The Pennsylvania State University. The Graduate School. College of Agricultural Sciences

Transcription

1 The Pennsylvania State University The Graduate School College of Agricultural Sciences THREE ESSAYS ON ECONOMIC DEVELOPMENT: INTERNATIONAL MIGRATION, SOCIAL NETWORKS, AND SOCIAL CAPITAL A Dissertation in Agricultural, Environmental, and Regional Economics by Vania Bitia Salas Garcia 2013 Vania Bitia Salas Garcia Submitted in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy December 2013

2 The dissertation of Vania Bitia Salas Garcia was reviewed and approved* by the following Stephan J. Goetz Professor of Agricultural Economics, Sociology and Education, and Demography Director of Northeast Regional Center for Rural Development Dissertation Advisor Chair of Committee David G. Abler Professor of Agricultural Economics, Sociology and Education, and Demography Edward Jaenicke Professor of Agricultural Economics, Sociology and Education, and Demography Kai Schafft Professor of Education in the College of Education Ann R. Tickamyer Head of the Department of Agricultural Economics, Sociology and Education *Signatures are on file in the Graduate School. ii

3 ABSTRACT Three essays are included in this dissertation; each approaches development strategies with different research methods and data. The focus of this dissertation is on analyzing the impacts of international remittances and social networks as development strategies pursued by households and small rural firms, respectively. This dissertation extends the study of social networks to determine the effect of population movement variables such as migration and commuting on social capital. Thus, the link among these three chapters is the analysis of development strategies with an emphasis on the role of these strategies in enhancing economic development. The first essay investigates the effect of international migration on children left behind in Peru. The research question posed is: what is the role of international remittances in human capital formation of children left behind, separating the effect of absenteeism of parents (due to migration, divorce or other) from the effect of remittances? The theoretical model is based on human capital theory and educational investment decisions linked to remittances. The model proposes to analyze the role of international remittances on the investment decision between sending children to a public or a private school, that is, the decision to acquire a higher quality of education. Using different econometric strategies to address endogeneity, I find that remittances have a positive effect on the likelihood of sending children to private schools, controlling for absenteeism of parents. The second essay uses unique survey data to identify the role of social networks on small rural firms performance located in the Northeast United States (Maryland, New York, and Pennsylvania); specifically, two networks are evaluated: social support and economic. There are two specific objectives: 1) to evaluate the effect of direct and indirect network ties within a iii

4 cluster on the economic viability of small rural business by using out-degree and closeness, respectively, and 2) to evaluate the impact of a network s topology on small rural firms by using density. After addressing sample selection and endogeneity biases, the estimation of the business success models shows that creating direct and indirect ties in the economic and the social support networks will improve economic performance as measured by the reported percent change in products sold and in net farm income. In addition, the results indicate that strong networks (i.e., a high density degree) will be detrimental to economic performance of small rural firms. The third essay follows the line of research of the second essay on social networks to extend the literature on social capital. This research includes migration, commuting (flow and time) and centrality network indicators as population movement indicators to examine the effects of those variables on the production of social capital, and its change over time. I exploit the directionality of these variables by including in- and out- per capita flows for migration and commuting. The results show that the effects of population movement variables on social capital vary according to the direction of the flow. To identify these effects, I use data from U.S. counties and employ three different econometric techniques: linear regression, spatial econometric, and panel data. I find that the migration variables (in- and out-) and the centrality network indicator closeness for migration and commuting have a significant effect whereas the commuting variables (in- and out-) do not have a clear effect. Out-migration positively affects the creation of social capital but in-migration has a negative effect. Finally, the closeness indicator for migration and commuting has a negative effect on social capital. iv

5 TABLE OF CONTENTS List of Tables..ix List of Figures....xii Acknowledgments.... xiii Chapter 1. INTRODUCTION Dissertation Outline References.2 Chapter 2. FIRST ESSAY: International remittances and human capital formation Introduction Literature review Peruvian background The model Data Variable definitions Estimation strategy and hypothesis Results Conclusions References.28 v

6 Appendix A.35 Chapter 3. SECOND ESSAY: Social networks and small rural firm performance Introduction Literature and framework Data and variable definitions Dependent variables: Economic performance Independent variables Social network indicators Network density and cluster organization Hypotheses Research methods Results Descriptive results Dependent variable: Percentage of volume of products sold Dependent variable: Percentage of net farm income Economic results Results for percentage change of products sold Results for percentage change of net farm income Conclusions References...67 Appendix B vi

7 Chapter 4. THIRD ESSAY: Population movement across counties and its effect on social capital Introduction Literature review Social capital Social network theory The social capital model Data and variable definitions Social capital variable Main independent variables Population movement variables Closeness centrality indicator Hypotheses Estimation strategy First model: OLS for social capital and its components Second model: The spatial regression model Third model: Panel data Results Fist model Second model Third model Conclusions vii

8 4.9. References.128 Appendix C viii

9 LIST OF TABLES Table 2.1. Descriptive statistics.31 Table 2.2. Educational investment equation..32 Table 2.3. Educational investment equation by gender of head of household..33 Table 2.A. Correlation matrix 35 Table 2.B. Educational investment and remittances equations using agencies as instrumental variable Table 2.C. Remittances equation...37 Table 2.D. Marginal effects for the educational investment equation...38 Table 2.E. Remittances equation by gender of the head of household..39 Table 2.F. Educational investment and remittances equations controlling for Lima 40 Table 3.1. Descriptive statistics. 71 Table 3.2. Impact of individual and group variables on % change of products sold.72 Table 3.3. Impact of out-degree and density on % change of products sold.73 Table 3.4. Impact of closeness and density on % change of products sold...74 Table 3.5. Impact of individual and group variables on % change of net farm income...75 Table 3.6. Impact of out-degree and density on % change of net farm income...76 Table 3.7. Impact of closeness and density on % change of net farm income.. 77 Table 3.A. Correlation matrix 84 Table 3.B. Selection equation for % change of products sold...85 Table 3.C. Selection equation for % change of net farm income..86 Table 3.D. First stage TOBIT for endogenous variable out-degree (for products sold).87 ix

10 Table 3.E. First stage TOBIT for endogenous variable closeness (for products sold) Table 3.F. First stage TOBIT for endogenous variable out-degree (for net income).89 Table 3.G. First stage TOBIT for endogenous variable closeness (for net income)..90 Table 4.1. Descriptive statistics Table 4.2. Migration per capita in 1990 and Table 4.3. Commuting per capita in 1990 and Table 4.4. Closeness centrality measure for commuting flows in 1990 and Table 4.5. Closeness centrality measure for migration flows in 1990 and Table 4.6. OLS results for social capital, Table 4.7. OLS results for social capital, Table 4.8. OLS results for components of social capital, Table 4.9. OLS results for components of social capital, Table Spatial econometric estimation for social capital, Table Spatial econometric estimation for social capital, Table Panel data estimation results (Random effects) 144 Table 4.A. OLS results for social capital 2009 including independent variables in 1990 and Table 4.B. Correlation matrix among betweenness and closeness centrality measures for migration and commuting in 1990 and Table 4.C. OLS results for social capital, Table 4.D. VIF analysis 1997 and Table 4.E. Complementary OLS regressions for social capital, 1997 and Table 4.F. Standardized (beta) coefficients for social capital estimations, 1997 and x

11 Table 4.G. Panel data with fixed effects..152 xi

12 LIST OF FIGURES Figure 2.1. Remittances and permanent emigration, Figure 3.1. Network ties within clusters...78 Figure 3.2. Network ties within clusters by percentage of volume of products sold.80 Figure 3.3. Network ties within clusters by percentage of net farm income. 82 Figure 4.1. Social capital index, xii

13 ACKNOWLEDGMENTS I would like to express my deep gratitude to all people who provided me with advice and support during the work on this dissertation. First of all, I would like to thank my advisor Dr. Stephan Goetz for his patience, support and guidance throughout the process of my doctoral education. I also thank Dr. Jill Findeis for her support during my first years as a Master student. I also gratefully acknowledge all the useful insights and suggestions from the members of my committee: Dr. David Abler, Dr. Edward Jaenicke, and Dr. Kai Schafft. I would like to thank Dr. Kathryn Brasier and The Small Farms Industry Clusters Project for providing me with the data for my second essay. Finally, this dissertation would not have been possible without the support of my friends and my family. Special thanks to my parents and sister in Peru. xiii

14 Chapter 1 INTRODUCTION Economic development remains as an important concern not only for developing countries such as Peru but also for developed countries such as the U.S. where rural areas are not showing improvements in economic performance 1. This dissertation addresses economic development in three different essays that include three potential development strategies: international remittances, social networks, and social capital. These strategies are undertaken with the expectation of producing positive changes in a particular population, region or country. Using data from Peru, small rural firms in the Northeast United States, and U.S. counties, my dissertation explores the effects of these strategies on three different development outcomes. The first essay focuses on the effect of international remittances on improving human capital of children left behind in Peru. Using Peruvian data for the period , this study estimates a two-step human capital equation to address endogeneity of remittances, and the findings shows that remittances have a positive effect on the likelihood of sending children to private schools controlling for absenteeism of parents. The second essay exploits a unique data from small rural firms located in the Northeast United States to evaluate the impact of creating ties within social support and economic networks on farm success. The econometric analysis is based on a two-stage model with instrumental variables to address selection and endogeneity biases. The analysis shows that social network variables such as out-degree, closeness and density are relevant to explain business success of 1 According to the U.S. Department of Agriculture, for the first half of 2012, employment in non-metro areas remains 3.2% below its peak in 2007 whereas metro areas show a higher employment growth with 2.1% below its 2007 peak. Similarly, the poverty rate in 2011 for non-metro areas was 17% whereas poverty in metro areas was 14.6%. 1

15 small rural firms. Creating direct and indirect ties has a positive and significant effect on the percentage change of products sold and net farm income, whereas network density negatively affects economic performance of small rural firms. The last chapter studies the effect of population movement variables and their network indicators on the formation of social capital at the county level in the U.S. Based on three different economic techniques, the results reveal that commuting (flow and time), migration and their network indicators play an important role in the creation of social capital. Out-commuting, out-migration and closeness indicators are significant under the three econometric models whereas in-commuting and in-migration per capita variables are sensitive to the econometric model employed References Economic Research Service (2012). Rural America at a glance. Economic Brief Number 21, USDA, 6p. 2

16 Chapter 2 FIRST ESSAY International remittances and human capital formation 2.1. Introduction One of the main problems facing policy-makers in Peru is improving the quality of public education in order to enhance human capital in the country, and thus leave the bottom rank in the Programme for International Student Assessment (PISA) evaluation compiled by the Organisation for Economic Co-operation and Development (OECD) for more than 60 countries. Although educational coverage has increased to almost the same levels as in developed countries (UNESCO-UIS/OECD, 2005), the quality of Peruvian education remains low. PISA (2003) results show that Peru has an average score of 327 points which is the lowest score in the Latin American region (Argentina, Chile, Mexico, Brazil, and Peru) and much lower than the average for the OECD countries (500 points) (UMC, 2004). As a consequence of PISA s results the Peruvian education has been declared to be in a state of emergency by the national government. The educational system in Peru reflects the high inequality experienced within Peruvian society 2 ; private schools provide better education compared to that provided by public schools. According to the results on reading literacy skills reported in PISA (2003), the difference in student performance within Peru for the year 2000 between the highest and the lowest quintile 3 is 314 points, while the average difference in the PISA study overall is 327 points (UMC, 2004) 4. This gap is explained by the differences among schools, and 2 Gini index in 2003 was (ECLAC). 3 The quintile is based on student s performance. 4 PISA (2003) constructs a combined reading literacy scale with an average score of 500 points as level 5 which is the highest level (UMC, 2004). 3

17 around 58% of student s performance in PISA is explained by school factors (UMC, 2004). This fact highlights that a better quality of education is provided by private schools compared to public schools. Attending school, especially private schools, has become a pathway to move up and improve socio-economic status (SES) which may have a larger impact on children from lowand middle-income households. Evidence from the U.S. suggests that children from private schools have a higher probability of accessing a better college education (Griffith & Rothstein, 2009), which in turn increases their job opportunities. Children from low-income households are doomed to attend public schools which in turn make them less competitive than children educated in private schools. The main constraint to access not only private but also public education in Peru is its cost and the opportunity cost of children s work. Sending children to school becomes prohibitive for families in the lower group of Peruvian s income distribution. International remittances might loosen this income constraint allowing families left behind by, firstly sending children to school, and secondly affording a private education for their children. International remittances have become an important source of income in developing countries, reaching $ billion in 2011 (World Bank, 2012). However, there is still not a clear effect of remittances on schooling (Borraz, 2005; Hu, 2012) and to my knowledge all the studies on remittances effects have evaluated only quantity without including quality of schooling. Additionally, there is almost no research on migration s effect on education for the Peruvian case (Frisancho & Oropesa, 2011). Only one paper (Calero et al., 2008) approaches the quality decision for Ecuador but it uses one-year information and fails to control for absenteeism of parents. Hence, the research question proposed here is: what is the role of receiving international remittances in human capital 4

18 formation of children left behind by separating the effect of absenteeism of parents (due to migration, divorce or other) from remittances? The Peruvian case is different from other cases such as Mexico or Central American countries. Unlike those cases, migration entails a high cost for Peruvian migrants due to the geographical distance to the United States, which is the main destination country (32.6% of Peruvians migrated to the U.S. over the period ); thus, relatively more affluent people are more likely to migrate abroad (Frisancho & Oropesa, 2011) and they are less likely to have liquidity constraints to afford schooling costs. Education in Peru is compulsory and provided at negligible cost by the government; parents only need to pay for a few additional expenses such as uniforms, transportation, and school supplies. Hence, relatively small and even negative effects of remittances on schooling have been found in previous studies for Peru 5. Yet international remittances in the Peruvian case should be explored beyond the quantity effect, i.e., years of schooling, and the effect of remittances on the quality of education should be included in the analysis. In sum, Peru is an interesting case due to several factors. First, inequality in the Peruvian school system is high. Second, the flow of non-returning emigrants and remittances is increasing. Third, there is scarce literature on remittances effects for Peru. And fourth, the data provide unique information for a panel study ( ) including the amount of international remittances received by each household member at the national level which is an important difference to previous studies (not only for Peru) which generally only used a categorical indicator for remittances 6. The contributions of my research are to: 1) fill the gap in the literature by examining the role of remittances on human capital beyond education quantity by including education quality; 2) consider a broader range of departments in Peru and not only the capital 5 See Acosta et al. (2007) and Frisancho & Oropesa (2011). 6 Only one study, Hu (2011) for internal migration in China includes amount of remittances but it does not correct for endogeneity. 5

19 department as in previous studies; 3) disentangle the effect of international remittances from the absenteeism of parents, unlike previous studies for the Peruvian case that focused only on one factor; and 4) the amount of remittances is included in the analysis instead of a dummy indicator of receiving remittances. The remainder of this chapter is organized as follows: section 2 presents a review of the literature, section 3 includes a brief background of Peru related to remittances, section 4 outlines the model, and the data and variable definition are presented in section 5 and 6, respectively. The estimation strategy and hypothesis are discussed in section 7 and, results are presented in section 8, and section 9 concludes Literature review Migration and remittances may have different effects on migrants and relatives left behind; their effects go beyond economic outcomes such as income or productivity: they include health, education, and subjective well-being. My research will focus on the effect of remittances on human capital investment of children left behind in the home country. Research on migration has found two opposite effects on human capital depending on the country and outcome studied. Migration changes both the marginal benefit and the marginal cost of education making unclear its effect on human capital (Berker, 2009). On the other hand, literature linking remittances and human capital has been addressed through a quantity approach such as measuring years of education, grade repetition or school attendance rather than focusing on education quality. This paper intends to fill that gap in the literature by including the differences between public and private schools as a proxy for education quality. Most migration studies have addressed migration s effect on human capital through three channels: (1) remittance s effect by relaxing income and borrowing constraints thereby 6

20 making the cost of school affordable, (2) absenteeism of parents as a lack of parental control or emotional support for children, and (3) immigrant population s density as a factor of changing demographic characteristics in the receiving and origin areas. However, there is insufficient research on migration s consequences for South American immigrants and their relatives left behind, especially for the Peruvian case (Frisancho & Oropesa, 2011). Studies on borrowing constraints such as Jacoby (1994) and Akee et al. (2010) have highlighted a positive relationship between income and school progress but without including remittances in their analyses. Hence, recent studies have focused on the first channel that proposes migration as a strategy to loosen family income and borrowing constraints by sending back remittances. According to the New Economics of Migration, remittances are sent by a household member living abroad as a repayment for the investment on migration which was undertaken as a family strategy for income diversification (Stark & Bloom, 1985) or for purely altruistic reasons (Lucas & Stark, 1985). Most findings in the literature have stated a positive relationship between remittances and human capital but there is no consensus. The positive effect of remittances on financial constraints changes the opportunity cost of acquiring more schooling; thus, families may find it optimal to send children to school instead of sending them to the labor market. Researchers have explored remittances effects on different indicators of human capital. Cox & Ureta (2003) employed dropping out of school as an indicator to explain the large and positive effect of remittances on human capital in comparison to other income sources for El Salvador. Amuedo-Dorantes & Pozo (2010) focused on children s school attendance from the Dominican Republic and found a positive effect among secondary school-age children and higher order of birth siblings. In a similar vein, Calero et al. (2008) found that international remittances have a positive effect on school enrollment in Ecuador. On the contrary, Acosta et al. (2007) found a negative effect of remittances on educational 7

21 attainment in the Dominican Republic, and Meza & Pederzini (2008) stated that remittances in rural Mexico have a negative effect on school achievement. In the second channel, migration of parents may produce a negative effect on children s educational outcome due to the lack of parental control and by producing a change in children s expectations. First, a household experiencing migration is similar to a disrupted family that has a negative psychological effect on children which in turn affects their educational performance (Bennett et al., 2012; Kandel & Kao, 2001). In addition, migration places rearing and housework responsibilities on children left behind, especially on the older one, affecting their allocation of time to school work. Children left behind must take the parents role as a provider by entering the labor force at earlier ages and becoming a parent figure for younger siblings (Booth & Tamura, 2009; McKenzie & Rapoport, 2010). Second, if children perceive that their immigrant parents (or relatives and friends) gain higher wages by working in unskilled jobs in the receiving country then children may have no or less incentive to pursue higher levels of education. According to Kandel & Kao (2001), migration is perceived as an alternative to achieve economic success without having higher levels of education. Children with migrant parents may increase their likelihood to migrate (Kandel & Kao, 2001; McKenzie & Rapoport, 2010), and thus they do not acquire more education because the marginal return to education from the origin country is lower in the receiving country. Evidence from Mexico found that children left in migrant households obtain less years of schooling in comparison to those children living in non-migrant households (McKenzie & Rapoport, 2010). In a similar vein, Frisancho & Oropesa (2011) found a negative impact on educational attainment for children in Peru living in households with a high risk of migration; however, they did not control for amount of remittances and only employed data for Lima (capital of Peru). 8

22 On the other hand, the brain gain hypothesis in the migration literature posits a positive relationship between labor migration and human capital formation. Evidence for Tajikistan shows that the long-term migration of parents increases the enrolment rate of children left behind, but this study did not control for remittances (Bennett et al., 2012). Theoretical studies such as Vidal (1998) pointed out that in a dynamic system there exists a threshold of human capital, above # h, such that sending countries with an initial human capital # h will invest more on education and will convert to a high level equilibrium. A country with a highly educated population will send emigrants who are more likely to earn a higher return to their education in the receiving country, which may have a positive influence on their relatives and friends remaining in the sending country to invest more in education. Likewise, Stark & Wang (2002) state that migration may have a positive effect on educational investments in developing countries as it is used as a substitute for education subsidies. The literature has found opposite effects from the first and second channels, and the net effect will depend on whether remittances or absenteeism of parents have the larger effect. Therefore, recent studies have included both channels to isolate each effect of migration. It is not clear a priori which effect will be larger; after controlling for the absenteeism of migrant family members, the positive effect of remittances may lose its statistical significance. Based on Dominican Republic data, Amuedo-Dorantes & Pozo (2010) concluded that migration of family members offsets the positive effect of remittances. On the other hand, Hu (2012) found that for the case of internal migration in China, the positive effect of remittances partially compensates for the negative effect of absenteeism of parents. And, Borraz (2005) found a positive but small effect of remittances on years of school for children living in remittances-receiving households in Mexico controlling for absent parents. 9

