704) (0.0384) GDP per capita (in USD 1,000) −0.0041∗∗∗ −0.0037∗∗∗ 0.0005 −0.0071∗∗∗ (0.0014) (0.0012) (0.0029) (0.0025) Average years of schooling 0.0177∗∗ −0.0007 –– (0.0075) (0.0169) Unemployment rate (in %) – – −0.0004 −0.0089∗∗ (0.0049) (0.0042) Total migrant stock (% of population) – – −0.0087 0.0057 (0.0054) (0.0047) MIPEX: Labor market mobility – – 0.0002 – (0.0014) Relationship between source and host country Source-country migrant stock (% of population) 0.0141∗ −0.0096 0.0176∗∗∗ −0.0062 (0.0075) (0.0083) (0.0065) (0.0058) Colonial ties 0.0282 0.0359 −0.0223 −0.0146 (0.0304) (0.0272) (0.0297) (0.0342) Geographic distance 0.0091 −0.0053 0.0623∗ 0.0098 (0.0066) (0.0139) (0.0361) (0.0251) Genetic distance 0.0275 −0.0776 −0.0268 −0.1064∗ (0.0372) (0.0559) (0.0547) (0.0634) Linguistic distance 0.0006 0.0005 −0.0001 0.0010∗∗ (0.0006) (0.0005) (0.0005) (0.0004) Right of free movement of workers 0.1392† – 0.0613 – (0.0377) (0.0450) Individual controls yes yes yes yes Host-country FE yes yes no no Source-country FE no no yes yes Year dummies yes yes yes yes

Log likelihood -1981.6 -1158.6 -1936.6 -1121.2 Pseudo R2 0.126 0.116 0.145 0.144 Observations 4,545 2,628 4,545 2,628 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the source- country level (Model 2) and host-country level (Model 3), respectively. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 175

Table 6.8: Models 2 & 3 – Ratio of FLFPR to MLFPR Model 2 Model 3 1st-Generation 2nd-Generation 1st-Generation 2nd-Generation Immigrants Immigrants Immigrants Immigrants ME/StdE ME/StdE ME/StdE ME/StdE Source-/host-country characteristics FLFPR/MLFPR 0.0018∗∗∗ 0.0031∗∗ 0.0048† 0.0097† (0.0007) (0.0014) (0.0013) (0.0014) Total fertility rate 0.0346 0.1000∗∗ −0.0399 0.0508 (0.0252) (0.0496) (0.0614) (0.0827) GDP per capita (in USD 1,000) −0.0047† −0.0021∗∗ −0.0003 −0.0038 (0.0013) (0.0010) (0.0017) (0.0025) Average years of schooling 0.0205∗∗∗ −0.0026 –– (0.0071) (0.0100) Unemployment rate (in %) – – −0.0004 −0.0038 (0.0037) (0.0060) Total migrant stock (% of population) – – −0.0041 0.0063 (0.0038) (0.0047) MIPEX: Labor market mobility – – 0.0007 – (0.0007) Individual controls yes yes yes yes Host-country FE yes yes no no Source-country FE no no yes yes Year dummies yes yes yes yes Bilateral variables yes yes yes yes Log likelihood -2329.6 -1457.4 -2297.9 -1440.0 Pseudo R2 0.125 0.097 0.137 0.108 Observations 5,187 3,064 5,187 3,064 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the source-country level (Model 2) and host-country level (Model 3), respectively. – Host-country population weights are applied.

Table 6.9: Model 2 – Source-Country Characteristics at Year of Migration 1st-Generation Immigrants ME StdE Source-country characteristics FLFP rate (in %) 0.0021∗∗ (0.0008) Total fertility rate 0.0190 (0.0155) GDP per capita (in USD 1,000) −0.0043∗∗∗ (0.0015) Average years of schooling 0.0136 (0.0092) Individual controls yes Host-country FE yes Year dummies yes Bilateral variables yes Log likelihood -2337.4 Pseudo R2 0.122 Observations 5,187 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – Standard errors are clustered at the source-country level. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 176

Table 6.10: Models 2 & 3 – Bias-Reduced Linearization of Standard Errors Model 2 Model 3 1st-Generation 2nd-Generation 1st-Generation 2nd-Generation Immigrants Immigrants Immigrants Immigrants Coef/StdE Coef/StdE Coef/StdE Coef/StdE OLS with clustered standard errors Source-/host-country characteristics FLFPR (in %) 0.0021∗∗∗ 0.0038∗∗∗ 0.0060† 0.0098† (0.0007) (0.0013) (0.0011) (0.0006) Total fertility rate 0.0285 0.1010∗∗ −0.0341 0.0919∗∗ (0.0230) (0.0416) (0.0579) (0.0367) GDP per capita (in USD 1,000) −0.0042† −0.0020∗∗ 0.0001 −0.0040∗ (0.0011) (0.0009) (0.0026) (0.0020) Average years of schooling 0.0173∗∗∗ −0.0043 –– (0.0065) (0.0098) Unemployment rate (in %) – – 0.0014 −0.0023 (0.0039) (0.0038) Total migrant stock (% of population) – – −0.0058 0.0042 (0.0047) (0.0040) MIPEX: Labor market mobility – – 0.0005 – (0.0011) Individual controls yes yes yes yes Host-country FE yes yes no no Source-country FE no no yes yes Year dummies yes yes yes yes Bilateral variables yes yes yes yes Log likelihood -3100.3 -1754.2 -3041.1 -1710.1 Adjusted R2 0.143 0.101 0.157 0.124 Observations 5,187 3,065 5,187 3,065

Bias-reduced linearization of standard errors Source-/host-country characteristics FLFPR (in %) 0.0021∗∗ 0.0038∗∗ 0.0060† 0.0098† (0.0009) (0.0017) (0.0013) (0.0006) Total fertility rate 0.0285 0.1010∗ −0.0341 0.0919 (0.0268) (0.0589) (0.0704) (0.0582) GDP per capita (in USD 1,000) −0.0042∗∗∗ −0.0020 0.0001 −0.0040∗∗ (0.0013) (0.0014) (0.0017) (0.0017) Average years of schooling 0.0173∗∗ −0.0043 –– (0.0086) (0.0157) Unemployment rate (in %) – – 0.0014 −0.0023 (0.0036) (0.0034) Total migrant stock (% of population) – – −0.0058∗∗ 0.0042∗ (0.0029) (0.0025) MIPEX: Labor market mobility – – 0.0005 – (0.0012) Individual controls yes yes yes yes Host-country FE yes yes no no Source-country FE no no yes yes Year dummies yes yes yes yes Bilateral variables yes yes yes yes Log likelihood – – – – Adjusted R2 –––– Observations 5,187 3,065 5,187 3,065 Notes: – † p < 0.001; ∗∗∗ p < 0.01; ∗∗ p < 0.05; ∗ p < 0.1. – The standard errors of the OLS regression are clustered at the source-country level (Model 2) and host-country level (Model 3), respectively. – Host-country population weights are applied. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 177 6.A Appendix

Table 6.A1: Explanatory Power of Source- & Host-Country Fixed Effects 1st-Generation Immigrants 2nd-Generation Immigrants Full Model Restricted Model Full Model Restricted Model Excl. SC-FE Excl. HC-FE Excl. SC-FE Excl. HC-FE R2 0.1739 0.1370 0.1617 0.1170 0.1032 0.1049 Semipartial R2 – 0.0369 0.0122 – 0.0138 0.0121 Expl. Power SC-FE 21.22% –– 11.80% –– Expl. Power HC-FE 7.02% –– 10.34% –– Observations 5,187 3,064 Notes: – Results are obtained from OLS regressions of Model 1. – Host-country population weights are applied. – The explanatory power of the source- and host-country fixed effects is computed as the difference between the R2 of the full model and the R2 of the respective restricted model. The values represent the proportion of the explained variance that can be explained by the sum of the source- and host-country fixed effects, respectively.

Table 6.A2: Explanatory Power of Source- & Host-Country Characteristics A. 1st-Generation Immigrants Model 2 Model 2 Model 3 Model 3 Restricted Restricted R2 0.1526 0.1364 0.1714 0.1610 Semipartial R2 – 0.0162 – 0.0104 Expl. Power SC Vars. 10.62% ––– Expl. Power HC Vars – – 6.07% – Host-country FE yes yes no no Source-country FE no no yes yes Host-country Vars. no no yes no Source-country Vars. yes no no no B. 2nd-Generation Immigrants Model 2 Model 2 Model 3 Model 3 Restricted Restricted R2 0.1151 0.1027 0.1401 0.1044 Diﬀ. in R2 – 0.0124 – 0.0357 Expl. Power SC Vars. 10.77% ––– Expl. Power HC Vars – – 25.48% – Host-country FE yes yes no no Source-country FE no no yes yes Host-country Vars. no no yes no Source-country Vars. yes no no no Notes: – Results are obtained from OLS regressions of Model 2 and 3. – Host-country population weights are applied. – The explanatory power of the source- and host-country variables is computed as the diﬀerence between the R2 of the full model and the R2 of the respective restricted model. The values represent the proportion of the explained variance that can be explained by the sum of the source- and host-country variables, respectively. CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 178

Table 6.A3: List of Source Countries 1st-Generation 2nd-Generation Immigrants Immigrants