23 Research on the third channel focuses on demographic changes due to immigration. Two important changes are mostly studied, change in the school-age population and change in the labor force age (Berker, 2009). Generally, immigrant children need to attend school levels below their corresponding age, increasing the pressure on school s assets which in turn negatively affects native children (Gould, et al., 2009; Schwartz & Stiefel, 2006). On the other hand, the workforce age will change due to the departure of family members that forces children left behind to enter the job market earlier to meet monetary needs (McKenzie & Rapoport, 2010). Evidence for Turkey and Israel suggests a negative impact on middleschool and high-school completion rates for native students in the receiving area, especially for native children from low SES households (Berker, 2009; Gould, et al., 2009). However, U.S. evidence suggests that immigrant children take advantage of the school system in the receiving area. Results for elementary and middle schools in U.S. (New York) show immigrant children obtain better scores on math and reading than native-born children 7 (Schwartz & Stiefel, 2006) Peruvian background Peru shows a high out-migration flow of non-returning Peruvians emigrants, which has rapidly increased since 2000 although there was a slight reduction in 2009 as shown in Figure 2.1. The stock of Peruvians living abroad who did not return during the period is 7.0% over the total expected population for 2009, and 51.0% are from Lima, the capital of Peru (OIM, 2010). The main destination country of Peruvian migrants is the U.S.: over the period , 32.6% migrated to the U.S., 16.6% to Spain, 13.5% to Argentina, 10.0% to Italy, 7.8% to Chile, and 4.2% to Japan (OIM, 2010). 7 Native-born children refer to children who were born in U.S. but their parents are immigrants. 10

24 The high number of emigrants produces a higher volume of remittances, which rose from 700 million to over 2 billion dollars during the period (BCRP 8 ). Figure 2.1 shows that remittances have followed the same increasing trend of permanent emigration in Peru, especially since And, the importance of international remittances in the national economy has increased from 0.29% of GDP in 1990 to 1.89% (BCRP) in According to figures from ENAHO, for the period , around 81.1% of international remittances are sent to the Peruvian coast, and Lima is the main destination area receiving around 41.5%; 11.2% are sent to Peruvian highlands and 7.7% are sent to Peruvian jungle 9. Hence, the main impact of international remittances is on the Peruvian coast. However, official figures collected by BCRP are downward biased because 8.7% 10 of international remittances during were received through friends and relatives instead of financial institutions. For the same period of time, 78.6% were allocated to regular household expenses such as consumption whereas 13.4% were allocated to education, 3.9% to housing and 4.1% to savings 11. Hence, remittances might have an important impact on education investment decisions, although the main motivation for sending remittances is not to invest on education The model The theoretical model is based on human capital theory and educational investment decisions linked to remittances. The model is used to analyze the role of international remittances on the investment decision between sending children to a public or a private school, that is, the decision to acquire a higher quality of education. Human capital may be acquired through formal schooling and on-the-job experience in order to increase the probability of gaining a 8 BCRP stands for Central Bank of Peru in Spanish. 9 Own elaboration based on ENAHO. 10 Idem. 11 Idem. 11

25 higher salary and moving up in the social status and economic ladder (Chiswick & Miller, 2009). The model in my research is focused only on formal primary and secondary school. International remittances may be used to acquire more years of formal education and a higher quality of schooling but according to Rapoport & Docquier (2006) there are several motives for sending back money to the origin country. In the case of Peru, 79% of total remittances are used to cover regular household expenses and the rest is used on education, housing, and savings (see Peruvian Background). The pooled income in a household is allocated between consumption and investment goods; however, remittances are more likely to be used to acquire investment goods such as education and health (Yang, 2006). Based on the permanent income hypothesis, if remittances are considered as a temporary income they will be invested rather than spent (McKenzie & Sasin, 2007). Hence, the allocation of remittances income is not perfectly fungible with other income sources of the household. Enrollment at a public school or private school is considered an investment project in human capital to be evaluated on the expected return and cost. Investment in human capital is not only an individual decision but also a decision across generations (Vidal, 1998); parents and children decide how many financial and non-financial resources 12 to invest in schooling. The schooling-decision is influenced by individual and family characteristics, and household financial resources (Berker, 2009; Cox & Ureta, 2003; Jacoby, 1994). Assuming a continuum of school alternatives according to school quality then a higher value of h reflects a higher quality of education. Parents will select the school (where to send their children) based on the following education production function from Catsiapis (1987) but including migration: h h M, T Mg ; H p 0 12 Non-financial resources may be thought as children s effort or time to do schoolwork, or time for studying at home, or parents involvement such as helping their children with their schoolwork. 12

26 where H 0 is the initial stock of human capital that measures parent s education, M are market resources, Mg is migration, and T is parents time. The model does not include p children s time T c because it is assumed that T c affects educational performance but not the school choice between public and private schools. This model captures the effect of absent parents through T p Mg. Generally, remittances- receiving households have a parent living abroad who sends remittances, which T p negatively affect the parent s time allocated to educating his children: < 0. Thus, Mg children do not have emotional support to affect their performance at school. Likewise, it is assumed that parents have a positive influence on the schooling decisions of their children; a parent who invest time in his children s education is more likely to send them to private h schools: > 0. In the empirical model, the number of parents present in the household is T p included to control for the effect of a parent s migration (second channel in the literature review). On the other hand, the cost of producing education depends on 1) direct costs p* M where p is the price of market goods (school tuition or school supplies are examples of direct costs), and it is assumed that private schools have higher direct costs than public schools, 2) forgone earnings for children w c * T c : in developing countries such as Peru children have to p T p work 13, 3) and forgone earnings of parents w * ; w c and w are the opportunity cost of p children s and parent s time in the production of education, respectively. Assuming an initial 0 cost C that can be understood as expenses on subsistence needs, the total cost function is: TC p * M wc * Tc wp * T p Mg 13 In 2007, 2.8% of the children population between 6 and 14 years old participate in the labor market and contribute to the 21.7% of the household budget (INEI, 2009). 13

27 TC HB R Mg C and 0 where HB is the household budget, R is remittances, that is, total cost cannot exceed family resources including remittances after covering subsistence goods C 0, e.g. cloths, food, and R housing. It is assumed that remittances depend positively on migration, > 0 : a Mg household with a migrant member will receive a positive amount of remittances. The model incorporates the two channels reviewed in the literature section, remittances and absenteeism of parents. The objective is to maximize human capital investments by acquiring a higher quality of education subject to education costs, thus: plugging 0 max h = max h M T, T Mg, ; H s.t. TC p * M w * T w T Mg c p c c p * HB R Mg C into the restriction, it follows that: HB R Thus, the Lagrangean equation to maximize is: Mg C p* M w * T w T Mg 0 c c p * max L = max h M, Tc, Tp Mg; H0 HB RMg C0 p* M wc * Tc wp * Tp Mg From the first order condition, the total effect of migration is obtained as follows: p 0 p L Mg = h T p Tp R * w Mg Mg p Tp * Mg Based on the assumptions of this model, L Mg has an ambiguous effect since: h 1. > 0 T p T p h Tp and < 0, thus * < 0, and Mg T Mg p 14

28 R 2. > 0 Mg T and p < 0, but theoretically is not possible to determine whether Mg R Mg Tp > wp * or Mg R Mg Tp < wp *, thus the sign of Mg R Mg T w * p p Mg is undetermined. From the theoretical part, it is clear that an empirical estimation is needed to determine the net effect of migration and remittances on human capital of children left behind. The empirical strategy will account for gender differences of the head of household. There exist gender differences in the intra-household allocation of remittances (Nguyen & Purnamasari, 2011; World Bank, 2008) which suggests a different impact on human capital investments of children left behind depending on the gender of the head of household Data This subsection describes the data collected on households and their educational investment decisions. The empirical work discussed in this chapter is based on the National Survey of Households (ENAHO 14 ) conducted by the National Institute of Statistics and Computing (INEI 15 ) in Peru. The ENAHO is a yearly survey, nationally representative, undertaken by the INEI since 1995 which provides information at national, dominion, and stratum level of statistical inference. Dominion level refers to departments and stratum level refers to ruralurban areas within a department. The information is collected at the individual level for each member of the household 16, although there is some information at the household level provided only by the qualified informant who is a 12 years old or older member of the household. 14 ENAHO stands for National Survey of Households in Spanish. 15 INEI stands for National Institute of Statistic and Computing in Spanish. 16 It does not include members of the Army, Navy or Air Force who are living in barracks. 15

29 The respondents are interviewed every year, and a panel sample is followed through the time. However, some households are randomly dropped and new households are added to the sample. Although, data have been collected since 1995, the panel data analysis proposed in this research only considers ENAHO 2007 through 2010, because the methodology of compiling data was changed in 2004 and the INEI stopped following panel households to re-start a new data of panel households in Before 2004, the ENAHO was executed on a quarterly basis whereas since 2004 it was executed on a monthly basis. The analysis in this paper follows children living in households with and without remittances who reside on the Peruvian coast since it is the main recipient area of international remittances. Including information for the Peruvian highlands and the Peruvian jungle may bias the analysis due to the high percentage of households without any remittances: 99.4% for the Peruvian highlands, and 99.2% for the Peruvian jungle. The balanced panel consists of 559 observations at the individual level for each year from 2007 to 2010 to make a total of 2,236 observations. The unit of analysis in this paper is the child of school-age in primary and secondary schools who is actually attending school instead of using an indicator of being registered. In 2007, children between 3 to 18 years old are included in the analysis to allow for grade advance and delay and they are followed through time. In Peru, children start primary school at 6 years of age and they need to complete 11 years of schooling: 6 years of primary and 5 years of secondary. Children between 4 and 19 are considered in 2008, children between 5 and 20 in 2009, and children between 6 and 21 are considered in This research only includes children who are single and never-married to examine those who depend on their parents support. 16

30 2.6. Variable definitions This research proposes quality of education as the main indicator for human capital formation. The analysis is focused on the role of international remittances in the decision to acquire not only education as measured by more years of education or the enrolment school rate but also a better education through attending private schools instead of public schools. Accordingly, the dependent variable employed is a dichotomous variable, whether the child is attending a private school (high quality) or a public school (low quality). The independent variables included in the analysis are divided into three groups: child, household, and head of household characteristics, and they are based on the theoretical model and similar studies cited in the literature review. The variables sex, only child, age, age squared, and years of education are measured for each child in the data. Household characteristics include an index for household assets, number of children, number of parents present, and logarithm of international remittances received. The characteristics of the head of household included are sex, years of education, age, age squared, and employment status. The perception of the returns to education depends on parents education; thus the head s education is included in the analysis (Acosta et al., 2007; Bennett et al., 2012; Cox & Ureta, 2003; Frisancho & Oropesa, 20122; Kandel & Kao, 2001; Hu, 2012; McKenzie & Rapoport, 2010). A more-educated parent will consider education as an investment good rather than as a consumption good; and he or she will place a higher return on private education than on public education, which makes him or her more likely to send children to private schools. The variable related to the employment status of the head of household is an indicator for having a job (Amuedo-Dorantes & Pozo, 2010). Employment includes both independent and dependent jobs. Additionally, the decision to send children to a private school is associated with the household income level (Cox & Ureta, 2003; Calero et al., 2008; Hu, 2012). Including 17

31 household income, however, may pose an endogeneity problem (Hu, 2012), and a multicollinearity problem arising from the potential correlation with remittances. The endogeneity problem arises due to the correlation of income levels with unobserved factors affecting schooling decisions; for instance, a household suffering a layoff that reduces its income may withdraw children from school. Likewise, a causality problem arises because households with school-age children may undertake certain strategies to increase their income to send their children to school; and the decision to send remittances may be affected by the low income level of the household in the origin country. To address these problems, I follow the literature and include an index for household assets as a proxy for household income. This index is created by using principal component analysis which considers five factors: internet (i.e., has a computer), TV cable (i.e., has a television), land phone, cell phone, and electricity. These five household factors measure household wealth which reflects a more permanent economic status than income level that changes according to current circumstances (Hu, 2012). Households receiving remittances are experiencing migration of at least one member who can be the father or mother of the children in school-age. The literature has found that the positive effect of remittances is offset by the negative effect of absenteeism of parents. To control for parents living outside the household, this research includes children and grandchildren of the head of household in the sample. To measure absenteeism of parents, the variable number of parents living in the household with a range from 0 to 2 is constructed; 0 means that neither of the parents are in the household, 1 means that only one parent (either father or mother) lives in the household, and two means that both parents live in the household, that is, there are no migrant parents; although, it is possible that a disrupted household may be due to a divorce or widowhood rather than migration. Due to limitations 18

32 in the data, however, international migration of parents cannot be identified 17 ; besides, remittances may be sent by a non-member of the household, such as an aunt or a cousin. As Table 2.1 shows for the period of study (2007 to 2010), households receiving remittances have on average more than 1 parent present in the household. The variable indicating being the only child and the number of children in the household (Amuedo-Dorantes & Pozo, 2010; Cox & Ureta, 2003; Frisancho & Oropesa, 2011; Hu, 2012) are included since a higher number of children will reduce the probability of attending a private school. More children in the household pose higher liquidity constraints which makes a higher quality education less affordable. Finally, years of education for children is included to control for any difference between primary and secondary school (Amuedo-Dorantes & Pozo, 2010; Lopez-Cordova, 2005). Table 2.1 reports the descriptive statistics and the test of means for the variables included in the econometric analysis for the pooled sample and for children living in remittances receiving and non-receiving households. A higher proportion of children living in remittances- receiving households attend private schools and they have more years of education. Households receiving remittances have a higher index of wealth, and the head of household has more years of education. Table 2.A in the Appendix presents the correlation matrix. Only years of education and age for children, and number of parents in the household and sex of the head of household show a high correlation coefficient, 0.96 and 0.72, respectively Estimation strategy and hypothesis Two main concerns arise in the estimation of educational investment decisions: selection and endogeneity. The decision of sending children to a private or public school is only observed 17 ENAHO does not gather information of absent members; just indicates that a member of the family is not living in the household. 19

33 for those children who are enrolled in the school (primary and secondary school) which poses a potential selection bias problem. However, since education in Peru is compulsory and public education is provided at negligible cost by the Peruvian government, almost 100% of the children are observed for the econometric analysis of quality of education. Based on the ENAHO, only 1.8% of the pooled data does not attend school for the period The hypothesis to test is that receiving international remittances improves the quality of human capital by sending children left behind to private schools in spite of absenteeism of parents. To incorporate remittances in the analysis, a continuous variable is created to measure the amount of international remittances. Two survey questions are employed: 1) frequency of remittances sent by individuals or households overseas, and 2) amount of remittances sent by individuals or households overseas. However, the endogeneity problem arises in the estimation of the causal effect of remittances on educational investment decisions. The correlation between remittances and the error terms may be explained by two reasons. First, there exists a reverse causality between educational investment and remittances. An individual may decide to migrate and send remittances because he has school-age children, and in turn remittances affect educational investment by loosening liquidity constraints. And, second, there exist unobserved characteristics included in the error terms that may be correlated with both the decision to send remittances and the decision to send children to a private school; for instance, individual attitudes such as ambition may sway the decision to send children to schools with a better quality (private schools) and the decision to migrate for sending remittances. 20

34 To address the endogeneity problem, I follow the literature on remittances, and the historical department-level migration rate is used as an instrument for remittances shocks (Hu, 2012; McKenzie & Rapoport, 2010; Nguyen & Purnamasari, 2011). The migration rate refers to the percentage of permanent out-migrants by department who left Peru and have not returned to Peru during the period The identification assumption is that the historical migration rate at department level influences current migration behavior which is a good indicator for receiving remittances but the historical migration rate at the department level does not affect current schooling-decision at the household level. On the other hand, Calero et al. (2008) proposed to use the number of Western Union offices as an instrumental variable, based on the identification assumption that the higher number of agencies will reduce the transaction costs and increase the accessibility to remittances; although the number of agencies may capture unobserved characteristics of the region such as the local economy. The results using the number of agencies (MoneyGram and Western Union) are presented in Appendix Table 2.B, and similar results were found with those using historical migration rate. The remittances equation needs to be estimated by a tobit model because remittances are a left-censored variable. Households that receive remittances record a positive amount whereas households that do not receive remittances always record zero. The money received by households is always positive since it is not possible to send a negative amount of money. I follow Vella & Verbeek (1999) to estimate the education investment equation by a two-step panel data model with a censored endogenous variable. Unlike other panel data estimators, Vella & Verbeek s methodology proposes to separate time-invariant individual effects from the individual time effects and captures the state dependence producing endogeneity. This method estimates asymptotically efficient estimators within a limited information framework (LIML). 21

35 The education investment model to estimate is the primary equation (Eq. 1) whereas the reduced form for remittances (endogenous variable) is the auxiliary equation (Eq. 2). Consider the model for N children observed through T consecutive time periods: (Eq. 1) * y it = 0 xit1 R it 2 i it where * y it is unobservable but y it = 1 if * y it > 0, i.e., children attend private school; and zero otherwise, i.e., children attend public school. (Eq. 2) * R it = 0 xit1 Rit 12 NBit3 i it where * R it is partly unobservable as R it = * R it if * R it > 0 and zero otherwise, i indexes children i 1,..., N, and t indexes time t 1,...,T. xit is the vector of independent exogenous variables described in Table 2.1, NB it is the number of household members, and R it is remittances which are endogenous. The lagged remittances included in Eq. 2 controls for dynamics and state dependence. R it1 and NB it are the exclusion restrictions needed to estimate the model, and they are not included in the educational investment equation. R it1 is excluded under the assumption that its effect on school investment operates only through its current values since school decisions depends on the current liquidity constraints (assuming no savings). NB it is excluded on the basis that the higher number of members in the household will pose a higher requirement of remittances and although it may pose a restriction on the education quantity, it will not sway the quality schooling decision. The errors include random invariant individual effects, i and i, and random individual specific time effects, it and it which are assumed to be independent across individuals. The errors are defined as follow: it i it ; uit i it ; 1 T ui T u t it, E 2 x it, u u u is assumed to allow for heteroskedasticity and and it it 1 it i 22

36 autocorrelation in it. To correct for endogeneity, u it and u i are included as additional explanatory variables with coefficients 1 and 2 to obtain consistent estimators for the primary equation. The endogeneity test is the significance of 1 and 2 ; indicates that there is no endogeneity. The term u i indicates the time-invariant unobserved individual effects, and u it indicates the time varying effects. The lagged value of remittances included in Eq. 2 cannot be assumed truly exogenous due to the presence of individual effects i. Vella & Verbeek (1999) propose to follow Heckman (1981) and include pre-sample information of all the exogenous variables x as instrumental variables to estimate the lagged value of remittances. The lack of data prior to 2007 makes it necessary to use a different approach but following in a similar vein. Following the literature, the historical rate of department migration is considered as instrumental variable for lagged values of remittances. Additionally, lagged values of independent exogenous variables are included in the estimation of the following equation: R x Z NB (Eq.3), and the predicted values are used to it1 0 it1 1 it1 2 it 3 it1 it estimate Eq.2. Eq.2 and Eq.3 are estimated with a random effect tobit model, and the primary equation is estimated with a random effect probit model. Finally, the pooled panel data with a standard two-stage method with instrumental variables for remittances and the random effect probit model are estimated to highlight the differences and ensure the robustness of the results. To test the significance of the historical rate of department migration as instrumental variable, the first-stage F-statistic was employed (rule of thumb: F>10). 23

37 2.8. Results In this section, four sets of regression results are reported over the full sample to estimate the dependent variable. The first one shows the results from the Vella & Verbeek (1999) model considering lagged remittances as exogenous, the second one accounts for the endogeneity of lagged remittances, the third one presents the two-stage estimation of the pooled panel data correcting for endogeneity with instrumental variables method, and the last set is the random effect probit model considering remittances as exogenous. Finally, two additional sets of estimates are reported dividing the sample by gender of the head of household. The results for the remittances equation (Eq. 2) are presented in the Appendix Table 2.C. Three tobit models are estimated to calculate the correction terms ( u it and u i ), and the predicted values for the estimation of the primary equation (Table 2.2). The instrumental variable, historical rate of department migration, is significant at the 99% confidence level with a F-stat of 9.54 and (rule of thumb: F>10) for the corresponding models 2 and 3, respectively. Table 2.2 reports the estimated coefficients for the four sets of regressions, while marginal effects are presented in the Appendix Table 2.D. The results from these models are similar in their significance, although the variables: only child, age, age squared, and job status of head of household are no longer significant when endogeneity is corrected for in columns (2) and (3). However, in all the models, the logarithm of international remittances received has a statistically significant and positive effect on the human capital investment of children left behind. The models using the Vella & Verbeek (1999) correction shows that only u i is statistically significant, and its negative effect suggests that the time-invariant unobserved individual effects make it less likely to send children to a private school as remittances increase. For instance, if u i reflects the unmeasured productivity of a family business, this 24

38 will increase remittances and decrease the probability of sending children to a private school. On the other hand, u it is not significant which suggests that the time varying effects generating the simultaneity between school decisions and remittances do not affect the school investment decision. However, according to the Wald test the null hypothesis 0 is 1 2 rejected for both models (columns 1 and 2) with a statistic of (prob > chi2 = ) and (prob > chi2 = ), respectively. This suggests that correcting for endogeneity is the best approach to estimate the educational investment equation (Eq. 1), and the second model (column 2) is preferred. The international remittances effects on educational investment are similar regardless of the model employed but the magnitude varies: the Vella & Verbeek correction considering lagged remittances as exogenous estimates the larger effect, and the smaller effect is estimated by the pooled probit. These results suggest that receiving a higher amount of remittances increases the probability of children attending private schools whether or not the endogeneity of remittances is controlled for. Furthermore, remittances remain significant regardless of the number of parents living in the household (which does not have a significant effect). The results for the second model correcting for endogeneity of lagged remittances are explained below. Children s age does not have a linear effect on the probability to attend a private school; age has a negative effect and age squared has a positive effect which reflects the greater investment in secondary school-age children. In Peru, the entrance examination to the university is based on the secondary education; it is expected that parents would decide to acquire a higher quality education for their children attending secondary school. The household variables, index for household assets and number of children, have a significant effect on the probability to send children to a private school. Children living in households with a higher index of assets are more likely to attend a private school by 2.3% since index of 25

39 assets is a proxy for household wealth then a wealthier household will face less liquidity constraints. On the other hand, the greater number of children in the household will decrease the likelihood to attend a private school by 2.3%: more children pose a competition for the financial resources. Finally, only two variables of the head of household group have a significant effect on the schooling decision. Children with a male head of household are more likely to attend a private school by 7.9% while the positive effect of a head s education is only 0.6%. As stated in the theoretical section, a well-educated parent will invest more in his children s education through private education. The previous results suggest that there is a gender difference in the decision to send children to private schools. To complete the analysis, two samples are created by dividing by the gender of the head of household to analyze the effect of remittances. The estimated coefficients presented in Table 2.3 replicate the second model for the full sample (column 2 in Table 2.2). The main problem in the estimation is that the sample for female head of household is very small, only 153 observations and the instrumental variable, historical rate of department migration, is not significant (Appendix Table 2.E). Remittances have a positive and significant effect for households with a female-head in comparison to the nonsignificant effect for households with a male-head which suggests that remittances in a female-head household are more likely to be employed to improve quality of education for children. To test the sensitivity of the results from the second model (Column 2, Table 2.2), one additional model was estimated. A dichotomous variable for Lima (capital of Peru) is included in the independent variables set (see Apendix Table 2.F). Ten departments are included in the analysis; however, Lima receives 41.5% of the international remittances, and being the capital of a highly centralized country such as Peru, makes it more relevant to control for Lima than for department effects. The results remain similar with a significant 26

40 and positive effect of remittances on the decision to send children to private schools and Lima does not have a significant effect Conclusions This research has provided theoretical and empirical analyses of the effect of remittances controlling for absenteeism of parents on the human capital investments on children left behind. Using Peruvian data for the period , this study estimated a two-step human capital equation to address endogeneity of remittances. Unlike previous studies, this research evaluates the effect of international remittances on the probability of acquiring a higher quality of education by sending children to private schools. Much of the literature has focused on the effects of either remittances or migration, and only a quantitative educational outcome has been considered in previous analysis. Migration in Peru, unlike from other countries close to the main destination country (U.S.), is undertaken mostly by relatively affluent people; hence the effect of migration and remittances should be evaluated not on quantity but quality of education. Furthermore, education coverage in Peru is at the same levels as in developed countries but the quality remains at among the lowest places in international evaluations such as PISA. The results suggest that families receiving remittances are more likely to send their children to private schools. Thus, unlike other studies for Peru, this research finds a positive and statistically significant effect of remittances on human capital which is not only an individual benefit but also a key factor for growth and development for the country. Although, international remittances have an important role on improving human capital, there is little room for policy interventions on private decisions made by receiving-remittances households. However, new financial instruments to reduce transaction costs of remittances can be created by governments and private firms such as the cell phone-based remittance 27

41 service implemented by Kenyan telecom provider (Yang, 2011). Evidence based on Salvadoran migrants in the U.S. shows that reducing transaction fees increases the amount of remittances (Aycinena et al., 2010) References Acosta, P., P. Fajnzylber & H. Lopez (2007). The impact of remittances on poverty and human capital. Evidence from Latin American households surveys. Journal of Population Economics, 19 (2), pp Amuedo-Dorantes, C. & S. Pozo (2010). Accounting for remittance and migration effects on children s schooling. World Development, 38 (12), pp Aycinena, D., C. Martinez & D. Yang (2010). The impact of remittance fees on remittance flows: Evidence from a field experiment among Salvadoran migrants. personal.umich.edu/~deanyang/papers/aycinena%20martinez%20yang%20- %20remittances.pdf BCRP - Banco Central de Reservas del Peru. Cuadros trimestrales históricos. Bennett, R., D. Clifford & J. Falkingham (2012). Household members migration and the education of children left behind : Empirical findings from Tajikistan and reflections for research practice. Population, Space and Place, DOI: /psp Berker, Ali (2009). The impact of internal migration on educational outcomes: Evidence from Turkey. Economics of Education Review, 28, pp Booth, A. & Y. Tamura (2009). Impact of paternal temporary absence on children left behind. IZA, 4381 Borraz, Fernando (2005). Assessing the impact of remittances on schooling: The Mexican experience. Global Economy Journal, 5 (1), pp Calero, C., A. Bedi & R. Sparrow (2008). Remittances, liquidity constraints and human capital investments in Ecuador. World Development, 20, pp Catsiapis, George (1987). A model of educational investment decisions. The Review of Economics and Statistics, 69 (1), pp Cox, A. & M. Ureta (2003). International migration, remittances, and schooling: Evidence from El Salvador. Journal of Development Economics, 72, pp ECLAC The Economic Commission for Latin America. 28

42 Frisancho, V. & R. Oropesa (2011). International migration and the education of children: Evidence from Lima, Peru. Population Research and Policy Review, 30 (4), pp Gould, E., V. Lavy & D. Paserman (2009). Does immigration affect the long-term education outcomes of natives? Quasi-experimental evidence. The Economic Journal, 119, pp Griffith, A. & D. Rothstein (2009). Can t get there from here: The decision to apply to a selective college. Economics of Education Review, 28, pp Hu, Feng (2012). Migration, remittances, and children s high school attendance: The case of rural China. International Journal of Educational Development, 32 (3), pp INEI-National Institute of Statistics and Computing. INEI (2009) -National Institute of Statistics and Computing. Perú: Niños, niñas y adolescentes que trabajan, Lima, 175p. IOM International Organization of Migration (2010). Peru: Estadisticas de la emigracion internacional de peruanos e inmigracion de extranjero, , 278p. Jacoby, Hanan (1994). Borrowing constraints and progress through school: Evidence from Peru. The Review of Economics and Statistics, 76 (1), pp Kandel, W. & G. Kao (2001). The impact of temporary labor migration on Mexican children s educational aspirations and performance. International Migration Review, 35 (4), pp Lopez-Cordova, E. (2005). Globalization, migration and development: The role of Mexican migrant remittances. Economia, 6 (11), pp Lucas, R. & O. Stark (1985). Motivations to remit: Evidence from Botswana. Journal of Political Economy, 53, pp McKenzie, D. & H. Rapoport (2010). Can migration reduce educational attainment? Evidence from Mexico. Journal of Population Economics, 24 (4), pp McKenzie, D. & M. Sasin (2007). Migration, remittances, poverty, and human capital: Conceptual and empirical changes. World Bank Policy Research Working Paper, Meza, L. & C. Pederzini (2008). Migracion internacional y escolaridad como medios alternativos de movilidad social: el caso de Mexico. Paper presented at the European population conference, Barcelona. Nguyen, T. & R. Purnamasari (2011). Impacts of international migration and remittances on child outcomes and labor supply in Indonesia: How does gender matter? World Bank Policy Research Working Paper, Schwartz, A. & L. Stiefel (2006). Is there a nativity gap? New evidence on the academic performance of immigrant students. American Education Finance Association. 29

43 Stark, O. & D. Bloom (1985). The new economics of labor migration. American Econmic Review, 75, pp Stark, O. & Y. Wang (2002). Inducing human capital formation: migration as a substitute for subsidies. Journal of Public Economics, 86 (1), pp Vella, F. & M. Verbeek (1999). Two-step estimation of panel data models with censored endogenous variables and selection bias. Journal of Econometrics, 90, pp Vidal, Jean-Pierre (1998). The effect of emigration on human capital formation. Journal of Population Economics, 11, pp UMC Unidad de medicion de la calidad educativa (2004). Una aproximacion a la alfabetizacion lectora de los estudiantes peruanos de 15 años. Resultados del Peru en la evaluacion internacional PISA. Documento de trabajo #6. World Bank (2008). The international migration of women, edited by A. Morrison, M. Schiff, and M. Sjoblom. Washington, DC. Yang, Dean (2006). International migration, human capital, and entrepreneurship: Evidence from Philippine migrants exchange rate shocks. World Bank Policy Research Working Paper, Yang, Dean (2011). Migrant remittances. Journal of Economic Perspectives, 25 (3), pp

44 Table 2.1. Descriptive statistics Dependent variable Remittance Non-remittance Pooled Data ( ) households households Obs Mean Std. Dev. Min Max Mean Mean Child attending private school (=1) *** Independent variables Child characteristics Sex (1=male) * Only child (=1) * Age Age squared Years of education *** Household characteristics Index for household assets *** Number of children Number of parents present *** Log of international remittances received Head of household characteristics Test of means 1/. Sex (1=male) *** Years of education for head of household *** Age for head of household *** Age squared for head of household *** Head of household has a job (=1) *** 1/. z-test for dichotomous variables and t-test for continuous variables. ***p<0.01, **p<0.05, *p<0.1 31

45 Table 2.2. Educational investment equation Method Random effect probit Random effect probit Pooled probit Random effect probit Vella & Verbeek (1999) correction Yes Yes No No Remittances equation Dynamic Dynamic First-stage Lagged values of remittances Exogenous Endogenous Instrumental variable Historical rate of migration Historical rate of migration Dep. Var.:Child attending private school (=1) (1) (2) (3) 1/. (4) Child characteristics Sex (1=male) (1.54) (0.73) (1.30) (-0.13) Only child (=1) ** (-1.09) (-0.00) (0.16) (2.54) Age * ** *** ** (-1.83) (-2.51) (-3.38) (-2.46) Age squared 0.029*** 0.025*** 0.009*** 0.014*** (3.49) (3.16) (4.33) (2.66) Years of education (-0.87) (0.44) (-0.21) (1.06) Household characteristics Index for household assets *** 0.299*** 0.536*** (-0.11) (3.09) (7.61) (6.94) Number of children *** *** *** (-2.98) (-2.95) (-4.94) (-1.46) Number of parents present (1.15) (-0.29) (-0.44) (-0.14) Log of international remittances received 0.328** 0.125* 0.032*** 0.111*** (2.26) (1.77) (3.59) (2.69) Head of household characteristics Sex (1=male) 1.874** 1.438** 0.271* (2.33) (2.10) (1.91) (1.38) Years of education for head of household * 0.034** 0.137*** (-0.42) (1.69) (2.56) (4.10) Age for head of household *** (-0.05) (-0.16) (-1.00) (-3.48) Age squared for head of household *** (-0.50) (-0.01) (0.71) (2.91) Head of household has a job (=1) 1.569* (1.78) (0.98) (1.16) (-0.65) u it (-1.50) (0.01) u i *** *** (-3.37) (-3.73) Constant 6.783* ** (1.79) (1.17) (2.09) (-0.77) Observations 1,616 1,615 2,171 2,171 Number of groups /. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 32

46 Table 2.3. Educational investment equation by gender of head of household Method Vella & Verbeek (1999) correction Remittances equation Lagged values of remittances Instrumental variable Dep. Var.:Child attending private school (=1) Child characteristics Sex (1=male) (0.52) (0.98) Only child (=1) (0.33) *** (-2.77) Age * (-1.79) *** (-2.82) Age squared 0.023** (2.49) 0.203** (2.23) Years of education (0.27) 4.608** (2.14) Household characteristics Index for household assets 0.506*** (2.87) 2.597* (1.86) Number of children *** (-2.76) (0.11) Number of parents present (1.19) *** (-2.88) Log of international remittances received (-0.15) 1.935*** (3.42) Head of household characteristics Years of education for head of household 0.140** (2.24) (-0.90) Age for head of household (-1.56) *** (-2.82) Age squared for head of household (1.48) 0.178*** (2.73) Head of household has a job (=1) 1.899* (1.96) *** (-3.24) u it Random effect probit Male sub-sample (1.12) ** (-2.49) u i *** (-5.29) *** (-2.77) Constant (1.51) *** (2.94) Observations 1, Number of groups z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 Yes Dynamic Endogenous Historical rate of migration Random effect probit Yes Dynamic Endogenous Historical rate of migration Female sub-sample 33

47 Millions of US Dollars Number of permanent emigrants Figure 2.1. Remittances and permanent emigration, ,000 3, , ,000 Remittances Emigrants 2,500 2, ,000 1, ,000 1,000 50, Source: OIM (2010) and BCRP. 34

48 Appendix A Table 2.A. Correlation matrix Dependent variable Dep. Var Attending private school Child attending private school (=1) 1 Child characteristics Sex (1=male) Sex (1=male) Only child (=1) Only child (=1) Age Age Age squared Age squared Years of education Years of education Household characteristics Index for hh assets Index for household assets Number of children Number of children No. parents present Number of parents present Log of remittances Log of international remittances received Head of household characteristics Child characteristics Household characteristics Head of household characteristics Sex (1=male) Sex (1=male) Years education head Years of education for head of household Age for head of household Age for head of household Age squared for head of household Age sq. for head of hh Head of hh has a job Head of household has a job (=1)

49 Table 2.B. Educational investment and remittances equations using agencies as instrumental variable Method Random effect tobit Random effect tobit Random effect probit Vella & Verbeek (1999) correction Yes Remittances equation Dynamic Lagged values of remittances First-stage Endogenous Instrumental variable Number of agencies Lagged remittances 1/. Remittances 2/. Attending private school Child characteristics Sex (1=male) (-1.43) (-0.48) (0.76) Only child (=1) (1.54) (1.26) (-0.04) Age ** (-0.86) (-0.20) (-2.48) Age squared *** (0.52) (-0.37) (3.14) Years of education (0.94) (0.81) (0.43) Household characteristics Index for household assets 2.489*** 1.449* 0.408*** (3.27) (1.72) (3.06) Number of children *** (0.67) (0.47) (-2.99) Number of parents present * (-1.47) (-1.85) (-0.25) Number of household members (0.66) (0.20) Log of international remittances received *** 0.126* (-0.78) (3.15) (1.79) Head of household characteristics Sex (1=male) 0.718** ** (2.11) (-0.59) (2.11) Years of education for head of household *** (-3.30) (1.05) (1.64) Age for head of household 0.012*** (3.54) (0.52) (-0.16) Age squared for head of household * (-1.81) (-0.41) (-0.02) Head of household has a job (=1) * (-1.91) (1.01) u it (-0.03) u i *** (-3.68) Number of agencies 0.034*** (3.11) Constant (-1.08) (-0.52) (1.19) F-stat 9.68 Observations 1,658 1,645 1,615 Number of groups /. Lagged independent variables are included in the estimation. 2/. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 36

50 uti Table 2.C. Remittances equation Method Random effect tobit Random effect tobit Random effect tobit Pooled Tobit Lagged remittances Exogenous First-stage Endogenous Instrumental variable Historical rate of migration Historical rate of migration Dep. Var. Remittances Lagged remittances 1/. Remittances 2/. Remittances Model (related to Table 2) (1) (2) (3) Child characteristics Sex (1=male) (-1.37) (-1.43) (-0.49) (-1.44) Only child (=1) ** (1.62) (1.55) (1.26) (2.26) Age (-0.38) (-0.88) (-0.20) (-1.39) Age squared (-0.35) (0.53) (-0.37) (0.07) Years of education ** (1.15) (0.97) (0.82) (2.12) Household characteristics Index for household assets 2.667*** 2.482*** 1.461* 2.371*** (3.39) (3.26) (1.74) (4.39) Number of children *** (0.57) (0.67) (0.47) (3.32) Number of parents present ** * *** (-2.57) (-1.48) (-1.86) (-4.12) Number of household members (0.52) (0.67) (0.20) Lagged log of international remittances received *** (0.63) (3.13) Head of household characteristics Sex (1=male) (-0.94) (-0.78) (-0.60) (-0.47) Years of education for head of household 0.838** 0.727** *** (2.18) (2.13) (1.06) (2.63) Age for head of household *** *** (-0.20) (-3.31) (0.52) (-3.36) Age squared for head of household *** *** (0.62) (3.55) (-0.41) (4.30) Head of household has a job (=1) * * * ** (-1.93) (-1.79) (-1.92) (-2.45) Historical rate of department migration 0.209*** 0.240*** (3.09) (5.72) Constant (-0.82) (-1.02) (-0.52) (-0.68) F-stat Observations 1,646 1,658 1,645 2,212 Number of groups /. Lagged independent variables are included in the estimation. 2/. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 ui 37

51 Table 2.D. Marginal effects for the educational investment equation Method Random effect probit Random effect probit Pooled probit Random effect probit Vella & Verbeek (1999) correction Yes Yes No No Remittances equation Dynamic Dynamic First-stage Lagged values of remittances Exogenous Endogenous Instrumental variable Historical rate of migration Historical rate of migration Dep. Var.:Child attending private school (=1) (1) (2) (3) 1/. (4) Child characteristics Sex (1=male) (1.56) (0.73) (1.29) (-0.13) Only child (=1) ** (-1.11) (-0.00) (0.16) (2.48) Age * ** *** ** (-1.79) (-2.46) (-3.39) (-2.37) Age squared 0.002*** 0.001*** 0.002*** 0.001** (3.46) (3.08) (4.36) (2.57) Years of education (-0.88) (0.44) (-0.21) (1.05) Household characteristics Index for household assets *** 0.057*** 0.036*** (-0.11) (2.97) (7.96) (5.51) Number of children *** *** *** (-2.93) (-2.72) (-5.14) (-1.40) Number of parents present (1.17) (-0.29) (-0.44) (-0.14) Log of international remittances received 0.018** 0.007* 0.006*** 0.007*** (2.32) (1.76) (3.58) (2.63) Head of household characteristics Sex (1=male) 0.102** 0.079** 0.052* (2.32) (2.04) (1.91) (1.37) Years of education for head of household * 0.006*** 0.009*** (-0.42) (1.73) (2.58) (3.76) Age for head of household *** (-0.05) (-0.16) (-1.00) (-3.29) Age squared for head of household *** (-0.50) (-0.01) (0.71) (2.79) Head of household has a job (=1) 0.086* (1.80) (0.98) (1.16) (-0.65) u it (-1.53) (0.01) u i *** *** (-3.39) (-3.73) Observations 1,616 1,615 2,171 2,171 Number of groups /. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 38