Source Country Observations Frequency (in %) Observations Frequency (in %) Albania 121 2.33 –– Algeria 54 1.04 61 1.99 Argentina 32 0.62 –– Australia 36 0.69 –– Austria 49 0.94 72 2.35 Belgium 73 1.41 28 0.91 Bolivia 18 0.35 –– Brazil 111 2.14 –– Bulgaria 48 0.93 –– Canada 36 0.69 –– Chile 26 0.50 –– China 27 0.52 –– Colombia 33 0.64 –– Congo 32 0.62 –– Czechoslovakia 135 2.60 239 7.80 Denmark 38 0.73 35 1.14 DR Congo 15 0.29 –– Ecuador 41 0.79 –– Finland 104 2.01 95 3.10 France 224 4.32 123 4.01 Germany 385 7.42 310 10.12 Ghana 17 0.33 –– Greece 32 0.62 22 0.72 Hungary 38 0.73 89 2.90 India 67 1.29 28 0.91 Indonesia 32 0.62 64 2.09 Iran 49 0.94 –– Iraq 35 0.67 –– Ireland 26 0.50 73 2.38 Italy 141 2.72 286 9.33 Jamaica – – 17 0.55 Japan 16 0.31 –– Kenya 17 0.33 –– Mauritius 18 0.35 –– Morocco 112 2.16 47 1.53 Mozambique 18 0.35 –– Netherlands 66 1.27 49 1.60 Norway 31 0.60 32 1.04 Pakistan 33 0.64 –– Peru 20 0.39 –– Philippines 63 1.21 –– Poland 215 4.14 143 4.67 Portugal 188 3.62 31 1.01 Republic of Korea 16 0.31 –– Romania 152 2.93 59 1.93 South Africa 35 0.67 –– Spain 67 1.29 67 2.19 Sri Lanka 31 0.60 –– Sweden 90 1.74 34 1.11 Switzerland 31 0.60 16 0.52 Thailand 30 0.58 –– Tunisia 24 0.46 23 0.75 Turkey 179 3.45 72 2.35 United Kingdom 307 5.92 108 3.52 USA 98 1.89 48 1.57 USSR 755 14.56 582 18.99 Venezuela 19 0.37 –– Viet Nam 24 0.46 –– Yugoslavia 457 8.81 211 6.89 Total 5,187 100.00 3,064 100.00 Note: To form a consistent list of source countries, we aggregate source countries that split or combined over time (i.e., Czechoslovakia, the USSR, and Yugoslavia). CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 179 LABORSTA Internet. http://laborsta.ilo.org Development Indicators. http://data.worldbank.org/indicator/ SP.DYN.TFRT.IN National Accounts Main Aggregates Database. http://unstats.un.org/unsd/ snaama/introduction.asp http://www.barrolee.com 1982–2011 ILO Department of Statistics, 1982–2011 World Bank Database, World 1982–2011 United Nations Statistics Division, 1982–2011 Barro and Lee (2010). Macroeconomic Data – Sources and Descriptions to a woman ifbear she children were in to accordanceinterpolate live with missing to current values the age-specific for end fertility intervening of rates. years her from We childbearing theestimated years available rates and data. of changecountry. derived from available data for the respective age population that engages activelysize in of the the supply ofand labor labor services available market, to engage duringcalculated either in for a the by females production specified and ofaged males goods time-reference 26 by to 5-year period. 59 agefrom years. group for the We The the interpolate available population rates missing data.impute values are for missing When intervening linear years valuesavailable interpolation data using is for estimated other not age possible, rates groups we of in the change respective derived country. from by the population.national Data currency in using constant thebase prices annual year period-average in for exchange USD all rate are years. of converted the into attained by an averagemary, secondary, person and at tertiary). allgroup These for levels data the of are population schooling measured5-year aged intervals by combined 26 for 5-year (pri- to the age for 59 years years. intervening 1980–2010. years The We from data interpolateand are the missing extrapolate available values available for in data the for year 2011. the period 1980–2010 working or looking for work. It provides an indication of the relative When linear interpolation is not possible, we impute missing values using Table 6.A4: I. Source- & host-country variables FLFPR & MLFPR Labor force participation rate is the proportion of a country’s working- GDP per capita (in USD 1,000) Per capita GDP is GDP in constant 2005 prices in USD 1,000 divided Total fertility rate Total fertility rate represents the number of children that would be born Average years of schooling Average years of schooling represents the number of years of schooling Variable Description Years Source CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 180 Development Indicators. http://data.worldbank.org/indicator/ SL.UEM.TOTL.ZS Development Indicators. http://data.worldbank.org/indicator/ SM.POP.TOTL.ZS Migration Database. http: global-bilateral-migration-database http://www.cepii.fr/anglaisgraph/bdd/ distances.htm //data.worldbank.org/data-catalog/ 2010 http://www.mipex.eu Years Source constant Mayer and Zignago (2011). 2002–2011 World Bank Database, World 2002–2011 World Bank Database, World 1982–2011 World Bank Database, Global Bilateral Macroeconomic Data – Sources and Descriptions (Continued) ment rate from the available data for the year 2011. cators grouped intomigrant 6 workers are broad eligible policyin for areas. most the same sectors. opportunities Laboring The as that market index EU migrants mobility varies nationals have betweenavailable to measures more for 0 work rights if the and in years 100, thein 2004, with corresponding our 2007, higher policy and sample values area. 2010. mean- are MIPEX As is only some included of from the countries the included 2010 version onwards, we use 2010 national migrant stock by countryregister of records. birth in Weavailable data interpolate 10-year for missing intervals the values for period for 1980–2010 the intervening and years years extrapolate for from thetionship. the year 2011. than that in whichfive-year they intervals. live. We interpolate Theavailable missing data data values are for for available the intervening for period years the from 2000–2010 years the and 2000–2010 extrapolate at for the year 2011. values for all years. work but available for and seeking employment. We extrapolate the unemploy- 1980–2010. The data are mostly based on population censuses and population Table 6.A3: Unemployment rate Unemployment rate represents the share of the total labor force that is without MIPEX: Labor market mobility The Migrant Integration Policy Index (MIPEX) considers over 140 policy indi- II. Bilateral variables Source-country migrant stock Source-country migrant stock provides information on the host country’s inter- Colonial ties Binary variable that is unity if the country-pair have ever had a colonial rela- Total migrant stock International migrant stock is the number of people born in a country other Variable Description CHAPTER 6. FEMALE IMMIGRANT LABOR SUPPLY 181 http://www.cepii.fr/anglaisgraph/bdd/ distances.htm http://www.anderson.ucla.edu/faculty_ pages/romain.wacziarg http://www.eva.mpg.de constant Mayer and Zignago (2011). constant Spolaore and Wacziarg (2009). constant Bakker et al. (2009). 2002–2011 European Commission (2003, 2005) index ST F index is based on the ST F is associated with larger differences. ST F genetic distance index measures the genetic differences between pop- Macroeconomic Data – Sources and Descriptions (Continued) ST F distances between the countries. ulations as a fractionare of collected the by total Cavalli-Sforzafrequency genetic et variance. of al. The 128 (1994). genetic alleles distance related The data to 45 genes. By construction, the ranges between 0Genetic and distance 1; reports a the calculated higher distance divided by 1,000. matic Similarity Judgment ProgramPlanck (ASJP), Institute developed for by Evolutionaryphonetic the Anthropology, similarity German automatically between Max evaluates allcompare the of pairs the of world’saccording words languages. to having their The the pronunciation. basic same Foradditions idea each or meaning word is subtractions pair, in to are it necessary is twointo evaluated to different the how transform many languages same one word wordand in divided in one Levenshtein another language distance. language.of We use each The the country approach most to prevalent is calculate native called the language distance. normalized right of freefree movement movement of of workersemployed, workers in and permits to a workers resideit given to in generally source search any applies Member for country.there State to employment, is of all The a to clause the immigrants right about be EuropeanStates migrating a of transition Union. can within period be While before the employed workers onStates. European from equal, the non-discriminatory Union, new Citizens terms Member in ofSwitzerland the the have old the Member Member same Statesare right of treated of the as freedom old European of Member Economic movement and States Area these inside and countries the EEA. which uses the geographic coordinates of the capital cities for calculating the 1,000km. Geodesic distances are calculated following the great circle formula, Table 6.A2: Note: The macroeconomic indicators for the combined countries (i.e., Czechoslovakia, the USSR, and Yugoslavia) are calculated as a population-weighted average of the single-country values. Geographic distance Geographic distance is the geodesic distance between country capitals in Genetic distance The Linguistic distance The linguistic distance measure is drawn from linguistic research. The Auto- Right of free movement of workers Binary variable that is unity if citizens of a given host country underly the Variable Description Years Source 182

Bibliography

Adsera, A. and M. Pytlikova (2012). The role of language in shaping international migration. CReAM Discussion Paper No. 06/12.

Aggarwal, R., C. Kearney, and B. Lucey (2012). Gravity and culture in foreign portfolio investment. Journal of Banking & Finance 36 (2), 525–538.

Ahearne, A. G., W. L. Griever, and F. E. Warnock (2004). Information costs and home bias: an analysis of US holdings of foreign equities. Journal of International Economics 62 (2), 313–336.

Aldashev, A., J. Gernandt, and S. L. Thomsen (2009). Language usage, participation, employment and earnings: Evidence for foreigners in West Germany with multiple sources of selection. Labour Economics 16 (3), 330–341.

Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg (2003). Fraction- alization. Journal of Economic Growth 8 (2), 155–194.

Anderson, J. E. (1979). A Theoretical Foundation for the Gravity Equation. American Economic Review 69 (1), 106–116.

Anderson, J. E. (2011). The Gravity Model. Annual Review of Economics 3 (1), 133–160.

Anderson, J. E. and E. van Wincoop (2003). Gravity with Gravitas: A Solution to the Border Puzzle. American Economic Review 93 (1), 170–192.

Anderson, J. E. and E. van Wincoop (2004). Trade costs. Journal of Economic Litera- ture 42 (3), 691–751.

Anderson, J. E. and Y. V. Yotov (2010a). Specialization: Pro-and Anti-Globalizing, 1990-2002. NBER Working Paper No. 16301.

Anderson, J. E. and Y. V. Yotov (2010b). The Changing Incidence of Geography. American Economic Review 100 (5), 2157–2186. BIBLIOGRAPHY 183

Angrist, J. D. and V. Lavy (1997). The eﬀect of a change in language of instruction on the returns to schooling in Morocco. Journal of Labor Economics 15 (1), S48–S76.