52 Table 2.E. Remittances equation by gender of the head of household Method Random effect tobit Random effect tobit Random effect tobit Random effect tobit Lagged remittances First-stage Endogenous First-stage Endogenous Instrumental variable Historical rate of migration Historical rate of migration Dep. Var. Lagged remittances 1/. Remittances 2/. Lagged remittances 1/. Remittances 2/. Model (related to Table 3) Child characteristics Sex (1=male) (-1.51) (0.11) (-0.45) (-0.42) Only child (=1) (1.09) (1.13) (0.46) (0.92) Age (-1.43) (-0.29) (0.81) (0.84) Age squared (0.17) (-0.12) (-1.03) (0.03) Years of education 2.919* (1.92) (0.65) (0.52) (-1.41) Household characteristics Index for household assets 2.873*** 2.321** (3.24) (2.33) (0.20) (0.37) Number of children 6.020** (2.06) (1.41) (-1.11) (-0.75) Number of parents present (-0.45) (-0.71) (-0.13) (-0.34) Number of household members ** (-1.54) (-1.41) (2.16) (1.53) Lagged log of international remittances received 0.476** (2.42) (0.02) Head of household characteristics Male Sample Years of education for head of household *** (0.89) (-0.04) (1.31) (2.65) Age for head of household *** ** (-3.03) (0.54) (0.39) (2.06) Age squared for head of household 0.014*** ** (3.38) (-0.27) (-0.37) (-2.04) Head of household has a job (=1) ** (-2.04) (-0.77) (-0.44) (0.02) Historical rate of department migration 0.186** (2.50) (0.79) Female Sample Constant * (0.42) (-0.52) (-1.82) (-0.27) F-stat Observations 1,461 1, Number of groups /. Lagged independent variables are included in the estimation. 2/. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 39

53 Table 2.F. Educational investment and remittances equations controlling for Lima Method Random effect tobit Random effect tobit Random effect probit Vella & Verbeek (1999) correction Yes Remittances equation Dynamic Lagged values of remittances First-stage Endogenous Instrumental variable Historical rate of migration Lagged remittances 1/. Remittances 2/. Attending private school Child characteristics Sex (1=male) (-1.49) (-0.69) (0.97) Only child (=1) 6.903* 7.852* (1.74) (1.88) (-0.65) Age * (-0.81) (-0.45) (-1.94) Age squared *** (0.61) (0.02) (2.97) Years of education (0.77) (0.86) (-0.03) Household characteristics Index for household assets 2.452*** 1.568* 0.330* (3.24) (1.90) (1.72) Number of children *** (0.85) (0.81) (-2.68) Number of parents present * (-1.52) (-1.89) (0.31) Number of household members (0.57) (0.42) Log of international remittances received * (1.27) (1.94) Head of household characteristics Sex (1=male) ** (-0.89) (-1.15) (2.03) Years of education for head of household 0.745** 0.697* (2.17) (1.81) (0.54) Age for head of household *** (-3.37) (0.14) (-0.19) Age squared for head of household 0.012*** (3.59) (0.15) (-0.10) Head of household has a job (=1) * * (-1.70) (-1.87) (1.37) u it u i Historical rate of department migration 1.807** (2.11) (-0.91) ** (-2.45) Lima (Capital of Peru) * 7.761** (-1.89) (2.52) (-1.11) Constant (-1.50) (-0.89) (1.35) F-stat 4.45 Observations 1,658 1,645 1,615 Number of groups /. Lagged independent variables are included in the estimation. 2/. Predicted remittances from the first-stage are included as predictor in the independent variables set. z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 40

54 Chapter 3 SECOND ESSAY Social networks and small rural firm performance 3.1. Introduction This research is based on the Small Farms Clusters (SFC) Project that began in 2006 and employed data from small rural firms located in the Northeast United States (Maryland, New York, and Pennsylvania). One of the main disadvantages small firms face is limited access to production inputs including information about markets beyond the local economy (Gronum et al., 2012; Malecki, 1993). In that sense, clusters are important because they can provide an environment in which to exchange information and knowledge among firms or enterprises with similar business interests that compete but also cooperate (Brasier et al., 2007; Porter, 2000), as well as clusters are recognized as an important strategy for rural development (Das & Das, 2011). Moreover, a cluster facilitates relationships among small firms with government and international organizations, which allows its members to acquire new resources such as training and information (Das & Das, 2011). However, empirical studies do not agree on the impact of clusters on entrepreneurship and development (Rocha, 2004). Cluster membership does not guarantee benefits. Hence, network ties or joint ventures within an organization such as a cluster or an industry have been underscored in the literature on organizational strategy and regional development as a mechanism to improve the economic performance of firms (Chell & Baines, 2000; Goerzen & Beamish, 2005; Rocha, 2004; Walter et al., 2006), and as a strategy to overcome the lack of resources especially for small firms 41

55 (Atterton, 2007; Parida et al., 2010). In the case of small rural firms, building ties arises from their geographical closeness and it tends to consist of stronger but homogenous ties (Atterton, 2007; Johannisson, 1986; Malecki, 1993). Nevertheless, there is a lack of research of the impacts of network ties on small rural firms economic performance. This research will test the hypothesis that building social support and economic relationships 18 within a cluster enhances the performance of small rural firms belonging to that cluster. There are two specific objectives: 1) to evaluate the effect of network ties within a cluster on the economic viability of small rural business, and 2) to evaluate the impact of a network s topology on small rural firms. There are three main contributions of my research. First, prior research on social networks mostly focuses on intra-firms and inter-firms linkages within the industrial sector including customers, suppliers or alliances. My research extends the literature by analyzing network linkages among small rural firms at the individual level represented by entrepreneurs within agricultural clusters rather than the ties to organizations resources. Second, the applied literature on social network theory is small and there is not a consensus about the impacts on firm performance (Goerzen & Beamish, 2005; Johannisson, 1986). My research employs unique survey data to contribute to the empirical literature by analyzing the role of network ties on the economic performance of small rural firms. Third, although there is literature focusing on network ties, there is little attention to their characteristics and directionality, especially for indirect ties (Martinez & Aldrich, 2011). To the best of my knowledge, Sparrowe et al. (2001) includes an in-degree indicator to analyze individual performance based on a leader s evaluation but for an intra-organizational study, and Baldwin et al. (1997) analyze closeness on student s performance. This research will include centrality indicators to capture direct and indirect ties 18 In this research, relationships and network ties are used interchangeably. 42

56 among rural entrepreneurs, and network density to assess the relevance of network ties within each rural cluster studied for the firm s performance. The paper is organized as follows. The literature and framework relating economic growth and social networks is explained in section 2, while data and variables are explained in section 3. Hypotheses are proposed in section 4, and research methods outlined in section 5. Results and conclusions are discussed in section 6 and 7, respectively Literature and framework Neoclassical theory states that economic growth understood as a quantitative change is a function of two traditional inputs: physical capital and labor. However, those two factors have failed to explain about 50% of historical economic growth in industrialized countries, which is known as the Solow residual (Nelson, 1981; Rocha, 2004). Hence, the literature has started to focus on new inputs of production beyond the traditional ones. Technological change and human capital embedded in entrepreneurship have been added as variables by the new endogenous growth theory to explain economic growth (Rocha, 2004). Entrepreneurship is identified with the capacity to create new organizations or new economic activities of small or medium size (Rocha, 2004). Likewise, social capital has arisen as an important input of economic growth. Putnam (2000, p.67) points out that social capital refers to features of social organization, such as networks, norms, and trust that facilitate coordination and cooperation for mutual benefits. Following this line, management s studies employ network indicators to evaluate the effect of social capital on leadership, employment, creativity, and individual and team performance (Borgatti & Foster, 2003). Studies such as Greve et al. (2010) use degree centrality as an 43

57 indicator of social capital, and they found a positive relationship with individual productivity. This research employs network centrality indicators (out-degree and closeness) to analyze the effect of the direct and indirect linkages created by a member on his or her firm s performance. The location or position of individuals within social networks has become relevant to explain firm organization and performance (Granovetter, 1985). Social network research helps us to understand the agent s social environment and uncovers influence processes and leveraging processes (Borgatti et al., 2009). A social network is a structure or system that embeds agents and their social interactions with other agents, where an agent is represented by a node, and the interactions among agents are represented by edges. These interactions may or not have directionality. Undirected interactions reflect a link between two agents but without an orientation, such as a geographic relationship (e.g., being neighbors). Meanwhile, directed interactions reflect an orientation from one node to another, for instance to give advice from A to B, but it does not necessarily work the other way, that is, it is possible that B does not give any advice to A (Borgatti & Foster, 2003; Wasserman & Faust, 1994). In this research, agents are the entrepreneurs who own small rural firms, and the edges are directed interactions of advice. The literature on social networks recognizes that network ties may affect firm performance directly or indirectly through innovation, in the same vein as the classification proposed by Borgatti & Foster (2003) between structuralist and connectionist studies. According to structuralist studies, network ties have a direct effect regardless of the information transmitted through contacts, and it focuses on the network s characteristics. On the other hand, for connectionist studies the effect of network ties is indirect through the flow of information. A network is seen as a small-world that is globally and locally efficient in the exchange of information among its members (Latora & Marchiori, 2001). 44

58 Network ties play an important role in creating and spreading knowledge, innovation, and learning processes among entities, persons, and regions, which in turn influences their management, production techniques, profit maximization and productivity (Atterton, 2007; Borgatti & Foster, 2003; Goerzen & Beamish, 2005; Hassanein & Kloppenburg, 1995; Malecki, 1993; Parida et al., 2010; Reagans & Zuckerman, 2001). The macroeconomic theory of economic growth states that technology transfer has a market price whereas knowledge spillovers may be acquired at no cost (Audrestch, 2007). The idea of knowledge spillovers is applied to a micro level and it can be captured by interactions within social networks. Thus, network ties become an important strategy especially for small and medium firms to learn about new technologies and new products at a low cost from external sources which allow them to survive and keep growing (Brasier et al., 2007; Parida et al., 2010). The transmission of ideas and innovations within networks enhances entrepreneurship, and facilitates access to new opportunities, new knowledge and specialized production resources (Hoang & Antoncic, 2003; Johannisson, 1986; Parida et al., 2010). This interchange of ideas and contact with other entrepreneurs make network s members more likely to be aware of their limitations and abilities (McQuaid, 1996). Therefore, small and medium farmers who invest on their network ties will improve their firm performance and results because they may exploit the information received and take advantage of entrepreneurship opportunities transmitted by their direct and indirect contacts within the network (Atterton, 2007; Johannisson, 1986; Parida et al., 2010). Similarly, in physics the preferential attachment hypothesis states that individuals who already have a large number of connections will interact with other individuals at a higher rate compared to those with fewer connections within the social network (Barabasi & Albert, 199; 45

59 Barabasi et al., 2002; Newton, 2001). Barabasi et al. (2002) and Newton (2001) show that network ties can enhance the performance of members, for example, authors with more connections (measured by co-authorship) increase their number of journal publications (i.e., performance). This large number of connections gives a central position to the agent within the network which is related to structuralist studies that contend position drives performance results (Borgatti & Foster, 2003). Based on a qualitative study, Sandstrom & Carlsson (2008) found that an efficient network within the higher education policy sector has a high density score and its members have a high degree centrality. Likewise, Baldwin et al. (1997) show that centrality in friendship, communication, and adversarial networks affect grades (i.e., performance) of business administration students. However, some researchers contend that network ties may have a negative effect on firm s performance. Atterton (2007, p.231) points out that networks can also become restricting, conservative and stifling if they involve no new players or information, and thus can result in negative impacts. In the case of social capital, Portes (1998, p.15) identifies four negative outcomes: exclusion of outsiders, excess claims on group members, restrictions on individual freedoms, and downward leveling norms. In the same vein, Chell & Baines (2000) state that network ties may exclude and reinforce barriers to new members. The exclusion of outsiders limits the access to diverse information and redundant information will flow within the network. Likewise, excess claims on group members and restrictions on individual freedoms may force agents to take business decisions based on social relations rather than on economics arguments (Atterton, 2007; Johannisson, 1986). Research on network ties has been mostly applied to large enterprises such as multinational or multilevel firms. In those cases, network ties are related to the intra-firm 46

60 organization and inter-firm alliances or joint ventures which improve market access, innovation transfer, and the acquisition of complementary technological capabilities (Goerzen & Beamish, 2005). For instance, Goerzen & Beamish (2005) study 580 Japanese multinational enterprises and their network ties by using three measures of product entropy (i.e., diversity). They find that higher entropy has a negative effect on economic performance measured by return on assets, return on sales, and return on capital. In an intra-firm study, Sparrowe et al. (2001) show a positive relationship between job performance and those employees with a central location in an advice network, as is predicted by structuralist studies. In the case of small firms, Parida et al. (2010) investigate three types of relationships, with suppliers, with customers, and with competitors, universities, and government agencies, controlling for network size. They find that number of contacts and relationships with competitors, universities and government agencies have significant and positive effects on firm performance. Gronum et al. (2012) measure network ties as the frequency with which a business seeks information or advice and they find a small positive relationship with sales growth. However, both Parida et al. (2010) and Gronum et al. (2012) include network indicators that are not directional, unlike the indicators included in this research. On the other hand, based on a qualitative study in the Scottish Highlands and Islands, Atterton (2007) found that strong ties negatively affect future business development. Following the literature, the key assumption in this research is that network centrality indicators, such as degree and closeness, capture the transmission of knowledge and innovation among entrepreneurs who interact within a network. Hence, it is expected to find a positive effect of network ties on the performance of the small rural firms. To analyze the effect of network ties, this research lies on the framework developed by Brueckner (2006; p.851) that 47

61 models two types of ties: direct and indirect, which positively depend on the effort of each agent involved in the interaction. This effort may reflect the participation of the members in the cluster, such as to serve on a project team or committee. It is assumed that the probability P of building a linkage among entrepreneurs within a network positively depends on their effort e, with 0 P 1; a higher effort implies more ties. For instance, a member who participates in a project of the cluster is more likely to create more ties with other entrepreneurs. The expected benefit from creating ties is expressed as follows: ij, e ji * uij ihpe jh ehj Ce ij B i P e, j hi j where B is negative if the costs C i j e ij are bigger than the expected benefits of the direct u Pe, e plus the indirect P e e * Pe, e ij j ij ji j the direct and indirect benefits for the entrepreneur i, respectively. ij, ji ih jh hj ties, u ij 0 and ih 0 are hi The entrepreneur i will choose effort level e il that maximizes his expected benefits. The first-order condition states that: Bi e il P eil, e e il li u il hi P e The effect of direct ties is captured by ih, e Ce lh hl il P eil, e e il li * u il, and the effect of indirect ties is captured by P eil, e e il li * hi Pe, e. If the marginal cost ih lh hl C exceeds e il 48

62 , e Pe, e P eil e il li * u il il e il li * hi Bi Bi effort will be 0 ; otherwise > 0. e e il Pe, e then the benefits of building more ties through ih il lh hl 3.3. Data and variable definitions A self-completion survey questionnaire was mailed to members of five clusters in the Northeast of the United States from which a total of 105 usable answers were received. Cluster participants were asked about their experiences in the cluster activities, assessment of the cluster, network ties with other members within the cluster, benefits of the cluster, and socio-economic characteristics Dependent variables: Economic performance The dependent variable to assess business success is based on the members perception of the cluster s effect on their businesses. Previous studies such as Kale et al. (2002) and Walter et al. (2006) have shown that subjective and objective data of firm performance produce similar results; hence perception is a good proxy to measure economic success. There are 3 relevant questions regarding participation in the cluster: 1) Makes your business more profitable, with 5- point Likert scales responses grouped into three categories: agree (strongly agree and agree), disagree (strongly disagree and disagree), and unsure; 2) Percentage change of volume of products sold; and 3) Percentage change of net farm income. The latter two dependent variables are based on questions with three categories as responses including a quantitative response, as well. The three categories are: i) Has not made any difference to my total volume of products sold (net farm income) which is considered as a 49

63 change of 0%. ii) Has contributed to a decrease in my total volume of products sold (net farm income). Since I began participating in the cluster my total volume of products sold (net farm income) has decreased by roughly %. And iii) Has contributed to an increase in my total volume of products sold (net farm income). Since I began participating in the cluster my total volume of products sold (net farm income) has increased by roughly %. These three responses are combined into one continuous response; thus, dependent variables 2) and 3) are treated as continuous variables rather than categorical variables. On the other hand, question 1 is treated as a categorical variable with three outcomes or categories. The majority of the cluster s member responses for question 1 are skewed toward positive outcomes, even no members of G5 has a negative outcome (i.e., disagree with question makes your business more profitable ). Therefore, based on the descriptive analysis, I decided to employ the percentage change of products sold (question 2) and net income (question 3) as dependent variables. Although questions 2 and 3 have only few cases of a negative percentage change, the responses for those cases exhibit a broad percentage variation allowing a better analysis of the economic impact for the cluster s members Independent variables The variables included in the econometric analysis are divided into three groups: 1) individual characteristics of the entrepreneur: age, age squared, education with three levels, years in the business, non-farm experience, and partner s job; 2) cluster characteristics: dichotomous variables for each cluster and organization type of cluster; and 3) network centrality indicators: degree, closeness, and network density. The correlation matrix (see Appendix Table 3.A), as expected, shows a high correlation (higher than 0.6) between age and age squared, among 50

64 education levels, between gender and cluster indicators since G3 has only female members, and among the centrality indicators. Table 3.1 presents the descriptive statistics for the variables included in the econometric analysis Social network indicators Three centrality network indicators are well-known in the literature degree, betweenness, and closeness to evaluate the importance of node or agent within the social structure or network. Freeman s betweenness captures the frequency of a node employed along the shortest paths to connect two other nodes and it is normalized by (n-1)*(n-2)/2. This indicator, however, does not take into account the directionality of the relationship. Two social networks built in each cluster are considered in the econometric analysis, one based on economic interactions, and the other one on social support interactions. Both networks are considered as advice networks since members involved are exchanging assistance and collaboration with other members similar to a horizontal network. To identify an economic network, the survey asks: Who are the individuals you most rely on to give you the information you need to manage your farm business? These are the people you go to for information about production, business management, and/or marketing your farm products, and the question for social support network is: Who are the people you most rely on to give you social support as a farmer? These are the people who really understand what you are trying to do on your farm and support your efforts to achieve those goals. Thus, an entrepreneur of the cluster is considered as a member of the economic or social support network if he seeks or gives advice regarding management or social support, respectively. In the survey, the entrepreneur may list all of his/hers contacts without any limitation in the number of members mentioned. 51

65 The nature of the network ties, numbers and types, created among cluster members are relevant for the knowledge flow (Koka & Prescott, 2002). Two centrality indicators, degree and closeness, are employed to capture ties among members within each network. The UCINET software is used to calculate the centrality indicators. Centrality indicators capture the power or popularity of each member within the network; these metrics indicate that a central location allows faster and broader access to information (Oh, Chung & Labianca, 2004). In the case of an advice network, central members are more likely to accumulate knowledge and expertise related to their business from other members (Baldwin eta al., 1997; Sparrowe et al., 2001). Therefore, it is expected that more benefits are acquired by the member with a more central position within the network. The number of network ties is captured by the degree centrality that counts the number of direct ties among members within each cluster. A high degree centrality means that a member is well-connected and has access to knowledge through several direct sources (i.e., other members). There are two degree indicators, out-degree that measures centrality and influential position, and in-degree that measures prestige. In the literature and in the software UCINET, the nature of the network is similar to a friendship relationship (i.e., give friendship), and out-degree counts the ties of a given actor directed to other actors within the network, that is, the choices (or friends) made by an actor. On the other hand, in-degree measures the popularity or prestige; a member with a high in-degree is that who is selected as a friend more times (i.e., receiving friendship from others). The nature of the interactions in this research, however, is an advice relationship (i.e., receive advice) where a given actor asks for advice to another, that is, receive an advice or information. For instance, if member A asks for advice to B and C (i.e., ties to B and C), out- 52

66 degree for A is equal to 2 but it means that A receives advice from two sources (B and C), whereas in-degree for B is equal to 1 but it means that B sends his advice just to one member (A). Thus, in this research in-degree measures the centrality and influence position of a member within the network, whereas out-degree measures the prestige or advice/information received. Since this research is focused on the individual benefits for the members rather than the collective benefits for the cluster, out-degree is preferred to in-degree to measure the impact of information received. The formula to measure the standardized 19 out-degree of member i is: d( i) i x ij n 1, where x ij is the ties from member i to member j, and n is the number of members in the network (Wasserman & Faust, 1994). Closeness centrality is related to the type of network ties, and it can be interpreted as indirect ties acquired via a sequence of links beyond the immediate ties within the network. Martinez & Aldrich (2011, p.26) state that indirect ties link entrepreneurs to people who are otherwise mostly invisible to them. Closeness centrality calculates the inverse of the shortest distance from one node to all the other nodes. The formula is: n 1 CC( i) n, which d i, j j1 represents the normalized inverse of the sum of the distance from member i to all the other members, which is converted to a scale by: CC ( i) CC ( i)* 100. A higher closeness centrality index indicates that a member has a better location within the network to quickly reach new information or knowledge from different members without having to depend on other members (Wasserman & Faust, 1994), that is, an entrepreneur with more indirect ties may have easy access to broader resources and innovation ideas (Baldwin et al., 1997; Martinez & Aldrich, 19 The standardized measure is employed to control for network size. 53

67 2011). Due to the advice nature of the relationship in this research, out-closeness will measure the ability of a member to acquire advice or information from other members within the network. However, due to the lack of a good instrument to address the endogeneity problem for outcloseness, the average of in- and out-closeness is included in the econometric analysis, and this may be interpreted as a position of advantage within the network Network density and cluster organization Two group indicators are considered in the analysis: density and organization type. Density is a group-level indicator related to social network theory whereas organization type refers to the business nature of the cluster considering three types: commodity, organic, and ethnic/cultural 20. For this research, network density is the ratio between existing ties and all the possible ties within a network for each cluster, and the standardized formula is: d( N) xij n n 1, that is, density measures the overall interaction within the network (Sparrowe et al., 2001). A network with high density represents a strong network whereas a weak network has a low density score. A strong network experiences cohesive social ties that facilitate the exchange of information among members because knowledge flows faster and with less distortion (Coleman, 1988; Kenis & Knoke, 2002). However, there are potential negative effects from a high density level; for instance, redundant and obsolete knowledge can be easily transmitted (Koka & Prescott, 2002), and it can limit the access to diverse information (Martinez & Aldrich, 2011). Two density measures based on economic and social support relationships are employed in the analysis. The UCINET software is employed to calculate density, and NETDRAW to 20 Ethnic/cultural is used as the reference category in the econometric analysis. 54

68 graph the ties within each of the five clusters analyzed. According to the graphs in Figure 3.1, clusters G2 and G4 are strong networks with a scale-free network structure whereas clusters G1 and G3 show a mix structure with sub-graphs in both cases, and a scale-free structure in G1 and a star structure in G3. Finally, cluster G5 shows a scale-free structure for the economic network, and a mix between sub-graphs and star structure for the social support network. Csermely (2009) points out that the topology of a network evolves from a strong networks represented by random graphs with more resources and fewer stress levels toward weak networks represented by sub-graphs with fewer resources and higher stress levels. Hence, the clusters analyzed in this research show strong and weak structures with G2, G4 and economic network of G5 showing a strong structure, G1 and social support network of G5 showing a mixed structure, and G3 with a weak structure Hypotheses Three hypotheses are proposed in this research. The first and second hypotheses concern the type of relationships created within the cluster whereas the third one relates to the nature of the cluster. 1. Prestige of a member within the network measured by the standardized out-degree as a mechanism to receive knowledge and information through direct ties will have a positive effect on the business success of the small rural farmers. 2. The member s location within the network measured by closeness as a tool to reach knowledge and information from several sources beyond the immediate ones (i.e., indirect ties) will have a positive effect on the business success of the small rural farmers. 55

69 3. Networks characteristics (i.e., weak or strong networks) measured by the standardized density will have a significant and negative impact on business success of small rural farmers. As discussed in the previous section, a high density degree may lead to redundant information flowing in the network and restrict the entry of new members Research methods The econometric analysis poses two problems: endogeneity and selection biases. To address the selection bias problem, the statistical procedure is to use a Heckman selection model; and instrumental variables (IV) are employed to address the causality (i.e., endogeneity problem). Thus, the analysis needs to address a sample selection bias in the context of IV-estimates. A two-stage model with instrumental variables is undertaken to address the hypotheses proposed, and robust variance estimators were included to address heteroskedasticity problems 21. The selection problem arises when the dependent variable is observed only for a restricted and non-random sample. In this research, 367 small rural farmers were interviewed but only 124 provided information for the dependent variables (i.e., % change of products sold and % change of net income). Hence, the % change is only observed for those who provide information, and a probit model is used for modeling the decision to provide or not information. Similar to social capital studies, the main concern in the econometric analysis is the direction of the causality. An endogeneity problem arises in the estimation of the effect of social network ties (i.e., out-degree and closeness) included in that is, out-degree and closeness are correlated with the error term N i (see eq. 3.1.) on business success, i. This correlation is explained by the reverse causality between social ties and business success. An entrepreneur 21 The Huber-White sandwich estimators were employed. 56

70 who has a successful business is more likely to have more social ties since other entrepreneurs will seek for his/her help and advice, and in the other hand, having more social ties will affect positively the business performance as proposed in the hypotheses of this research. Two models of business success are estimated by the following general model: BS * i 0 1Ii 2Ci 3 N i i (eq. 3.1.) where * BS i is business success (percentage of volume of products sold and percentage of net farm income), I i is the vector of individual characteristics, characteristics, and N i is the vector of centrality network indicators. C i is the vector of cluster The methodology described by Wooldridge (2010, ch. 19) and Semykina & Wooldridge (2010) is followed to estimate equation 3.1. Assume that information with the following reduced form of the selection equation: * s i is the propensity to provide s * i 0 1Ii 2Ci 3Ei 4 IV i i (eq. 3.2.) and, BS BSi NA * * if si > 0 otherwise where and E i is the interaction term between age and gender that represents the exclusion restriction, IV i is the instrumental variable represented by served on a project team or committee. IV is included in the equation 3.2 since it is not assumed that IV is exogenous to the decision of providing information. The Inverse Mills Ratio (IMR) is calculated from equation 3.2 which is included to estimate the fitted values from the first stage regression of the endogenous variables on IV. To address the endogeneity problem, I use the instrumental variables method to estimate the fitted values. The IV reflects the social interaction of the entrepreneurs within the cluster 57

71 which in turn will positively affect the likelihood of increasing the number of contacts, but it is expected that those variables will not affect the dependent variables measuring business success since social attitudes are not necessarily related to business management. To test the validity of the instruments, I estimate a tobit model for out-degree and closeness in the first stage. The rule of thumb is that if the F-test > 10 then the IV employed is a strong instrument. The tobit model is employed rather than a linear model because out-degree and closeness are left censored variables with no negative values. Additionally, I run a 2SLS using ivreg2 in STATA to take advantage of the tests associated with the linear regression with instrumental variables to test their validity. The ivreg2 provides an under-identification test and weak identification Stock- Yogo (SY) test for instrumental variables (rule of thumb: SY > 10%), and an exogeneity test for the model. Finally, the primary equation is estimated with a linear regression model including the fitted values and the IMR as follow: BS * i Ii Ci Nˆ i 4 IMR i where Nˆ i is the fitted values from the first stage for out-degree and closeness Results Descriptive results The first analysis is based on UCINET and NETDRAW software which provides a graphic instrument to represent the relationship between the dependent variables and the position of the cluster s member in the two networks analyzed. The figures show a square for each node (i.e., cluster member) that increases in size with respect to the dependent variables. For instance, a 58

72 higher percentage of products sold are represented with a bigger square. Finally, an econometric analysis is undertaken to test the three hypotheses proposed Dependent variable: Percentage of volume of products sold The descriptive results in Figure 3.2 show that within structurally strong network such as those of G2, G4 and economic network of G5 a better location of their members is relevant to report a higher percentage of volume of products sold and net income. Meanwhile, both networks of G1 and social support network of G5 with a less strong structure network show that benefits are spread among their member regardless of their location within the network. On the other hand, cluster members in G3 do not show a clear pattern: the node with a higher percentage change of products sold does not have the best location in either of the two networks. This may be the result of the topology of the G3 s networks that has a star shape mix with sub-graphs that contains two central agents and peripheral members who are not connected with any of those two central agents. A star network reflects the moderate rather than low costs of creating linkages (Brueckner, 2006) and consequently the difficult of reaping the potential benefits of network ties for their members, that is, the benefits are more likely to be concentrated on just few members. The graphs shed some light regarding the role of network density in the perception of business success for cluster s members since strong networks are related to denser ties. Additionally, for networks with extreme topology structure cases, strong or weak, the location of their members is important to determine their economic success. 59

73 Dependent variable: Percentage of net farm income Similar to the results showed in Figure 3.2, descriptive for % change on net income point out the relevance of a central location within networks with a strong structure. Networks of G2 and G4, and economic network of G5 showed in Figure 3.3 reveal that members with a central location report having a higher % change on net income. Likewise, networks of G3 show a similar pattern that that in Figure 3.2 where members with a central location do not report the highest benefits. On the other hand, members in networks of G1 and in social support network of G5 with a mixed topology structure show a different behavior than that on Figure 3.2. In this case, the benefits on net income are not perceived by all the cluster s members as well as location does not have a clear effect Econometric results Following the methodology explained in section 3.5, a selection equation is estimated to compute the IMR. Table 3.B (see Appendix) shows that the exclusion restriction, age*gender, is relevant to model the decision to provide information regarding the % change of products sold. Similarly, Table 3.C (see Appendix) shows that the exclusion restriction is relevant to explain the decision to provide information of % change of net income. The results show that there is a selection problem when the sample observed is that for % change of products sold except when closeness of social support network is included in the analysis. On the other hand, there is not a sample selection bias problem when the sample observed is % change of net income. The next step is to address the endogeneity problem by an IV method including the IMR from the selection equation. The instrumental variable employed is served on project team on 60

74 committee. Tables 3.D, 3.E, 3.F, and 3.G (see Appendix) show that the IV is significant and relevant to estimate the fitted values for the endogenous variables (i.e., out-degree and closeness). The first-stage is estimated with a Tobit model, and according to the rule of thumb (F>10) the IV is a strong instrument for both dependent variables when the effect of closeness of both networks and out-degree of economic network are analyzed but it is a weak instrument when out-degree of social support network is analyzed. These results are similar to those found with the Stock-Yogo (SY) test, IV is weak for both dependent variables when closeness and outdegree of social support network are analyzed but it is strong when closeness and out-degree of economic network is analyzed. Finally, the test of exogeneity (Hausman and Sargan-Hansen tests) states that there is an endogeneity problem in the estimation of % change of products sold and of net income, then a two stage method is followed Results for percentage change of products sold The first analysis includes only the individual and group variables. Table 3.2 shows that individual variables do not have a significant effect on % change of products sold, but education variables become significant when cluster variables are included. Having high school and college education with respect to some schooling has a positive and significant effect on products sold, and members in G5 report a higher % change in comparison to members in G3. The next two tables, 3.3 and 3.4, present the effect of out-degree and closeness on % change of products sold, respectively. Four equations were estimated for each table. The first two columns show the effect of each network indicator to test hypotheses 2 and 3 controlling by 61

75 cluster membership, and the last two columns includes density and network indicators controlling by type of cluster organization to test hypothesis 3. The out-degree of both networks, economic and social support, has a positive and significant effect on the perception of % change of products sold which support the hypothesis 1 proposed, that is, creating direct ties will increase products sold of cluster s members regardless if their ties are created within an economic or social support network. A unit increase in outdegree of economic network will increase % change of products sold by 26.3%, and of social support network will increase it by 37.2%. These results are consistent with Sparrowe et al. (2001) where in-degree of an advice network has a positive effect on individual performance, and Parida et al. (2010) where the number of direct ties within a horizontal network has a positive effect on small firms performance. Additionally, the variable years managed/own a farm has a positive effect (see first two columns of Table 3.3) which is expected since owning a farm will develop farm-business attitudes that makes a difference in the benefits perceived for cluster s members. Other relevant results show that belonging to clusters G1 and G4 has a negative impact in comparison to G3. This suggests that networks with a stronger structure play a negative role on the % change of products sold. To extent this analysis, the last two columns of Table 3.3 incorporates density measures for both networks and type of cluster s organization. When density is included, out-degree of both networks remains significant but years managed/own a farm is no longer significant. The analysis for the economic network (third column) shows that density is not significant, and belonging to a commodity cluster decreases the % change of products sold in comparison to those who belong to an ethnic/cultural cluster. On the other hand, the analysis for the social 62

76 support network (last column) shows that density has a significant and negative impact on the % change of products sold which supports the hypothesis 3 proposed. Likewise, belonging to an organic cluster will increase the % change of products sold with respect to those who belong to an ethnic/cultural cluster. A similar analysis was undertaken for closeness indicators in Table 3.4. The results provide evidence to support hypothesis 2 that indirect ties measured by closeness have a positive and significant effect on % change of products sold regardless the nature of the network. These results are consonant with those found by Baldwin et al. (1997) where closeness of communication network has a positive impact on students performance. Indirect ties give access to a broader range of contacts in comparison to direct ties that includes only similar peers; thus an entrepreneur with a higher closeness degree will be in a better position within the network to capture information and a broader market to benefit his/her small rural firm. Additionally, only spouse has a non-farm job has a significant but negative effect on products sold (second column) which is in line with the positive effect of years owning a farm. A spouse with no prior experience in farm jobs may negatively affect business performance. Likewise, belonging to G1, G2, G4, and G5 negatively affects the % change of products sold in comparison to membership in G3 which confirms the results previously found that structurally weak networks show a better economic performance. The last two columns in Table 3.4 include density and type of organization. Closeness remains significant and density has a negative effect on products sold which confirms the hypothesis 3. Percentage of products sold will decrease if a member belongs to a network with high density degree, that is, cohesive ties will decrease the quantity sold. A network with a high density restricts the influence of outsiders, which limits the market for the entrepreneur for 63

77 selling his/her products. Other results show that respect to ethnic/cultural clusters, belonging to a commodity cluster has a negative effect (third column) and to organic clusters has a positive effect (fourth column). Finally, spouse has a non-farm job and years managed/owned a farm have a significant impact on products sold (fourth column) Results for percentage change of net farm income The first analysis presented in Table 3.5 compares the outcomes from individual and group variables. The results show that years managed/owned a farm has a positive effect whereas spouse has a non-farm job has a negative effect on the dependent variable. After adding the group variables, education variables become significant and belonging to G5 with respect to G3 has a positive effect on net income. Four equations are presented in Tables 3.6 and 3.7 to test hypotheses 1 and 2. The first two columns in Table 3.6 show that out-degree has a significant and positive effect on % change of net income. The % change net income will increase by 25.7% and by 36.5% if the out-degree of economic network and of social support network increases by one unit, respectively. These results and those found for % change of products sold confirm hypothesis 1, cluster s members who invest in creating direct ties will obtain higher benefits. In the same vein with the results found for % change of products sold, Table 3.6 also shows that acquiring expertise on farm jobs measured by the variable years managed/owned a farm will increase the % change of net income, the lack of a farm job for the spouse and membership in G1 and G4 in comparison to G3 will negatively affect net income. In addition, there is a non linear relationship between net income and age of the small farmer. 64

78 The last two columns in Table 3.6 introduce density and type of organization. Unlike the results for % change of products sold, density has a negative affect regardless the nature of the network, % change of net income will decrease by 63.7% and by 107.1% if density of economic network and of social support network increases by one unit. These results confirm hypothesis 3, a structurally strong network is detrimental to obtain higher economic benefits. A network with a high density degree may incur in excess claims on group members (Portes, 1998) which may impose decisions based on social relations rather than on economic behavior. The results also show that belonging to an organic organization will increase net income in comparison to an ethnic/cultural cluster organization. The same approach previously explained is undertaken to present the analysis for closeness and density in Table 3.7. The first two columns show that closeness has a significant and positive effect on % change of net income which support hypothesis 2 proposed that indirect ties will benefit economic returns for small rural farmers. As it was pointed out in the literature section, through indirect network ties entrepreneurs may reach a broader set of information to improve their management skills and production techniques which in turn will increase their profit maximization and productivity. Other results show that a spouse without a farm job will negatively impact net income, and there is not a linear relationship between age and net income. Similar to previous results, clusters with stronger structure networks (i.e., G1, G2, G4, and G5) has a negative impact in comparison to G3 that has weaker structure networks. To complete the analysis and test hypothesis 3, last two columns include density and type of organization. Density has a negative effect regardless the nature of the network, net income will decrease by 465.8% and by 274.7% if density of economic network and of social support 65

79 network increases by one unit. Finally, an organic organization produces better results whereas a commodity organization will reduce economic benefits Conclusions The results of this research support the hypotheses proposed that centrality network indicators have a significant impact on a firm s performance, and they confirm the relevance of considering social network research in the analysis of small firms performance. Network indicators are found to have different impacts on business success. First to notice is that education variables are not longer significant when centrality network indicators are included in the analysis. The relevance of network ties is also confirmed by the individual analysis of each kind of centrality indicator, out-degree and closeness; both have a positive and significant effect on the economic performance of small rural firms. Creating direct and indirect ties in the economic and in the social support networks will improve % change of products sold and of net income. The significant effect of centrality network indicators (i.e., out-degree and closeness) remains significant even after including density degree of each network to capture the topology of the networks. Here, the nature of the network and the type of network ties become relevant to determine the effect of density degree. The analysis shows that strong ties (i.e., high density degree) within an economic network have no effect on the % change of products sold when outdegree is included in the econometric analysis. In the rest of the cases, the results show that a high density degree will reduce economic outcomes, which it is in line with the behavior of clusters. Those clusters that have structurally strong networks such as G1, G2, G4, and G5 produce lower economic benefits in comparison to structurally weak networks such as G3. A 66

80 network with a high density excludes outsiders and limits the market where to sell products or from where to get information which it may explain the negative results of density. Here, there is an important place for policy measures. Creating ties depends on private efforts, but governments and non-profit organizations can promote activities to encourage gatherings among entrepreneurs to develop informal contacts which can evolve into the creation of more formal networks (Malecki, 1993; Walter et al., 2006). Additionally, as is suggested in the literature and these results, direct and indirect ties have different effects on firm s performance; in particular, indirect ties provide access to broader information. In that sense, developing conferences to share experiences or bring experts will help to improve the knowledge flowing within the network and expand connections beyond direct ties. To conclude, the results reveal that centrality indicators have a positive effect on small firms performance but the nature of the network (i.e., economic or social support) and the characteristics of the network ties (i.e., direct or indirect) are relevant to understand the extent of this positive effect References Atterton, Jane (2007). The strength of weak ties: Social networking by business owner in the highlands and islands of Scotland. European Society for Rural Sociology, 47(3), pp Audretsch, David (2007). Entrepreneurship capital and economic growth. Oxford Review of Economic Policy, 23(1), pp Baldwin, T., M. Bedell & J. Johnson (1997). The social fabric of a team-base M.B.A. program: Network effects on student satisfaction and performance. Academy of Management Journal, 40(6), pp Barabasi, A.L. & R. Albert (1999). Emergence of scaling in random networks. Science, 286, pp