Angrist, J. D. and J.-S. Pischke (2009). Mostly harmless econometrics: An empiricist’s companion. Princeton, NJ: Princeton University Press.

Antecol, H. (2000). An examination of cross-country diﬀerences in the gender gap in labor force participation rates. Labour Economics 7 (4), 409–426.

Antecol, H. (2001). Why Is There Interethnic Variation in the Gender Wage Gap? The Role of Cultural Factors. Journal of Human Resources 36 (1), 119–143.

Aviat, A. and N. Coeurdacier (2007). The geography of trade in goods and asset holdings. Journal of International Economics 71 (1), 22–51.

Baier, S. L. and J. H. Bergstrand (2009). Bonus vetus OLS: A simple method for approximating international trade-cost eﬀects using the gravity equation. Journal of International Economics 77 (1), 77–85.

Baker, M. and D. Benjamin (1997). The Role of the Family in Immigrants’ Labor-Market Activity: An Evaluation of Alternative Explanations. American Economic Review 87 (4), 705–727.

Bakker, D., A. Müller, V. Velupillai, S. Wichmann, C. H. Brown, P. Brown, D. Egorov, R. Mailhammer, A. Grant, and E. W. Holman (2009). Adding typology to lexicostatistics: A combined approach to language classiﬁcation. Linguistic Typology 13 (1), 169–181.

Baldwin, R. and D. Taglioni (2006). Gravity for dummies and dummies for gravity equations. NBER Working Paper No. 12516.

Baltagi, B. H., P. H. Egger, and M. Pfaﬀermayr (2003). A generalized design for bilateral trade ﬂow models. Economics Letters 80 (3), 391–397.

Barro, R. J. and J.-W. Lee (2010). A New Data Set of Educational Attainment in the World, 1950–2010. NBER Working Paper No. 15902.

Basilio, L., T. K. Bauer, and M. Sinning (2009). Analyzing the labor market activity of immigrant families in Germany. Labour Economics 16 (5), 510–520.

Bauer, T. K., B. Dietz, K. F. Zimmermann, and E. Zwintz (2005). German migration: Development, assimilation, and labour market eﬀects. In K. F. Zimmermann (Ed.), European Migration: What Do We Know, pp. 197–261. Oxford, UK: Oxford University Press. BIBLIOGRAPHY 184

Bauer, T. K., G. S. Epstein, and I. N. Gang (2005). Enclaves, language, and the location choice of migrants. Journal of Population Economics 18 (4), 649–662.

Bauer, T. K., M. Lofstrom, and K. F. Zimmermann (2000). Immigration policy, assimilation of immigrants, and natives’ sentiments towards immigrants: evidence from 12 OECD countries. Swedish Economic Policy Review 7 (2), 11–53.

BBC (2010). Merkel says German multicultural society has failed (2010/10/17). http: //www.bbc.co.uk/news/world-europe-11559451.

BBC (2011). State multiculturalism has failed, says David Cameron (2011/02/05). http: //www.bbc.co.uk/news/uk-politics-12371994.

Becker, G. S. (1965). A Theory of the Allocation of Time. Economic Journal 75 (299), 493–517.

Bell, R. M. and D. F. McCaﬀrey (2002). Bias Reduction in Standard Errors for Linear and Generalized Linear Models with Multi-Stage Samples. Survey Methodology 28 (2), 169–181.

Belot, M. V. K. and S. Ederveen (2012). Cultural barriers in migration between OECD countries. Journal of Population Economics 25 (3), 1077–1105.

Belot, M. V. K. and T. J. Hatton (2012). Immigrant Selection in the OECD. Scandinavian Journal of Economics 114 (4), 1105–1128.

Bergstrand, J. H. (1985). The Gravity Equation in International Trade: Some Microeco- nomic Foundations and Empirical Evidence. Review of Economics and Statistics 67 (3), 474–481.

Bergstrand, J. H. and P. H. Egger (2011). Gravity Equations and Economic Frictions in the World Economy. In D. Bernhofen, R. Falvey, D. Greenaway, and U. Kreickemeier (Eds.), Palgrave Handbook of International Trade, pp. 532–570. London, UK: Palgrave Macmillan.

Bertrand, M., E. Duflo, and S. Mullainathan (2004). How Much Should We Trust Differences-In-Differences Estimates? Quarterly Journal of Economics 119 (1), 249–275.

Beugelsdijk, S. and B. Frijns (2010). A cultural explanation of the foreign bias in international asset allocation. Journal of Banking & Finance 34 (9), 2121–2131.

Bisin, A., E. Patacchini, T. Verdier, and Y. Zenou (2011). Ethnic identity and labour market outcomes of immigrants in Europe. Economic Policy 26 (65), 57–92. BIBLIOGRAPHY 185

Black, S. E. and P. J. Devereux (2011). Recent Developments in Intergenerational Mobility. In O. Ashenfelter and D. Card (Eds.), Handbook of Labor Economics, Vol. 4B, pp. 1487–1541. Amsterdam, NL: Elsevier.

Blau, F. D. (1992). The Fertility of Immigrant Women: Evidence from High-Fertility Source Countries. In G. J. Borjas and R. B. Freeman (Eds.), Immigration and the Workforce: Economic Consequences for the United States and Source Areas, pp. 93–134. Chicago, IL: University of Chicago Press.

Blau, F. D. and L. M. Kahn (2011). Substitution Between Individual and Cultural Capital: Pre-Migration Labor Supply, Culture and US Labor Market Outcomes Among Immigrant Women. NBER Working Paper No. 17275.

Blau, F. D., L. M. Kahn, and K. L. Papps (2011). Gender, Source Country Character- istics, and Labor Market Assimilation among Immigrants. Review of Economics and Statistics 93 (1), 43–58.

Bleakley, H. and A. Chin (2004). Language skills and earnings: Evidence from childhood immigrants. Review of Economics and Statistics 86 (2), 481–496.

Bleakley, H. and A. Chin (2010). Age at arrival, English proﬁciency, and social assimilation among US immigrants. American Economic Journal: Applied Economics 2 (1), 165–192.

Borjas, G. J. (1987). Self-Selection and the Earnings of Immigrants. American Economic Review 77 (4), 531–553.

Borjas, G. J. (1992). Ethnic Capital and Intergenerational Mobility. Quarterly Journal of Economics 107 (1), 123–150.

Borjas, G. J. (1995). Ethnicity, Neighborhoods, and Human-Capital Externalities. Ameri- can Economic Review 85 (3), 365–390.