81 Barabasi, A.L., H. Jeong, Z. Neda, E. Ravasz, A. Schubert & T. Vicsek (2002). Evolution of the social network of scientific collaborations. Physica A: Statistical Mechanics and its Applications, 311(3-4), pp Barkley, D., F. DiFurio & J. Leatherman (2004). The role of a public venture capital program in state economic development: The case of Kansas Venture Capital, Inc. The Journal of Regional Analysis & Policy, 34(2), pp Borgatti, S., A. Mehra, D. Brass & G. Labianca (2009). Network analysis in the social sciences. Science, 323(5916), pp Borgatti, S. & P. Foster (2003). The network paradigm in organizational research: A review and typology. Journal of Management, 29(6), pp Brasier, K., S. Goetz, L. Smith, M. Ames, J. Green, T. Kelsey, A. Rangarajan & W. Whitmer (2007). Small farm clusters and pathways to rural community sustainability. Journal of Community Development, 38(3), pp Brueckner, Jan (2006). Friendship networks. Journal of Regional Science, 46(5), pp Csermely, Peter (2009). Weak links. The universal key to the stability of networks and complex systems. Heidelberg: Springer; p. Chell, E. & S. Baines (2000). Networking, entrepreneurship and microbusiness behavior. Entrepreneurship & Regional Development, 12, pp Coleman, James (1988). Social capital in the creation of human capital. The American Journal of Sociology, 94, pp. S95-S120. Das, R. & K. Das (2011). Industrial cluster: An approach for rural development in North East India. International Journal of Trade, 2(2), pp Goerzen, Anthony & Paul Beamish (2005). The effect of alliance network diversity on multinational enterprise performance. Strategic Management Journal, 26, pp Goetz, Stephan (2008). Self-employment in rural America: The new economic reality. Rural Realities, 2 (3), pp Granovetter, Mark (1985). Economic action and social structure: The problem of embeddedness. American Journal of Sociology, 91(3), pp Greve, A., M. Benassi & A. Dag-Sti (2010). Exploring the contributions of human and social capital to productivity. International Review of Sociology, 20(1), pp

82 Gronum, S., M. Verreynne & T. Kastelle (2012). The role of networks in small and mediumsized enterprise innovation and firm performance. Journal of Small Business Management, 50(2), pp Johannisson, Bengt (1986). Network strategies: Management technology for entrepreneurship and change. International Small Business Journal, 5(19), pp Kale, P., J. Dyer & H. Singh (2002). Alliance capability, stock market response, and long-term alliance success: The role of the alliance function. Strategic Management Journal, 23(8), pp Latora, V. & M. Marchiori (2001). Efficient behavior of small-world networks. Physical Review Letters, 87, Malecki, Edward (1993). Entrepreneurship in regional and local development. International Regional Science Review, 16 (1 & 2), pp Martinez, M. & H. Aldrich (2011). Networking strategies for entrepreneurs: balancing cohesion and diversity. International Journal of Entrepreneurial Behaviour & Research, 17(1), pp McQuaid, R.W. (1996). Social networks, entrepreneurship and regional development. In: M. Danson (ed), Small firm formation and regional economic development. Routledge, London. Nelson, Richard (1981). Research on productivity growth and differences: Dead ends and new departures. Journal of Economic Literature, 19(3), pp Newman, M.E.J (2001). Clustering and preferential attachment in growing networks. Physical Review E., 64, (R). Parida V., M. Westerber, H. Ylinenpaa & S. Roininen (2010). Exploring the effects of network configurations on entrepreneurial orientation and firm performance: An empirical study of new ventures and small firms. Annals of Innovation & Entrepreneurship, 1: 5601 DOI: /ale.v Porter, Michael (2000). Location, competition, and economic development: Local clusters in a global economy. Economic Development Quarterly, 14 (1), pp Portes, Alejandro (1998). Social capital: Its origins and applications in modern sociology. Annual Review of Sociology, 24, pp Putnam, Robert (2000). Bowling alone: The collapse and revival of American community. New York: Simon Schuster. Rocha, Hector (2004). Entrepreneurship and development: The role of clusters. Small Business Economics, 23(5), pp

83 Sandstrom, A. & L. Carlsson (2008). The performance of policy networks: The relation between network structure and network performance. The Policy Studies Journal, 36(4), pp Semykina, A. & Wooldridge J.M. (2010). Estimating panel data models in the presence of endogeneity and selection. Journal of Econometrics, 157, pp Sparrowe, R., R. Liden, S. Wayne & M. Kraimer (2001). Social networks and the performance of individuals and groups. The Academy of Management Journal, 44(2), pp Walter, A., M. Auer & T. Ritter (2006). The impact of network capabilities and entrepreneurial orientation on university spinoff performance. Journal of Business Venturing, 22(6), pp Wasserman, S. & K. Faust (1994). Social Network Analysis: Methods and Applications. Cambridge, ENG and New York: Cambridge University Press. Wooldridge, J.M. (2010). Econometric analysis of cross section and panel data. Cambridge, MA: The MIT Press, 2 nd edition. 70

84 Table 3.1. Descriptive statistics Obs Mean Std. Dev. Min Max Dependent variables Volume of products sold Net farm income Independent variables Economic network Out-degree Closeness Density Social support network Out-degree Closeness Density Individual Age Age squared ,400 Gender (1=male) Some schooling (reference) High school, vocational or some college College and post-graduate Years managed/owned a farm Non-farm job Spouse non-farm job Clusters G G G3 (reference) G G Commodity Organic Ethnic/cultural (reference) Instrumental variable Served on a project team or committee Exclusion restriction Age*gender

85 Table 3.2. Impact of individual and group variables on % change of products sold Primary equation OLS % change products sold Individual Age 6.45 (1.41) 5.82 (1.51) Age squared (-1.20) (-1.37) Gender (1=male) (-0.19) (-1.08) High school, vocational or some college (=1) 7.44 (0.33) 79.43** (2.60) College and post-graduate (=1) (1.53) *** (2.78) Years managed/owned a farm 0.81 (0.79) 1.33 (1.29) Entrepreneur has a non-farm job (0.74) 6.24 (0.29) Spouse has a non-farm job (-1.30) (-1.29) Cluster (ref=g3) G (-0.19) G (1.53) G (0.56) G *** (3.20) Constant (-1.48) ** (-2.13) Observations R-squared Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<

86 Table 3.3. Impact of out-degree and density on % change of products sold Primary equation OLS % change products sold Economic network Predicted out-degree 26.27** (2.48) 24.85** (2.36) Density (-1.32) Social support network Predicted out-degree 37.23** (2.48) 40.77*** (2.65) Density ** (-2.50) Individual Age 7.35 (1.27) (-1.03) 7.33 (1.19) (-1.20) Age squared (-1.16) 0.09 (1.19) (-1.05) 0.11 (1.33) Gender (1=male) (-1.04) (-0.32) (-0.64) (-0.04) High school, vocational or some college (=1) (-0.03) (0.73) 4.27 (0.10) (0.33) College and post-graduate (=1) (0.71) (0.90) (1.00) (0.66) Years managed/owned a farm 2.16* (1.72) 2.40* (1.82) 1.85 (1.56) 2.09 (1.64) Entrepreneur has a non-farm job (0.89) (1.11) (0.79) (1.58) Spouse has a non-farm job (-1.28) (-1.62) (-1.11) (-1.55) Cluster (ref=g3) G ** (-2.25) ** (-2.50) G (-1.05) (-1.46) G ** (-2.40) ** (-2.43) G (-0.46) (-0.41) Type of cluster (ref=ethnic/cultural identity) Commodity * (-1.82) (-1.19) Organic (0.95) *** (2.78) Constant (-1.41) (0.69) (-1.37) (1.01) IMR ** (2.24) ** (2.13) ** (2.23) * (1.83) Observations R-squared Underidentification test Ho. Eq. Is underidentified (Ha. Instruments are relevant) Weak identification test (>SY10%=strong) Stock-Yogo ID 10% Stock-Yogo ID 15% Exogeneity test Ho. Closeness is exogenous Excluded instrument Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< Identified Identified Identified Identified Strong instrument Weak instrument Strong instrument Weak instrument Endogenous Endogenous Endogenous Endogenous Served on project team Served on project team Served on project team Served on project team 73

87 Table 3.4. Impact of closeness and density on % change of products sold Primary equation OLS % change products sold Economic network Predicted closeness ** (2.48) ** (2.36) Density ** (-2.23) Social support network Predicted closeness ** (2.50) *** (2.79) Density *** (-2.73) Individual Age 8.66 (1.45) (-1.34) 6.59 (1.08) (-0.82) Age squared (-1.30) 0.09 (1.42) (-0.94) 0.04 (0.70) Gender (1=male) (-1.30) (-1.30) (-1.35) (-0.98) High school, vocational or some college (=1) (0.41) (-1.19) 2.75 (0.06) 0.69 (0.03) College and post-graduate (=1) (0.98) (-0.39) (0.34) (0.51) Years managed/owned a farm 0.05 (0.05) 0.56 (0.57) (-0.01) 2.14* (1.71) Entrepreneur has a non-farm job 1.50 (0.07) (1.17) (0.43) (0.77) Spouse has a non-farm job (-1.02) ** (-2.04) (-1.06) * (-1.99) Cluster (ref=g3) G ** (-2.54) ** (-2.45) G ** (-2.20) ** (-2.22) G4-1,386.5** (-2.47) -1,666.1** (-2.45) G ** (-2.18) -1,002.7** (-2.16) Type of cluster (ref=ethnic/cultural identity) Commodity ** (-1.99) (-1.33) Organic ** (2.30) ** (2.37) Constant *** (-2.64) * (-1.74) ** (-2.47) (-0.69) IMR ** (2.51) ** (2.12) Observations R-squared Underidentification test Ho. Eq. Is underidentified (Ha. Instruments are relevant) Weak identification test (>SY10%=strong) Stock-Yogo ID 10% Stock-Yogo ID 15% Exogeneity test Ho. Closeness is exogenous Excluded instrument Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< Identified Strong instrument Endogenous Identified Weak instrument Endogenous Identified Weak instrument Endogenous Identified Weak instrument Endogenous Served on project team Served on project team Served on project team Served on project team 74

88 Table 3.5. Impact of individual and group variables on % change of net farm income Primary equation OLS % change net farm income Individual Age 3.47 (1.20) (1.06) Age squared (-1.48) (-1.25) Gender (1=male) 5.81 (0.30) (-0.06) High school, vocational or some college (=1) (0.97) ** (2.31) College and post-graduate (=1) (1.66) ** (2.57) Years managed/owned a farm 1.36* (1.73) 1.422* (1.84) Entrepreneur has a non-farm job 6.21 (0.37) (-0.14) Spouse has a non-farm job * (-1.79) (-1.58) Cluster (ref=g3) G (-0.94) G (1.48) G (-0.07) G ** (2.37) Constant (-1.03) (-1.61) Observations R-squared Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<

89 Table 3.6. Impact of out-degree and density on % change of net farm income Primary equation OLS % change net farm income Economic network Predicted out-degree ** (2.52) ** (2.52) Density * (-1.89) Social support network Predicted out-degree ** (2.54) *** (2.73) Density ** (-2.52) Individual Age (1.62) (-1.52) (1.59) (-1.63) Age squared * (-1.90) (1.52) * (-1.85) (1.62) Gender (1=male) (-0.24) (0.63) (-0.06) (0.73) High school, vocational or some college (=1) (-0.89) (-0.28) (-0.89) (-0.32) College and post-graduate (=1) (-0.49) (-0.29) (-0.42) (-0.34) Years managed/owned a farm 2.724*** (2.76) 2.949*** (2.84) 2.620*** (2.71) 2.944*** (2.81) Entrepreneur has a non-farm job (1.21) (1.48) (1.21) (1.52) Spouse has a non-farm job * (-1.91) ** (-2.42) * (-1.88) ** (-2.42) Cluster (ref=g3) G * (-1.96) ** (-2.35) G (-0.96) (-1.53) G ** (-2.28) ** (-2.38) G (-1.06) (-1.04) Type of cluster (ref=ethnic/cultural identity) Commodity (-1.16) (-1.31) Organic * (1.95) *** (2.64) Constant (-0.92) (1.51) (-0.74) * (1.69) Observations R-squared Underidentification test Ho. Eq. Is underidentified (Ha. Instruments are relevant) Weak identification test (>SY10%=strong) Stock-Yogo ID 10% Stock-Yogo ID 15% Exogeneity test Ho. Closeness is exogenous Excluded instrument Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< Identified Strong instrument Endogenous Identified Weak instrument Endogenous Identified Endogenous Identified Served on project team Served on project team Served on project team Served on project team Strong instrument Weak instrument Endogenous 76

90 Table 3.7. Impact of closeness and density on % change of net farm income Primary equation OLS % change net farm income Economic network Predicted closeness ** (2.52) ** (2.49) Density ** (-2.40) Social support network Predicted closeness ** (2.53) *** (2.75) Density *** (-2.70) Individual Age 7.43* (1.94) (-1.51) 5.21 (1.53) (-0.81) Age squared -0.08** (-2.05) 0.07 (1.36) -0.06* (-1.79) 0.01 (0.15) Gender (1=male) (-0.46) 2.67 (0.12) (-1.01) (0.61) High school, vocational or some college (=1) (-0.41) (-1.39) (-0.85) 2.24 (0.07) College and post-graduate (=1) (-0.59) (-1.09) (-0.92) (-0.20) Years managed/owned a farm 0.24 (0.27) 0.79 (1.08) 0.69 (0.97) 2.76*** (2.79) Entrepreneur has a non-farm job 7.79 (0.46) (1.32) (0.70) 3.66 (0.21) Spouse has a non-farm job ** (-2.05) *** (-2.83) * (-1.73) *** (-2.74) Cluster (ref=g3) G ** (-2.48) ** (-2.50) G ** (-2.33) ** (-2.38) G4-1,550.7** (-2.49) -1,767.4** (-2.50) G5-1,133.9** (-2.33) -1,166.0** (-2.35) Type of cluster (ref=ethnic/cultural identity) Commodity (-1.53) ** (-2.08) Organic ** (2.57) (0.66) Constant *** (-2.79) ** (-2.03) *** (-2.73) (-0.48) IMR * (1.71) Observations R-squared Underidentification test Ho. Eq. Is underidentified (Ha. Instruments are relevant) Weak identification test (>SY10%=strong) Stock-Yogo ID 10% Stock-Yogo ID 15% Exogeneity test Ho. Closeness is exogenous Excluded instrument Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< Identified Strong instrument Endogenous Served on project team Identified Weak instrument Endogenous Served on project team Identified Weak instrument Endogenous Served on project team Identified Weak instrument Endogenous Served on project team 77

91 Figure 3.1. Network ties within clusters N1 Who are the people you rely on to: Manage your farm business N2 Who are the people you rely on to: Social support as farmer 78

92 79

93 Figure 3.2. Network ties within clusters by percentage of volume of products sold N1 Who are the people you rely on to: Manage your farm business N2 Who are the people you rely on to: Social support as farmer 80

94 81

95 Figure 3.3. Network ties within clusters by percentage of net farm income N1 Who are the people you rely on to: Manage your farm business N2 Who are the people you rely on to: Social support as farmer 82

96 83

97 Appendix B Table 3.A. Correlation matrix Dep. Var. Indep. Var. Individual Cluster Economic network Social support network Volume Net of farm product income s sold Age Gender Age (1=mal squared e) Some schoolin g High school or some college Years manage d a farm College and postgraduat Nonfarm job Spouse nonfarm job G1 G2 G3 G4 G5 Commo Organic Ethnic dity /cultural Outdegree Closene Density ss Outdegree Closene Density ss IV Exc. Served as Age*ge project nder team Dep. Var. Volume of products sold 1 Net farm income Indep. Var. Individual Age Age squared Gender (1=male) Some schooling High school or some college College and postgraduate Years managed a farm Non-farm job Spouse non-farm job Cluster G G G G G Commodity Organic Ethnic/cultural Economic network Out-degree Closeness Density Social support network Out-degree Closeness Density Instrumental variables Served as project team Exclusion restriction Age*gender

98 Table 3.B. Selection equation for % change of products sold Selection equations - PROBIT Provide information about % change products sold Individual Group Network ind. Density and network ind. Individual Age (-0.68) (-1.20) (-0.62) (-0.59) (-0.67) Age squared (-0.73) (-0.83) (-1.24) (-1.26) (-1.24) Gender (1=male) *** *** ** ** ** (-2.70) (-2.60) (-2.40) (-2.50) (-2.47) High school, vocational or some college (=1) (1.18) (1.01) (0.90) (0.90) (0.93) College and post-graduate (=1) (1.25) (1.05) (0.89) (0.90) (0.89) Years managed/owned a farm (-0.29) (-0.24) (-0.49) (-0.48) (-0.47) Entrepreneur has a non-farm job (-0.58) (-0.38) (-0.22) (-0.24) (-0.23) Spouse has a non-farm job (0.70) (0.47) (0.39) (0.40) (0.37) Cluster G *** *** (-6.78) (-7.03) G *** *** (-6.98) (-7.15) G *** *** (-6.76) (-5.70) G *** *** (-7.23) (-7.36) Type of cluster Commodity *** *** (-6.48) (-5.34) Organic *** *** (-5.76) (-5.14) Economic network Density (-0.47) Social support network Density (-0.43) Instrumental variables Served on project team or committee (0.10) (0.10) (0.04) Exclusion restriction Age*gender 0.119*** 0.209*** 0.215*** 0.209*** 0.219*** (2.81) (2.80) (2.61) (2.70) (2.67) Constant 4.987* *** ** *** *** (1.74) (3.08) (2.56) (2.67) (2.65) Observations Robust z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 85

99 Table 3.C. Selection equation for % change of net farm income Selection equations - PROBIT Provide information about % change net farm income Individual Group Network ind. Density and network ind. Individual Age (0.50) (0.50) (0.50) Age squared (-0.19) (0.04) (-1.13) (-1.13) (-1.17) Gender (1=male) (-0.94) (-0.96) ** ** ** *** ** (-2.42) (-2.43) (-2.46) High school, vocational or some college (=1) (-2.75) (-2.36) (0.83) (0.82) (0.86) College and post-graduate (=1) (1.24) (0.97) (0.06) (0.05) (0.04) Years managed/owned a farm (0.72) (0.36) (-0.40) (-0.40) (-0.33) Entrepreneur has a non-farm job (-0.15) (-0.06) (0.43) (0.43) (0.40) Spouse has a non-farm job (0.33) (0.32) (-0.55) (-0.55) (-0.59) Cluster (-0.38) (-0.69) G (-0.95) G2 (-1.20) (-1.33) G4 (-1.55) (-1.21) G5 (-1.10) * * (-1.71) Type of cluster (-1.85) Commodity (-0.60) (-0.39) Organic (-0.48) (-0.52) Economic network Density (-0.64) Social support network Density (-0.34) Instrumental variables Served on project team or committee (0.56) (0.58) (0.41) Exclusion restriction Age*gender 0.098*** 0.085*** 0.083*** 0.083*** 0.085*** (2.84) (2.64) (2.59) (2.59) (2.63) Constant (1.39) (1.33) (0.87) (0.91) (0.93) Observations Robust z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 86

100 Table 3.D. First stage TOBIT for endogenous variable out-degree (for products sold) First stage - TOBIT Economic Network For % change products sold Social Economic Network Network Social Network Individual Age (-0.25) (1.51) (-0.20) (1.53) Age squared * * (0.22) (-1.67) (0.17) (-1.68) Gender (1=male) (-0.49) (-1.00) (-0.43) (-1.02) High school, vocational or some college (=1) 3.589*** *** (3.14) (1.55) (3.22) (1.53) College and post-graduate (=1) 3.523*** 2.295* 3.658*** 2.359** (2.91) (1.93) (3.18) (1.99) Years managed/owned a farm (-0.60) (-0.56) (-0.64) (-0.68) Entrepreneur has a non-farm job (-0.85) (-0.84) (-0.94) (-0.81) Spouse has a non-farm job (0.46) (0.74) (0.48) (0.77) Cluster G * 4.833*** (1.86) (3.19) G *** 4.376*** (2.73) (2.94) G *** 7.561*** (3.07) (3.85) G *** 4.617*** (4.14) (3.28) Type of cluster Commodity (0.50) (1.05) Organic (0.05) (0.49) IMR (-0.82) (-0.43) (-0.73) (-0.39) Instrumental variables Served on project team and committee 3.962*** 2.795** 3.933*** 2.912** (3.66) (2.10) (3.58) (2.35) Economic network Density 2.073** (2.43) Social support network Density (1.46) Constant * * (-0.45) (-1.76) (-0.60) (-1.85) F-test Observations Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 87