Borjas, G. J. (2000). Ethnic enclaves and assimilation. Swedish Economic Policy Re- view 7 (1), 89–122.

Brenton, P., F. Di Mauro, and M. Lücke (1999). Economic Integration and FDI: An Empirical Analysis of Foreign Investment in the EU and in Central and Eastern Europe. Empirica 26 (2), 95–121.

Brown, C. H., E. W. Holman, S. Wichmann, and V. Velupillai (2008). Automated classiﬁcation of the World’s languages: A description of the method and preliminary results. STUF-Language Typology and Universals 61 (4), 285–308. BIBLIOGRAPHY 186

Cameron, A. C., J. B. Gelbach, and D. L. Miller (2008). Bootstrap-Based Improvements for Inference with Clustered Errors. Review of Economics and Statistics 90 (3), 414–427.

Cameron, A. C., J. B. Gelbach, and D. L. Miller (2011). Robust Inference With Multiway Clustering. Journal of Business & Economic Statistics 29 (2), 238–249.

Carliner, G. (1981). Wage diﬀerences by language group and the market for language skills in Canada. Journal of Human Resources 16 (3), 384–399.

Casey, T. and C. Dustmann (2010). Immigrants’ Identity, Economic Outcomes and the Transmission of Identity across Generations. Economic Journal 120 (542), F31–F51.

Cavalli-Sforza, L. L., P. Menozzi, and A. Piazza (1994). The history and geography of human genes. Princeton, NJ: Princeton University Press.

Chan, K., V. Covrig, and L. K. Ng (2005). What Determines the Domestic Bias and Foreign Bias? Evidence from Mutual Fund Equity Allocations Worldwide. Journal of Finance 60 (3), 1495–1534.

Charette, M. and R. Meng (1994). Explaining language proﬁciency: Objective versus self-assessed measures of literacy. Economics Letters 44 (3), 313–321.

Chen, J., R. R. Sokal, and M. Ruhlen (1995). Worldwide analysis of genetic and linguistic relationships of human populations. Human Biology 67 (4), 595–612.

Chen, M. K. (2013). The eﬀect of language on economic behavior: Evidence from savings rates, health behaviors, and retirement assets. American Economic Review 103 (2), 690–731.

Chiswick, B. R. (1978). The Eﬀect of Americanization on the Earnings of Foreign-born Men. Journal of Political Economy 86 (5), 897–921.

Chiswick, B. R. (1991). Speaking, reading, and earnings among low-skilled immigrants. Journal of Labor Economics 9 (2), 149–70.

Chiswick, B. R. (1999). Are Immigrants Favorably Self-Selected? American Economic Review 89 (2), 181–185.

Chiswick, B. R. and P. W. Miller (1995). The endogeneity between language and earnings: International analyses. Journal of Labor Economics 13 (2), 246–288.

Chiswick, B. R. and P. W. Miller (1999). English language ﬂuency among immigrants in the United States. In S. W. Polachek (Ed.), Research in Labor Economics, Volume 17, pp. 151–200. Oxford, UK: JAI Press. BIBLIOGRAPHY 187

Chiswick, B. R. and P. W. Miller (2001). A model of destination-language acquisition: Application to male immigrants in Canada. Demography 38 (3), 391–409.

Chiswick, B. R. and P. W. Miller (2002). Immigrant earnings: Language skills, linguistic concentrations and the business cycle. Journal of Population Economics 15 (1), 31–57.

Chiswick, B. R. and P. W. Miller (2005). Linguistic Distance: A Quantitative Measure of the Distance Between English and Other Languages. Journal of Multilingual and Multicultural Development 26 (1), 1–11.

Chiswick, B. R. and P. W. Miller (2007). Modeling immigrants’ language skills. IZA Discussion Paper No. 2974.

Chiswick, B. R. and P. W. Miller (2010). Occupational language requirements and the value of English in the US labor market. Journal of Population Economics 23 (1), 353–372.

Coeurdacier, N. and H. Rey (2013). Home Bias in Open Economy Financial Macroeco- nomics. Journal of Economic Literature 51 (1), 63–115.

Crystal, D. (2010). The Cambridge Encyclopedia of Language (3rd ed.). Cambridge, UK: Cambridge University Press.

Cutler, D. M., E. L. Glaeser, and J. L. Vigdor (2008). When are ghettos bad? Lessons from immigrant segregation in the United States. Journal of Urban Economics 63 (3), 759–774.

Danzer, A. M. and F. Yaman (2010). Ethnic concentration and language ﬂuency of immigrants in Germany. IZA Discussion Paper No. 4742.

Daude, C. and M. Fratzscher (2008). The pecking order of cross-border investment. Journal of International Economics 74 (1), 94–119. de Coulon, A. and F.-C. Wolff (2007). Language proficiency of immigrants and misclassification. Applied Economics Letters 14 (12), 857–861.

Desmet, K., I. Ortuño-Ortín, and S. Weber (2009). Linguistic diversity and redistribution. Journal of the European Economic Association 7 (6), 1291–1318.

Dörnyei, Z. and R. Schmidt (2001). Motivation and second language acquisition, Volume 23 of Technical Report/Second Language Teaching & Curriculum Center. Honolulu, HI: University of Hawaii. BIBLIOGRAPHY 188

Dryer, M. S. and M. Haspelmath (2011). The World Atlas of Language Structures Online. Munich, DE: Max Planck Digital Library.

Dustmann, C. (1999). Temporary migration, human capital, and language ﬂuency of migrants. Scandinavian Journal of Economics 101 (2), 297–314.

Dustmann, C. and F. Fabbri (2003). Language proﬁciency and labour market performance of immigrants in the UK. Economic Journal 113 (489), 695–717.

Dustmann, C. and A. van Soest (2001). Language ﬂuency and earnings: Estimation with misclassiﬁed language indicators. Review of Economics and Statistics 83 (4), 663–674.

Dustmann, C. and A. van Soest (2002). Language and the earnings of immigrants. Industrial and Labor Relations Review 55 (3), 473–492.

Eaton, J. and S. Kortum (2002). Technology, Geography, and Trade. Econometrica 70 (5), 1741–1779.

Egger, P. H. and A. Lassmann (2012). The language eﬀect in international trade: A meta-analysis. Economics Letters 116 (2), 221–224.

Epstein, G. S. and I. N. Gang (2010). Migration and culture. In G. S. Epstein and I. N. Gang (Eds.), Frontiers of Economics and Globalization: Migration and Culture, Volume 8, pp. 1–21. Bingley, UK: Emerald Group Publishing Limited.

Esser, H. (2006). Migration, language and integration. AKI Research Review 4, Social Science Research Center Berlin, Berlin, DE.