101 Table 3.E. First stage TOBIT for endogenous variable closeness (for products sold) First stage - TOBIT Economic Network For % change products sold Social Economic Network Network Social Network Individual Age * (-0.62) (1.75) (-0.06) (1.37) Age squared (0.49) (-1.65) (0.02) (-1.19) Gender (1=male) (-0.26) (-0.16) (0.39) (-0.27) High school, vocational or some college (=1) 0.115** 0.162*** 0.132* (2.03) (3.24) (1.97) (1.52) College and post-graduate (=1) 0.124** 0.147*** 0.171** 0.200** (2.04) (2.93) (2.55) (1.99) Years managed/owned a farm 0.002* (1.85) (0.81) (1.39) (-0.88) Entrepreneur has a non-farm job (0.27) (-0.69) (-0.34) (0.02) Spouse has a non-farm job * ** (0.11) (1.88) (0.37) (2.09) Cluster G *** 0.899*** (16.60) (15.92) G *** 0.911*** (20.43) (17.01) G *** 1.865*** (22.81) (19.56) G *** 1.370*** (40.68) (26.89) Type of cluster Commodity * (0.79) (1.79) Organic *** (-4.70) (1.00) IMR * ** (-1.84) (0.79) (-0.49) (2.44) Instrumental variables Served as project team 0.157*** 0.123*** 0.141*** 0.241*** (5.26) (3.50) (3.26) (3.64) Economic network Density 0.647*** (17.77) Social support network Density 0.501*** (5.89) Constant 0.526*** * (3.40) (0.11) (1.74) (-0.80) F-test Observations Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 88

102 Table 3.F. First stage TOBIT for endogenous variable out-degree (for net income) First stage - TOBIT Economic Network For % change net farm income Social Economic Network Network Social Network Individual Age (-0.17) (1.60) (-0.13) (1.61) Age squared * * (0.12) (-1.78) (0.08) (-1.79) Gender (1=male) (-0.26) (-0.73) (-0.22) (-0.77) High school, vocational or some college (=1) 3.764*** 2.074* 3.806*** 2.049* (3.23) (1.73) (3.30) (1.70) College and post-graduate (=1) 3.695*** 2.398** 3.796*** 2.435** (3.08) (2.06) (3.30) (2.09) Years managed/owned a farm (-0.70) (-0.68) (-0.73) (-0.75) Entrepreneur has a non-farm job (-0.93) (-0.80) (-0.99) (-0.79) Spouse has a non-farm job (0.64) (0.78) (0.64) (0.79) Cluster G *** (1.65) (2.81) G ** 3.757** (2.47) (2.38) G *** 6.797*** (2.76) (3.23) G *** 3.757** (3.61) (2.28) Type of cluster Commodity (0.32) (0.88) Organic (-0.06) (0.28) IMR (-0.22) (0.42) (-0.14) (0.39) Instrumental variables Served as project team 4.034*** 3.007** 4.022*** 3.052** (3.59) (2.15) (3.54) (2.38) Economic network Density 2.012** (2.28) Social support network Density (1.39) Constant * * (-0.54) (-1.85) (-0.68) (-1.92) F-test Observations Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 89

103 Table 3.G. First stage TOBIT for endogenous variable closeness (for net income) First stage - TOBIT Economic Network For % change net farm income Social Economic Network Network Social Network Individual Age * (-0.56) (1.67) (-0.07) (1.25) Age squared (0.36) (-1.56) (0.00) (-1.02) Gender (1=male) (-0.01) (-0.38) (0.40) (-0.61) High school, vocational or some college (=1) 0.110* 0.156*** 0.127* (1.86) (3.07) (1.88) (1.40) College and post-graduate (=1) 0.134** 0.143*** 0.174** 0.183* (2.23) (2.89) (2.59) (1.80) Years managed/owned a farm 0.002* (1.87) (0.88) (1.44) (-0.82) Entrepreneur has a non-farm job (-0.06) (-0.66) (-0.47) (0.22) Spouse has a non-farm job * * (0.59) (1.74) (0.50) (1.71) Cluster G *** 0.914*** (16.48) (16.93) G *** 0.925*** (18.78) (17.52) G *** 1.883*** (22.37) (19.35) G *** 1.389*** (36.26) (25.85) Type of cluster Commodity ** (0.84) (2.15) Organic *** (-4.70) (1.22) IMR (-1.63) (0.11) (-0.76) (1.18) Instrumental variables Served as project team 0.148*** 0.120*** 0.136*** 0.244*** (4.95) (3.27) (3.21) (3.64) Economic network Density 0.650*** (17.59) Social support network Density 0.495*** (5.82) Constant 0.534*** * (3.45) (0.19) (1.88) (-0.71) F-test Observations Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 90

104 Chapter 4 THIRD ESSAY Population movement across counties and its effect on social capital 4.1. Introduction The term social capital is not new in the social sciences; sociological, political, and economic studies have used many different definitions and considered several effects of this form of capital on society. Schafft & Brown (2003) report that there was an increase of articles related to social capital from 2 or 3 in 1990 to more than 200 in However, there is no a consensus in the literature for a unique definition of social capital, which makes it difficult to establish its components. This research contributes to the literature on social capital by incorporating new factors as its determinants, such as population movement variables with a spatial econometric analysis which has rarely if ever been used in previous studies of this subject. This study incorporates not only migration and commuting flows but also commuting time and network centrality indicators to analyze the total effect of population movement on social capital. Despite the lack of consensus on the definition of social capital, its positive impact on economic outcomes such as on income growth, employment rate, entrepreneurship, educational attainment, and others is well documented (Bourdieu, 1986; Callois & Schmitt, 2009; Israel & Beaulieu, 2004; Levitte, 2004; Putnam, 2000). By creating social capital, members of a community are building a facilitating pathway by reducing transaction costs for enhancing economic development and human capital (Coleman, 1988; Dinda, 2008; Glaeser et al., 2002; Hanifan, 1916; Iyer et al., 2005; Rupasingha et al., 2006; Wallis et al., 2004). Hence, the study 91

105 of social capital is important as an instrument to reach a desirable outcome in economic terms 22. But it is not enough to state that increasing social capital will provide economic benefits. Social capital is not a solution for every social problem (Portes & Landolt, 2000); the analysis needs to incorporate why social capital has an effect. To accomplish this, it is important to understand where social capital comes from, and what factors or variables facilitate building social capital, if the goal is to delineate policy measures that may influence this form of capital. Social capital is a broad concept which refers to social relations among similar people who live in the same geographical area such as neighborhoods or communities; here it is known as bonding social capital, although its definition may be extended to social relations across distant geographical areas (e.g., counties) or across different ethnicities and socio-economic status, where it is known as bridging social capital (Flora & Flora, 2008; Iyer et al., 2005; Wallis et al., 2004). Social capital relations are based on the engagement of people who commit their time, their knowledge, and their effort to build community and civic organizations for meeting common objectives within a given area (Coleman, 1988; Putnam, 1995). For this reason, social capital is often identified as a place-specific characteristic (Glaeser & Redlick, 2009). Gertler (2000) points out that a complete analysis of social capital has to explain the determinants of social capital levels among regions and nations. Therefore, it is important to understand the creation of social capital not only at the individual but also at the aggregate level, and how factors such as migration and commuting may affect it. Migration and commuting affect social capital through changes produced on population characteristics, which are the main underlying inputs into producing social capital. In the short run, an immigrant is seen as a newcomer who may not share the social norms of the receiving 22 Portes & Sensenbrenner (1993) contend that social capital is generated by individuals with the anticipation of utilities associated with good standing in a particular collectivity (p. 1325). Hence, social capital is used as an instrument to obtain a reward for those who participate in the creation of social capital. 92

106 area; thus the trust required to produce social capital is broken. However, this immigrant may have some characteristics (learned in his/her origin area) that promote the production of social capital. On the other hand, commuting reduces available time to allocate to the production of social capital (Glaeser et al., 2002); but similar to migrants, commuters may take their knowledge from the residence (job) area to the job (residence) area to build social capital (Goetz et al., 2010). To determine the effect of migration and commuting variables on social capital it is necessary to address a causality problem (i.e., endogeneity) between them and social capital. The study of social capital poses an endogeneity problem since its causes and outcomes are not clearly separated, similar to other social processes, and it is not an easy task to state what comes first. For instance, in education research, there exists a reverse causality between income and educational attainment: a higher income has a positive effect on education but in turn education is a determinant of income level. Similarly, social capital may be the result of factors such as poverty rather than a cause of poverty. This research relies on lagged values to assess a causal relationship between social capital and the independent variables proposed in the analysis. Although previous studies have considered migration to explain social capital, they have not incorporated commuting or commuting time and only few have used a spatial econometric approach. Moreover, some empirical studies have failed to address the endogeneity problem of social capital. This research seeks to analyze the joint role of migration and commuting in the formation of social capital at the county level within the U.S. by using three econometric approaches, a linear regression model, a panel data analysis, and a spatial econometric model. Another important contribution is to incorporate network centrality indicators (i.e., closeness) in the analysis to determine the effect of migration and commuting beyond the number of people moving across counties. 93

107 The remainder of this research is structured as follows: a brief literature review of social capital and social networks is presented in section 2; the social capital model is described in section 3. Section 4 includes the data employed in the econometrics. Hypotheses are presented in section 5. The estimation strategy underlying the three econometric models and the endogeneity problem is explained in section 6. Results and conclusions are discussed in section 7 and 8, respectively Literature review Social capital The definition of social capital as coined by sociology theorists captures the main idea of the positive effects of social participation and it is not actually a new sociological concept (Portes, 1998). Portes (1998) cited Bourdieu s definition of social capital as the first contemporary analysis which posits social capital as an instrument to acquire actual or potential resources through relationships or networks. However, the concept of social capital has started to lose distinct meaning (Portes, 1998; Woolcock, 1998) and it is becoming a multifaceted concept (Sobel, 2002) 23 due to its intangible nature. Portes & Sensenbrenner (1993) found four sources of social capital in the literature: value introjection, reciprocity exchanges, bounded solidarity, and enforceable trust. They stated that the last two are more related with community and the international migration experience. One of the first references to the term social capital can be traced to Hanifan (1916) who argued that there is a positive association between social capital and school performance in rural 23 The literature shows that each author proposes different measures of social capital depending on the outcome to be explained. For instance, to explain the role of social capital in education, Coleman (1988) uses closeness defined as dense ties/relations among people, whereas Teachman et al. (1996) employ intergenerational closure as a measure for social capital to explain schooling outcomes. 94

108 U.S. communities. Hanifan (1916) stated that the term capital is used in a figurative sense and it references to that (i.e., good will, fellowship and families) in life which makes it better. After Hanifan (1916), social capital was rarely studied until Coleman (1988) and Putnam et al. (1993, 1995) who reignited interest in this concept. Coleman (1988) highlighted the role of trust as an aspect of social capital in the acquisition of human capital: he found that the higher level of social capital the lower the student dropout rate. Meanwhile, Putnam et al. (1993, 1995) analyzed community participation to explain the difference in economic success between the North and South of Italy, and factors associated with the decline of social capital within the U.S. These three studies have been broadly cited in most of the literature on social capital. Coleman s study (1988) incorporated obligations and expectations, information channels, and social norms to define social capital, and the importance of social relations among individuals within the family and in the community to make possible the achievement of higher human capital levels of the next generation by reducing the high school dropout rate. Putnam et al. s work (1993) delineated a definition of social capital as civic engagement in the community and they found a strong association between social capital and economic and political outcomes in Italy. However, they did not analyze the determinants affecting social capital. By 1995, Putnam extended social capital definition to networks, norms, and trust but kept the associational nature of social capital that allows collective action for common benefits. Putnam (1995) explored the determinants of social capital in the U.S. during , and his findings show that time spent watching television is the main reason for the decline in social capital among other factors such as education, mobility, the changing role of women, race, family, and welfare state. 95

109 Recent studies have tried to explain changes in social capital based on empirical approaches. Iyer et al. (2005) examine the determinants of social capital for 40 communities in the U.S. but they employ 8 aspects of social capital separately social trust, racial trust, civic participation, ethnic diversity, group involvement, organized interactions, faith-based and informal social interactions instead of an overall indicator as proposed in this research following Rupasingha et al. (2006). Although, they proposed geographical patterns as a determinant of social capital, their model does not include a spatial econometric analysis as undertaken in this research. Glaeser & Redlick (2009) explore the relationship between two forms of social capital group membership and voting in the last election and geographical mobility; they find a negative association between individual social capital and migration but a lack of association between area level social capital and migration flow. Moreover, Putnam (1995) pointed out that there is no causal relationship between mobility and social capital for the U.S. On the other hand, several studies have focused on the effects of social capital on different outcomes. Palloni et al. (2001) studied the impact of social capital on international migration; they found that social capital measured as family relationships increases the risk to migrate from Mexico to the U.S. Based on cross-country data, Dinda (2008) analyzed the role of social capital measured as trust in the formation of human capital and economic growth measured as per capita income; but unlike Coleman (1988), Dinda (2008) shows that education (average years of schooling) causes social capital which in turn affects economic growth. Callois & Schmitt (2009) classified social capital into three categories: bonding, bridging, and linking; their findings show that social capital has a positive effect on local economic growth measured as employment growth and population change for French rural areas. 96

110 However, there is a stream of literature that is critical of social capital basically due to the lack of clarity of the definition of the term and its use to explain outcomes without a further analysis of the causal relationship. The term is applied to so many events and in so many different contexts [that it is starting] to lose any distinct meaning (Portes, 1998, p.2). Portes (1998) and Durlauf (2002) point out that social capital has been defined over a broad range of concepts not necessarily related to each other and even mixing functional and causal concepts. Moreover, Portes & Landolt (2000) point out that social capital has been identified only with positive outcomes, and scholars have employed social capital as a general solution for any social problem. Schafft & Brown (2003) contend that using social capital as the main factor to explain development obscures the role of critical factors such as power and privilege. One way to address this shortcoming is to evaluate what factors motivate individuals to invest in social capital. An econometric analysis including those critical factors as independent variables will allow us to identify the determinants of social capital level across U.S. counties. Most of the literature on social capital and migration has focused on just one-directional effects, the impact of social capital on the international migration flow or the impact of migration on social capital, without controlling for endogeneity. One of the first references regarding the effect of (internal) migration on social capital is found in Putnam (1995) who argues that the decline of social capital in the U.S. is not totally explained by mobility; however, Rupasingha et al. (2006) find that immobility has a significant and positive effect on the creation of social capital. On the other hand, Schiff (2000) explores theoretically the effect of international migration on social capital formation in the receiving and sending countries but without including separate measures for in- and out- migration. Putnam (2007) proposes to explore diversity as a proxy of migration since he claims that immigration increases diversity. Based on 97

111 the Social Capital Community Benchmark Survey, he found that diversity negatively affects the level of social capital measured as trust. For international migration, Portes & Sensenbrenner (1993) describe the social confrontation between the host society and the immigrant which has two effects, reducing the social capital of the native group and increasing the social capital of immigrant groups by bounded solidarity (based on identity) and enforceable trust (based on rewards). Besides, the literature on social capital does not study the joint effect of migration and commuting. So far, most of the literature on social capital has focused on delineating a definition of social capital and its effects without analyzing its determinants or causes. For policy analysis, it is important to understand the factors driving the changes on social capital (Rupasingha et al., 2006). However, to my knowledge, only two studies Glaeser et al. (2002) and Rupasingha et al. (2006) have modeled social capital production based on an economic theory in order to identify factors affecting social capital formation. The social capital model proposed by Glaeser et al. (2002) incorporates a rate of depreciation of accumulated social capital to evaluate it as an individual investment decision but only considers two aspects of social capital: networks and status. Their theoretical model s predictions and findings show a positive relationship between social capital and discount factor, social skills, depreciation rate, and community social capital, whereas there is a negative relationship with age, mobility and relocation. Rupasingha et al. (2006) model social capital production based on a household time allocation model; thus the decision to produce social capital depends on the opportunity costs (of time and resources) and the marginal benefits. They construct a composite social capital index which includes not only associational densities (following Putnam, 1995) but also participation 98

112 in elections and the decennial census, and the number of non-profit organizations. Their main findings show that education, diversity, and family households with children have a significant impact on the production of social capital Social network theory Network theory is based on graph theory with a mathematical foundation and has been used more in physics and biological research than social science research. However, network theory is highly relevant to capture social interactions. A graph is used to represent undirected or directed relations between two agents. Undirected relations reflect a link between two agents but without an orientation; meanwhile, directed relations reflect an orientation from one node to another but not necessarily the other way. Generally, the relation between an agent and itself is not considered in the analysis (Wasserman & Faust, 1994). Network analysis is used to represent social structures and ties among agents, such as individuals, groups, microorganisms, households, communities or counties. These ties may represent different type of relationships such as co-authorship, economic, movement flows (e.g., migration and commuting), or political among others. The ties considered in this research are directed relations created by the population movement at the county level. Migrants and commuters move across counties to reach a new destination for a permanent and temporary basis, respectively. These movements from the origin to the destination may produce webs of social relationships among agents living in the counties involved in migration and commuting which may be understood as network externalities. Massey et al. (1993) state that migration networks are relations among migrants, former migrants, and non-migrants in origin and destination areas through ties of kinship, friendship, and shared community origin. Network 99

113 theory helps to understand the underlying social connections created by migration and commuting beyond the direct link and the number of people moving. The innovative concept developed by social network theory is to go beyond the agent s own characteristics to include agent s social environment for explanations. Network theory captures influence and leveraging processes (Borgatti et al., 2009). Influence processes may be understood as an imitation or contagion process such that an agent copies their friends ideas; meanwhile, leveraging processes represent the opportunity to have access to (private) information or resources due to connections with other agents in the social structure. The concept of centrality and prestige are used to study the prominence or importance of an agent through his direct and indirect connections within a social structure (Borgatti et al., 2009; Maya-Jariego & Armitage, 2007). The most important agent has a strategic location within the social structure; thus, even without a direct link an agent is visible to the other agents. Centrality indices use information on choices made; whereas prestige is calculated based on choices received (Wasserman & Faust, 1994). Network theory relies on matrices to compile network data that complete the graph analysis and allow computing centrality measures. In network terminology, an agent is named a node, the relationships among nodes are represented via edges, and the structure is the network which embeds the nodes and edges (Butts, 2009; Wasserman & Faust, 1994). The importance of a node is reflected in its position within the network which in turn affects the node s outcome in activities such as economic growth, health status, or social capital level. A powerful node will face fewer constraints to reach a better outcome (Borgatti et al., 2009). Three network measures are well-known in the literature degree, betweenness, and closeness to evaluate the importance of a node within the social structure. Centrality measures 100

114 need to be normalized to control for network size in order to compare with other networks of different sizes. Degree centrality is related to the size of the network and the prestige of a node because it reflects the number of direct edges for each node without capturing the importance of a given node within the network. A node may have a range of degree from 0 if it has no ties with the other nodes to a degree equal to n-1 if it has ties to all other nodes (n=nodes). A node with a high degree centrality is a node with a high activity in the network (Borgatti el al., 2009; Maya- Jariego & Armitage, 2007; Wasserman & Faust, 1994). For the case of migration and commuting, degree counts the number of migrants and commuters between two counties; this measure is already included in the migration and commuting flow variables in the econometric analysis. According to the direction of the edges, there are in-degree and out-degree indicators or counts of those who are moving in to a county (in-migrants) and those who are moving out from a county (out-migrants), respectively. Freeman s betweenness captures the frequency of a node employed along the shortest paths to connect two other nodes and it is normalized by (n-1)*(n-2)/2. This indicator does not take into account the direction of the edges, thus there is only one betweenness measure. Betweenness indicates the potential control of a node within the structure because all the information passes through that node, that is, a node with high betweenness is considered essential for the traffic flow. A high betweenness means that a node is used several times to connect two other nodes (Borgatti et al., 2009; Maya-Jariego & Armitage, 2007; Tutzauer, 2007; Wasserman & Faust, 1994). Closeness is the inverse of the normalized sum of the shortest distance to all other nodes. The distance is measured from a node to another using the edges as length. Closeness 101