European Commission (2003). Treaty Concerning the Accession of the Czech Republic, the Republic of Estonia, the Republic of Cyprus, the Republic of Latvia, the Republic of Lithuania, the Republic of Hungary, the Republic of Malta, the Republic of Poland, the Republic of Slovenia and the Slovak Republic to the European Union. Oﬃcial Journal of the European Union L 236, Luxembourg, LU.

European Commission (2005). Treaty Concerning the Accession of the Republic of Bulgaria and Romania to the European Union. Oﬃcial Journal of the European Union L 157, Luxembourg, LU.

European Commission (2010a). Consolidated versions of the Treaty on the Functioning of the European Union. Oﬃcial Journal of the European Union C 83, Luxembourg, LU.

European Commission (2010b). EUROPE 2020: A strategy for smart, sustainable and inclusive growth. European Commission. http://ec.europa.eu/europe2020/index_en.htm. BIBLIOGRAPHY 189

European Commission (2011). Migrants in Europe: A statistical portrait of the ﬁrst and second generation. Luxembourg, LU: Publications Oﬃce of the European Union.

Evans, M. D. R. (1986). Sources of immigrants’ language proﬁciency: Australian results with comparisons to the Federal Republic of Germany and the United States of America. European Sociological Review 2 (3), 226–236.

Fearon, J. D. and D. D. Laitin (2003). Ethnicity, Insurgency, and Civil War. American Political Science Review 97 (1), 75–90.

Felbermayr, G. J. and F. Toubal (2010). Cultural proximity and trade. European Economic Review 54 (2), 279–293.

Fernández, C. and C. Ortega (2008). Labor market assimilation of immigrants in Spain: employment at the expense of bad job-matches? Spanish Economic Review 10 (2), 83–107.

Fernández, R. (2007). Alfred Marshall Lecture Women, Work, and Culture. Journal of the European Economic Association 5 (2-3), 305–332.

Fernández, R. (2011). Does Culture Matter? In J. Benhabib, M. O. Jackson, and A. Bisin (Eds.), Handbook of Social Economics, Vol. 1A, pp. 481–510. Amsterdam, NL: Elsevier.

Fernández, R. (2013). Cultural Change as Learning: The Evolution of Female Labor Force Participation over a Century. American Economic Review 103 (1), 472–500.

Fernández, R. and A. Fogli (2009). Culture: An Empirical Investigation of Beliefs, Work, and Fertility. American Economic Journal: Macroeconomics 1 (1), 146–177.

Fernández, R., A. Fogli, and C. Olivetti (2004). Mothers and Sons: Preference Formation and Female Labor Force Dynamics. Quarterly Journal of Economics 119 (4), 1249–1299.

Fogli, A. and L. Veldkamp (2011). Nature or Nurture? Learning and the Geography of Female Labor Force Participation. Econometrica 79 (4), 1103–1138.

Freeman, R. B. (2006). People Flows in Globalization. Journal of Economic Perspec- tives 20 (2), 145–170.

French, K. R. and J. M. Poterba (1991). Investor Diversiﬁcation and International Equity Markets. American Economic Review 81 (2), 222–226.

Friedberg, R. M. (2000). You can’t take it with you? Immigrant assimilation and the portability of human capital. Journal of Labor Economics 18 (2), 221–51. BIBLIOGRAPHY 190

Friedman, M. (1962). Capitalism and Freedom. Chicago, IL: University of Chicago Press.

Galbraith, J. K. (1979). The Nature of Mass Poverty. Cambridge, MA: Harvard University Press.

Giuliano, P., A. Spilimbergo, and G. Tonon (2014). Genetic distance, transportation costs, and trade. Journal of Economic Geography 14 (1), 179–198.

Grinblatt, M. and M. Keloharju (2001). How Distance, Language, and Culture Inﬂuence Stockholdings and Trades. Journal of Finance 56 (3), 1053–1073.

Gronau, R. (1977). Leisure, Home Production, and Work – the Theory of the Allocation of Time Revisited. Journal of Political Economy 85 (6), 1099–1123.

Guiso, L., P. Sapienza, and L. Zingales (2006). Does Culture Aﬀect Economic Outcomes? Journal of Economic Perspectives 20 (2), 23–48.

Guiso, L., P. Sapienza, and L. Zingales (2009). Cultural biases in economic exchange? Quarterly Journal of Economics 124 (3), 1095–1131.

Hagen, S., J. Foreman-Peck, and S. Davila-Philippon (2006). ELAN: Eﬀects on the European economy of shortages of foreign language skills in enterprise. Brussels, BE: European Commission.

Haisken-DeNew, J. P. and J. R. Frick (2005). Desktop Companion to the Ger- man Socio-Economic Panel (SOEP): Version 8.0. http://www.diw.de/documents/ dokumentenarchiv/17/diw_01.c.38951.de/dtc.409713.pdf.

Haisken-DeNew, J. P. and M. Hahn (2006). Panelwhiz: A ﬂexible modularized stata interface for accessing large scale panel data sets. http://www.panelwhiz.eu/docs/ PanelWhiz_Introduction.pdf.

Hart-Gonzalez, L. and S. Lindemann (1993). Expected achievement in speaking proﬁciency. Foreign Service Institute, US Department of State, Washington, DC.

Head, K. and T. Mayer (2013). Gravity Equations: Workhorse, Toolkit, and Cookbook. In G. Gopinath, E. Helpman, and K. Rogoﬀ (Eds.), Handbook of International Economics, Vol. 4, pp. forthcoming. Amsterdam, NL: Elsevier.

Head, K., T. Mayer, and J. Ries (2010). The erosion of colonial trade linkages after independence. Journal of International Economics 81 (1), 1–14.

Helpman, E., M. Melitz, and Y. Rubinstein (2008). Estimating Trade Flows: Trading Partners and Trading Volumes. Quarterly Journal of Economics 123 (2), 441–487. BIBLIOGRAPHY 191

Holman, E. W., C. H. Brown, S. Wichmann, A. Müller, V. Velupillai, H. Hammarström, S. Sauppe, H. Jung, D. Bakker, P. Brown, O. Belyaev, M. Urban, R. Mailhammer, J.-M. List, and D. Egorov (2011). Automated dating of the world’s language families based on lexical similarity. Current Anthropology 52 (6), 841–875.

Hornok, C. (2012). Gravity or Dummies? The Limits of Identiﬁcation in Gravity Estima- tions. Department of Economics Working Papers.

Hutchinson, W. K. (2005). “Linguistic Distance” as a Determinant of Bilateral Trade. Southern Economic Journal 72 (1), 1–15.

IFPRI (2013). 2012 Global Food Policy Report. Washington, DC: International Food Policy Research Institute.

IOM (2013). World Migration Report 2013 – Migrant Well-being and Development. Geneva, CH: International Organization for Migration.

Isphording, I. E. and S. Otten (2011). Linguistic Distance and the Language Fluency of Immigrants. Ruhr Economic Papers #274.

Isphording, I. E. and S. Otten (2013). The Costs of Babylon – Linguistic Distance in Applied Economics. Review of International Economics, 21 (2), 354–369.

Isphording, I. E. and M. Sinning (2012). The returns to language skills in the US labor market. IZA Discussion Paper No. 7080.

Jaeger, D. A. (1997). Reconciling the old and new Census Bureau education questions: Recommendations for researchers. Journal of Business & Economic Statistics 15 (3), 300–309.

Jann, B. (2007). Making regression tables simpliﬁed. Stata Journal 7 (2), 227–244.

Johnston, D. W., S. Schurer, and M. A. Shields (2012). Maternal Gender Role Attitudes, Human Capital Investment, and Labour Supply of Sons and Daughters. IZA Discussion Paper No. 6656.

Kok, S., N. Bosch, A. Deelen, and R. Euwals (2011). Migrant Women on the Labour Market: On the Role of Home-and Host-Country Participation. IZA Discussion Paper No. 5817.

Kossoudji, S. A. (1988). English language ability and the labor market opportunities of Hispanic and East Asian immigrant men. Journal of Labor Economics 6 (2), 205–228. BIBLIOGRAPHY 192

Ku, H. and A. Zussman (2010). Lingua franca: The role of English in international trade. Journal of Economic Behavior & Organization 75 (2), 250–260.

Lane, P. R. and G. M. Milesi-Ferretti (2003). International ﬁnancial integration. IMF Staﬀ Papers 50 (S1), 82–113.

Lane, P. R. and G. M. Milesi-Ferretti (2008). International Investment Patterns. Review of Economics and Statistics 90 (3), 538–549.

Lewis, K. K. (1999). Trying to Explain Home Bias in Equities and Consumption. Journal of Economic Literature 37 (2), 571–608.

Lewis, P. M. (2009). Ethnologue: Languages of the World. Dallas, TX: SIL International.

Llach, A. M. P. (2010). An overview of variables aﬀecting lexical transfer in writing: A review study. International Journal of Linguistics 2 (1), E2.

Lohmann, J. (2011). Do language barriers aﬀect trade? Economics Letters 110 (2), 159–162.

Manser, M. and M. Brown (1980). Marriage and Household Decision-Making: A Bargaining Analysis. International Economic Review 21 (1), 31–44.

Martin, P. and H. Rey (2004). Financial super-markets: size matters for asset trade. Journal of International Economics 64 (2), 335–361.

Matyas, L., C. Hornok, and D. Pus (2013). The Formulation and Estimation of Random Eﬀects Panel Data Models of Trade. MPRA Paper No. 36789.

Mayer, T. and S. Zignago (2011). Notes on CEPII’s distances measures (GeoDist). CEPII Working Paper 2011-25.

McCallum, J. (1995). National borders matter: Canada-U.S. regional trade patterns. American Economic Review 85 (3), 615–623.

McElroy, M. B. and M. J. Horney (1981). Nash-Bargained Household Decisions: Toward a Generalization of the Theory of Demand. International Economic Review 22 (2), 333–349.

McManus, W., W. Gould, and F. Welch (1983). Earnings of Hispanic men: The role of English language proﬁciency. Journal of Labor Economics 1 (2), 101–130.

Mélitz, J. (2008). Language and foreign trade. European Economic Review 52 (4), 667–699. BIBLIOGRAPHY 193

Mélitz, J. and F. Toubal (2012). Native language, spoken language, translation and trade. CEPR Discussion Paper No. 8994.

Michler, I. (2011). Warum große Firmenfusionen immer wieder scheitern. Die Welt. http://www.welt.de/12878747.

Moulton, B. R. (1986). Random group eﬀects and the precision of regression estimates. Journal of Econometrics 32 (3), 385–397.

Moulton, B. R. (1990). An Illustration of a Pitfall in Estimating the Eﬀects of Aggregate Variables on Micro Units. Review of Economics and Statistics 72 (2), 334–338.

Münz, R. and R. Ohliger (1998). Long-distance citizens: Ethnic Germans and their immigration to Germany. In P. H. Schuck and R. Münz (Eds.), Paths to Inclusion: The Integration of Migrants in the United States and Germany, pp. 155–202. New York, NY: Berghahn Books.

Newport, E. L. (2002). Critical periods in language development. In L. Nadel (Ed.), Encyclopedia of Cognitive Science, pp. 737–740. London, UK: Macmillan Publishers Ltd./Nature Publishing Group.

OECD (2011). Trade for Growth and Poverty Reduction: How Aid for Trade Can Help. Paris, FR: OECD Publishing.

Orrenius, P. M. and M. Zavodny (2005). Self-selection among undocumented immigrants from Mexico. Journal of Development Economics 78 (1), 215–240.

Pencavel, J. (1998). Assortative Mating by Schooling and the Work Behavior of Wives and Husbands. American Economic Review: Papers and Proceedings 88 (2), 326–329.

Petroni, F. and M. Serva (2010). Measures of lexical distance between languages. Physica A: Statistical Mechanics and its Applications 389 (11), 2280–2283.

Portes, R. and H. Rey (2005). The determinants of cross-border equity ﬂows. Journal of International Economics 65 (2), 269–296.

Rauch, J. E. and V. Trindade (2002). Ethnic Chinese Networks in International Trade. Review of Economics and Statistics 84 (1), 116–130.

Reher, D. and M. Requena (2009). The national immigrant survey of Spain: A new data source for migration studies in Europe. Demographic Research 20 (12), 253–278.

Reimers, C. W. (1985). Cultural Diﬀerences in Labor Force Participation Among Married Women. American Economic Review 75 (2), 251–255. BIBLIOGRAPHY 194

Robinson, C. (1988). Language choice: The distribution of language skills and earnings in a dual language economy. In R. G. Ehrenberg (Ed.), Research in Labor Economics, Volume 9, pp. 53–90. Oxford, UK: JAI Press.

Rose, A. K. (2004). Do we really know that the WTO increases trade? American Economic Review 94 (1), 98–114.

Rose, A. K. and M. M. Spiegel (2009). Noneconomic Engagement and International Exchange: The Case of Environmental Treaties. Journal of Money, Credit and Bank- ing 41 (2-3), 337–363.

Rose, A. K. and M. M. Spiegel (2011). The Olympic Eﬀect. Economic Journal 121 (553), 652–677.

Roy, A. D. (1951). Some Thoughts on the Distribution of Earnings. Oxford Economic Papers 3 (2), 135–146.

Santacreu-Vasut, E., A. Shoham, and V. Gay (2013). Do female/male distinctions in language matter? Evidence from gender political quotas. Applied Economics Letters 20 (5), 495–498.

Santos Silva, J. M. C. and S. Tenreyro (2006). The Log of Gravity. Review of Economics and Statistics 88 (4), 641–658.

Serva, M. (2011). Phylogeny and geometry of languages from normalized levenshtein distance. Ithaca, NY: Cornell University Library.

Spolaore, E. and R. Wacziarg (2009). The Diﬀusion of Development. Quarterly Journal of Economics 124 (2), 469–529.

Spolaore, E. and R. Wacziarg (2012). Long-Term Barriers to the International Diﬀusion of Innovations. In J. Frankel and C. Pissarides (Eds.), NBER International Seminar on Macroeconomics 2011, pp. 11–46. Chicago, IL: University of Chicago Press.

Stock, J. H. and M. Yogo (2005). Testing for weak instruments in linear IV regression. In D. W. K. Andrews and J. H. Stock (Eds.), Identiﬁcation and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg. Cambridge, UK: Cambridge University Press.

Swadesh, M. (1952). Lexico-statistic dating of prehistoric ethnic contacts: With special reference to North American Indians and Eskimos. Proceedings of the American Philosophical Society 96 (4), 452–463. BIBLIOGRAPHY 195

Tainer, E. (1988). English language proﬁciency and the determination of earnings among foreign-born men. Journal of Human Resources 23 (1), 108–122.

Tesar, L. L. and I. M. Werner (1995). Home bias and high turnover. Journal of International Money and Finance 14 (4), 467–492.

Tinbergen, J. (1962). Shaping the World Economy: Suggestions for an International Economic Policy. New York, NY: The Twentieth Century Fund.

Tomz, M., J. L. Goldstein, and D. Rivers (2007). Do we really know that the WTO increases trade? Comment. American Economic Review 97 (5), 2005–2018.

United Nations (2000). United Nations Millennium Declaration. New York, NY: United Nations.

United Nations (2012). Trends in international migrant stock: Migrants by destination and origin. New York, NY: United Nations database, Department of Economic and Social Aﬀairs.

United Nations (2013). The Millennium Development Goals Report 2013. New York, NY: United Nations.

U.S. Department of Homeland Security (2010). Yearbook of immigration statistics: 2009. Washington, DC: U.S. Department of Homeland Security, Oﬃce of Immigration Statistics.

Van der Slik, F. W. P. (2010). Acquisition of Dutch as a second language. Studies in Second Language Acquisition 32 (3), 401–432.

Wacziarg, R. and K. H. Welch (2007). Trade Liberalization and Growth: New Evidence. World Bank Economic Review 22 (2), 187–231.

Winters, L. A., N. McCulloch, and A. McKay (2004). Trade Liberalization and Poverty: The Evidence So Far. Journal of Economic Literature 42 (1), 72–115.

World Bank (2011). Migration and Remittances Factbook 2011. Washington, DC: World Bank Publications.

Zimmermann, K. F. (1995). Tackling the European migration problem. Journal of Economic Perspectives 9 (2), 45–62. Acknowledgments

I want to express my deep gratitude to the many people who have supported me during my PhD years and thus contributed substantially to the completion of my dissertation. First of all, I would like to thank my supervisors Thomas K. Bauer and Christoph M. Schmidt for their excellent supervision, their valuable comments on my research projects, and their enduring patience during the last years. But more than that, I am grateful for their continuous advice and guidance throughout my first steps into academia. Furthermore, I am highly indebted to Thomas Bauer for supporting me from the very beginning and providing me the academic freedom I needed to find and perform my own research, which made my time at his chair an exciting and enjoyable experience. I would further like to thank my colleagues at the Ruhr University Bochum and the RWI Essen for creating an excellent working atmosphere and a fruitful research environment. In particular, I am grateful to Ronald Bachmann, Katja Görlitz, and Matthias Vorell for their continuing advice and complementary guidance over the years and to Sonja Kassenböhmer, Magdalena Stroka, Anna Talmann, and Lina Zwick for being always cooperative and helpful colleagues. I am also very thankful to Ira Gang, Michael Kvasnicka, Alfredo Paloyo, and Joel Stiebale for their valuable support and suggestions. My thesis has further benefited from many comments and suggestions from participants at various conferences and seminars. Furthermore, I want to thank those persons who supported me through their research and organizational assistance within the last six years. Explicitly, I would like to express my gratitude to Martin Burkert, Carsten Crede, Claudia Lohkamp, and Christian Rulff. Very special thanks go to my co-author Mathias Sinning, who constantly supported my research, and John Haisken-DeNew, who also mentored me throughout my doctorate. My deepest gratitude, however, I want to express to my co-authors, colleagues, and close friends Julia Bredtmann and Ingo Isphording. Working together with Julia and Ingo was always a pleasure and the smartest and most important decision I have made during my PhD. Their support was invaluable for me and the progress of my research. I hope that our friendship will accompany me in the future. Ingo’s competence in econometrics, his deep interest in research, and his overall curiosity enriched our joint projects immensely. Julia and I started with our PhDs and working at the chair at the same time. During the last six years, Julia was not only the best colleague I could have imagined, but also the person I benefited from the most. Her intelligence, talent, and work attitude was a continuous motivation for me. Finally, I am deeply grateful to my parents and my sister, who constantly supported me through my undergraduate and graduate studies. Furthermore, I owe thanks to my friends for their encouraging support and for always helping me to keep the balance. I dedicate this dissertation to the memory of my grandmother, who always kept the faith in me and motivated me to start my PhD. Curriculum Vitae

Personal Information

Date of Birth 21.11.1980 Place of Birth Rhede, Germany Citizenship German

Education

11/2007–02/2014 Ph.D. in Economics, Ruhr University Bochum Supervisors: Prof. Dr. Thomas K. Bauer, Prof. Dr. Christoph M. Schmidt

08/2006–12/2006 Studies in Economics, Saint Mary’s University Halifax, Canada 04/2002–09/2007 Diploma in Economics, Ruhr University Bochum

Professional Experience

11/2007–Present Research and Teaching Assistant, Ruhr University Bochum, Chair for Empirical Economics (Prof. Dr. Thomas K. Bauer)

11/2007–Present Research Aﬃliate, RWI Essen (Rhine-Westphalia Institute for Economic Research), Research Division “Labor Markets, Educa- tion, Population”