115 determines how well connected a node is with all others and its communication efficiency (Tutzauer, 2007; Wasserman & Faust, 1994). The most central node has higher values and it means that it has an advantaged position in the network. There are two measures: out-closeness is the sum of the rows, and in-closeness is the sum of the columns. For this research, outcloseness represents out-migration (out-commuting) and in-closeness represents in-migration (incommuting). Finally, eigenvector is other centrality measure but less popular in the network theory. This centrality measure is employed to capture the status of a node. A symmetric data set is needed to calculate the eigenvector measure; however, it is possible to work with asymmetric data taking into account that eigenvectors are not orthogonal (Bonacich & Lloyd, 2001). The eigenvector is calculated based on the first eigenvalue generated by a factor analysis. It is recommended that the first factor should explain at least the 70% of the overall variation in distances (Hanneman & Riddle, 2005). Centrality indicators capture different kind of flows; closeness and betweenness are useful for analyzing nodes importance within the network but they assume different flow characteristics (Borgatti, 2005). Population movement, migration and commuting, imply the transit of persons from one county to another and (perhaps) passing through other counties without repeating the same county (trajectory) until reach the destination. However, for this research, the population movement also embeds knowledge and attitude flows: migrants (Schaeffer, 2011) and commuters (Caragliu & Nijkamp, 2008; Goetz et al., 2010) share their ideas with those in the destination as well as with those left behind. The transmission of ideas is recognized as a key variable to promote entrepreneurship (Hoang & Antoncic, 2003) but it can be extended to social organizations, sharing new ideas to create associations or spreading the 102

116 importance of community participation. The interaction between migrants and their relatives/friends is the foundation for disseminating their ideas and new information (Haug, 2008), which can be applied to the commuting flow, as well. This idea sharing process is similar to an influence process: knowledge and attitudes are spread through interaction and simultaneously without having a target. In sum, for this research, population movement involves two characteristics: 1) transit of people (tangible), and 2) transit of knowledge and attitudes (intangible). The Bonacich eigenvector and Freeman closeness centrality measures are the best indicators of an influence process such as the sharing idea process. These measures are preferred over betweenness because the latter assumes a transfer process (Borgatti, 2005) but migrants and commuters do not lose their knowledge during the process unlike a transfer process. In this research, closeness centrality is used instead of an eigenvector because the movement process does not meet the symmetry required to calculate the eigenvector indicator. A county with a high closeness measure has a better location than other counties to receive information brought by migrants (commuters), and has a greater connectedness in the migration (commuting) flow. For regional studies, network theory is becoming relevant to analyze the role of interrelationships on regional outcomes such as economic development (Goetz et al., 2010; Schaeffer, 2011; Eagle et al., 2010). Findings from a network analysis may help to design policies to improve connections among entities such as counties, e.g., successful projects may be shared among counties. However, to my knowledge there is no applied research linking migration or commuting network centrality indicators with social capital. 103

117 4.3. The social capital model The theoretical framework is based on the household allocation of time models developed by Becker (1965, 1974). Although the empirical analysis proposed is at the county level, the production of social capital embeds decisions at the individual level: each person decides his/her degree of participation in the community life to create social capital. In this context, the allocation of time among several activities (including production of social capital) can be explained by time allocation models. Assume that population is normalized to one (at the county level) as the representative household of county i with utility maximization behavior; thus the objective is to maximize the utility level (Rupasingha et al., 2006) as follows: where i U U C, SK i i C i is the consumption of a composite commodity at the county level (as average), and SK i is the social capital good. Besides, it is assumed that social capital is a desirable good 24. Becker (1974) stated that inputs such as time ( t ), market goods ( G ) and other factors are required to produce C and SK. Thus, C C G C, t C SK SK G, t, S, S SK SK i j where S i stands for a county s socio-economic and demographic characteristics which includes migration and commuting flows among others, and S j stands for characteristics of other counties to capture the spatial effects (which will be complemented in the next section with a 24 Other studies (Callois & Schmitt, 2009; Glaeser et al., 2002; Iyer et al., 2005) have pointed out that individual social capital may have a negative effect on community social capital. 104

118 spatial econometric analysis). Unlike Glaeser et al. (2002), the theoretical model proposed in this research for social capital does not include a depreciation rate because it is assumed that social capital improves with its use (Dinda, 2008). Plugging C and SK into the utility function: i, t, SKG, t, S S U U C G, C C SK as suggested by Rupasingha et al. (2006), the budget constraint includes a time constraint; total time (T ) is employed in three activities, working ( t w ), producing C ( t C )which it is assumed different from t w, and producing SK ( t SK ), but an additional time constraint, commuting time ( SK i j t cm ), is included in the equation, as follows: T t w t C t SK t cm and the budget constraint is: where w stands for wage income, w* tw pc * C psk * SK T t t t p * C p SK w* C SK cm C SK * w* tw is the total monetary income from working, p c and p SK are the prices for a composite commodity and the social capital good, respectively; and p c is normalized to unity. The objective is to maximize utility subject to a budget constraint which in turn includes a time constraint. The aim is to decide on the optimal quantity of social capital ( SK ) that a household should produce to maximize utility. Assuming an interior solution, the first order condition shows that a household produces social capital only if the marginal utility exceeds the opportunity cost (Rupasingha et al., 2006), 105

119 U SK p SK After determining the optimal social capital level, it is necessary to understand how social capital changes. Based on the reduced form of social capital production, the contribution of this paper is to incorporate migration and commuting and their network indicators to analyze the overall effect of population movement variables on social capital formation Data and variable definitions Social capital variable First it is necessary to define the dependent variable: social capital index for 1997 (and 2009). Social capital has been studied as an individual and aggregate level concept (Coleman, 1990). Portes (1998) and Schafft & Brown (2003) point out that the use of social capital in the literature has passed from an individual or small level to an aggregate level of major geographical areas. People build their own social capital by their interaction with other people (Portes, 1998) within a spatial context. This interaction may be accumulated to construct an aggregate social capital 25. Hence, it is relevant to study social capital at the aggregate level to identify those variables susceptible to government intervention. Different studies have highlighted the relevance of social capital at aggregate levels ranging from community to the nation. Israel & Beaulieu (2004) employed community social capital to explain the decrease in the school drop-out rate among U.S. youth. Levitte (2004) found that social capital measured as relationships has a significant effect on the economic development of three Aboriginal communities in Canada. And, at a greater geographical level, Putnam (1993) employed social capital to explain regional differences in Italy. 25 Portes (1998) states that there is nothing intrinsically wrong with redefining it [i.e., social capital] as a structural property of large aggregates (p.21). 106

120 In this research, social capital is the collective action on social and civic activities undertaken by individuals who belong geographically to a given county in order to achieve a common objective. This collective action is represented by four components that are explained later in this section. The social involvement of each person in his community builds the social capital of his community and his own social capital. Following Rupasingha et al. (2006), a composite index is created based on a principal component analysis (PCA) which includes four components. Although there is a consensus in the literature that social capital produces positive economic outcomes, there is a possibility of a negative social outcome such as reproducing racism and discrimination. Portes (1998, p.15) identified in the literature four negative outcomes: exclusion of outsiders, excess claims on group members, restrictions on individual freedoms, and downward leveling norms. However, it is difficult to identify at the aggregate level those kinds of associations more likely to produce a negative social outcome. The four components proposed in my research to capture social capital can be related to private and social outcomes. Associations such as labor, professional, business, civic, and political included in the associational organizations component look for a positive outcome that benefits their group (and members) without considering the (possible negative) effect on the rest of the society (Fidrmuc & Gerxhani, 2008). On the other hand, non-profit organizations (religious and sport organizations) included in the associational component can be associated with positive social outcomes (Fidrmuc & Gerxhani, 2008). However, the components voterturnout and response rate can be identified with private and social outcomes, since the participation in both indicators may pursue not only a private but also a public positive outcome (benefits for the whole society). Each of the four components measures social capital from different perspectives, but there is not a consensus of which one is the best indicator. However, 107

121 it can be argued that those four components of social capital are instruments to achieve an objective, either private or public or both, but that objective is generally identified with economic benefits 26. Here, the PCA is a powerful statistical technique that allows us to reduce the data to avoid redundancy as well as to capture most of the variance and the underlying structure of the data 27. By using PCA it is possible to identify the patterns and characteristics of the data without losing much information. In this research, I use the first principal component that provides the most significant relationship among the four indicators employed. The four indicators of social capital included are: 1) associational organizations reflect the common spaces to share activities and knowledge within a county, and authors such as Glaeser & Redlick (2009), Krishna (2001) and Putnam (1995) identify social capital with social participation; 2) non-profit organizations which reflect solidarity and cooperation that are employed as social capital in the literature (Krishna, 2001; Paldam, 2000) but unlike Rupasingha et al. (2006) I do not include those organizations with an international mission because the focus of this study is on the concept of social capital for a community or county, 3) voter turnout (Coffe & Geys, 2006), and 4) response rate for the Census Bureau s decennial population. Both factors 3) and 4) represent civic participation and capture the engagement of citizens in the political aspects of their counties and civic values (Glaeser & Redlick, 2009; McLaren, 2010; Putnam, 1993; Keefer & Knack, 1997). Thus, Rupasingha et al. s index (2006) captures not only associational behavior but also civic participation and solidarity by using a principal component analysis to encompass the different measures of social capital in just one overall index. Each factor is standardized and measured at the county level. The data employed to construct the first component are from County Business Patterns (CBP). Ten establishments are 26 Although the impact of social capital on economic outcomes is important that analysis is beyond the scope of my research. 27 STATA 2011 manual. 108

122 considered for the associational organizations: religious organizations, civic organizations, business organizations, political organizations, professional organizations, labor organizations, bowling centers, fitness centers, golf clubs, and sport organizations. The descriptive statistics for the variables employed in the econometric analysis are presented in the Table 4.1. The variables are obtained from QuickFacts (U.S. Bureau of the Census) and are based on the literature review. The independent variables are divided into three groups: 1) socio-economic characteristics, 2) population movement variables, and 3) closeness centrality indicators. The socio-economic variables that need a description are income inequality which is calculated as the ratio of mean and median income; and ethnic diversity index (EDI) based on Olfert &Partridge (2011) to capture racial diversity. EDI is calculated as follow: EDI S ji 1 n 1 n j i 1 n, where i is the county, j is the ethnic group, S ij is the percentage of population in county i comprised in the ethnic group j, and n is the number of ethnic groups. I consider 5 ethnic groups: white alone, black alone, Asian alone, Hispanic or Latino, and other races. EDI is the probability that two random persons in a given county belong to different ethnic groups; the higher index, the higher racial diversity. The variables EDI and income inequality are an attempt to control for social divisions based on race and economic classes. There are two opposite effects of diversity on social capital; contact theory contends that diversity stimulates tolerance and solidarity whereas conflict theory contends that diversity stimulates distrust (Putnam, 2007). On the other hand, income inequality is associated with social polarization that makes individuals less likely to participate in 109

123 a collective action (Rupasingha et al., 2006; Fidrmuc & Gerxhani, 2008) and its negative effect remains regardless of when and where social capital is measured (Putnam, 2007). Finally, the variables urban and rural are codified according to the 2003 Rural-Urban Continuum Codes that has a range from 1 to 9, with codes from 1 to 3 labeled as urban, and codes 6 to 9 labeled as rural. Codes 1 to 3 are metro counties, and codes 6 to 9 are non-metro counties with less than 20,000 urban population. Codes 4 and 5 are grouped as the reference category Main independent variables Population movement variables The two key variables in this paper are commuting and migration, and the goal is to understand their effects on social capital at county level. Migration and commuting data are obtained from the U.S. Census of population (1990 and 2000). In-migration is measured as the per capita number of people who moved in to the county in 1990 (2000) during the last 5 years whereas out-migration is the per capita number of people who moved out to another county during the last 5 years. In the same fashion, in-commuting is the per capita number of people who travel toward the county for work whereas out-commuting is the per capita number of people who work in a different county. The effect of migration and commuting (in- and out-) on social capital is not clear and poses an econometric challenge (i.e., endogeneity) 28. In the short run, in-migrants are perceived as newcomers who do not share the same norms as residents of the receiving area and this can 28 Explained in the estimation strategy section. 110

124 reduce social capital. In the long run, however, in-migration has a positive effect on cultural, economic, fiscal, and development issues 29 (Putnam, 2007). Table 4.2 and Table 4.3 report commuting and migration flows for those counties included in the econometric analysis. The numbers show that counties with a high in-migration per capita also have a high rate of out-commuting per capita. This suggests that migrants are moving to their residence area to get closer to the area of their job which in turn raises the question about migrant s involvement in the creation of social capital at the residence or job area. For instance, in 1990 Douglas, CO has a high rate of in-migration and a high rate of outcommuting. The figures in Table 4.2 and Table 4.3 show that there is a consistent migration and commuting behavior over time (1990 and 2000) except for out-migration. Those counties with a high (low) per capita number of in-migrants, in-commuters, and out-commuters in 1990 remain within the highest (lowest) counties in However, those counties sending out more (lesser) migrants in 1990 do not remain within the highest (lowest) ones in Closeness centrality indicator Two 3045*3045 adjacency matrices were employed to calculate the independent variables of closeness centrality for migration and commuting in 1990 and UCINET program was used to calculate the in-closeness and out-closeness centrality measures. The formula is: n 1 CC( ci ), and represents the normalized inverse of the sum of the distance from one n d c, c j1 i j 29 However, the results for social capital in 2009 including migration in 1990 and in 2000 do not show a different effect (See Appendix Table 4.A). 30 Normalized betweenness measures for commuting and migration were calculated but they were not included in the econometric analysis because closeness measures are a better fit for the analysis. Besides, as the correlation matrix presented in the Table 4.B (Appendix) shows, both centrality measures are highly correlated. 111

125 i county c to all other counties, which is converted to a scale by: CC c ) CC( c )* 100. ( i i For in-closeness, c i is the county receiving people. For out-closeness, c i is the county sending out people. Migration (commuting) is the tie linking two counties, and counties are considered as the nodes in the network. Similar to migration and commuting flows, these two matrices or networks are overlapped because they share the same nodes, migrants and commuters with their knowledge pass through the same county; thus it is expected that centrality measures for migration and commuting capture similar properties. Data from the 1990 and 2000 Censuses of population were employed to calculate the absolute number of migrants and commuters to construct the two adjacency matrices. Rows contain the number of out-migrants (out-commuters) from a given county to another; whereas columns have the number of in-migrants (in-commuters) received by a given county. These two matrices or networks have 0 in the diagonal because it is assumed that there is no migration (commuting) within the same county, and the matrices are not symmetric because it is not (necessarily) the case that county A sending out-migrants (out-commuters) to county B will receive in-migrants (in-commuters) from county B. The closeness centrality captures the position of the county within the migration (commuting) network to develop interrelationships and attitudes between natives and migrants (commuters) as well as to receive information brought by them. A higher closeness centrality index indicates that the county has a better location within the network than other counties. There are two indicators, in-closeness and out-closeness, for each flow of migration and commuting. The in-closeness indicator captures whether a county is well-positioned to obtain information brought by those people in the in-migration (in-commuting) flow. In a similar 112

126 fashion, out-closeness captures the centrality of a county in the out-migration (out-commuting) flow. Including the closeness centrality is a new approach in the social capital literature. Table 4.4 reports closeness centrality measures for commuting flows in 1990 and 2000 for the counties included in the econometric analysis. In the case of in-closeness commuting, within the 20 counties with the highest indexes in 1990, only 4 counties (San Diego, CA, Bexar, TX, Orange, FL, and Tarrant, TX) do not remain among the 20 highest in For the case of out-closeness commuting, 6 counties in 1990 do not remain within the highest 20 counties in 2000 (Fairfax, VA, Shelby, TN, King, WA, Wayne, MI, Hamilton, OH, and DuPage, IL). Similarly, Table 4.5 reports closeness centrality for migration flow in 1990 and In the case of in-closeness migration, 5 counties in 1990 do not remain within the 20 highest counties in 2000 (Orange, FL, Orange, CA, Cumberland, NC, Duval, FL, Pierce, WA). Similarly, 4 counties of the highest out-closeness in 1990 do not remain within the 20 highest counties in 2000 (Wayne, MI, Denver, CO, Travis, TX, Oklahoma, OK). Comparing the figures of migration (commuting) per capita and closeness, it is clear that those indicators capture different types of information. In the case of in-closeness migration 2000, Maricopa, AZ; San Diego, CA; and Los Angeles, CA have the highest rate; meanwhile, Chesapeake, VA; Arlington, VA; and Gilpin, CO have the highest rate for in-migration per capita Hypotheses This research has as its main objective to investigate the role played by population movement variables i.e., migration and commuting in the production of social capital at the county level. Based on this objective and the theoretical discussion, three hypotheses are proposed: 113

127 1. Migration and commuting change the production of social capital. - A negative effect is expected for in- and out-migration. It is assumed that in-migrants do not share the same values with the natives and they will not be involved in the community life. On the other hand, out-migrants will disrupt the social life which may discourage the participation in the community life of those leave behind. - Commuting time (minutes) has a negative effect on the generation of social capital. Travelling from residence to work poses a time restriction on commuters who have less time to invest on social capital. - Out-commuting and in-commuting will have a negative effect on social capital. Commuters will have less time to be involved in their residence communities. However, incommuting will have a positive effect as well. In-commuters tend to spend more time in their workplaces which may stimulate their involvement in the community life at the job area. 2. Migration and commuting network centrality indicators have a significant statistical effect on the production of social capital. To my knowledge, there is no research on the subject of social capital that analyzes the effect of migration and commuting networks centrality indicators on social capital formation. There are two opposite effects: - The information passed through the migration (and commuting) network will have a positive effect. Those counties with a better location will capture new ideas and information related to organizations to improve and foster participation in the community. - Negative attitudes toward migrants, which are captured in the network flow, will reduce social capital. 114

128 3. The third hypothesis is that there exist spatial effects on social capital formation. The social capital of neighboring counties affects the social capital of a given county Estimation strategy To address the hypotheses proposed in this research, I employ different econometric techniques. For hypotheses 1 and 2, I rely on traditional econometric techniques such as linear regression and panel data models. Meanwhile, for the third hypothesis I employ a spatial regression model. The empirical analysis poses an endogeneity or reverse causality problem between the independent variables and the social capital index that makes unclear the cause-effect relationship. Portes (1998) points out the need to include controls to address the causality problem in studies of social capital. Although Portes (1998) refers to social capital as a predictor rather than an outcome, two points are relevant in any study dealing with endogeneity or causality problem: to include controls for directionality, and to identify historical origins. Econometric techniques can address endogeneity problems by introducing beginning-of-period values of the explanatory variables to determine the direction of the causality, as an attempt to avoid circular reasoning The estimation of social capital: SK i, t 0 1X i, t 2MVi, t it where it MV includes all the movement variables of the analysis, poses an endogeneity problem because, 0 E. The major concern is related to the main independent variables, i.e., population MV it it movement variables and network centrality indicators. The endogeneity or reverse causality may arise between migration and social capital, the lack of social capital may foster out-migration but migration may negatively affect the creation of social capital as well. To address the potential endogeneity problem, migration for 1990 (2000) is employed to explain social capital in

129 (2009). In general, beginning-of-period values for explanatory variables are employed in the analysis to reduce endogeneity bias since it is assumed that the future (social capital in 2009) E MV, i, t1 it. cannot explain the past (independent variables in 2000), thus 0 This estimation is complemented with a panel data analysis which allows identifying exogenous variations in the independent variables that are not caused by social capital (to avoid endogeneity problems) and to analyze inter-temporal changes. Finally, to avoid heteroskedasticity problems, the Huber-White sandwich estimators were employed First model: OLS for social capital and its components The first model is estimated by a linear regression, using ordinary least square (OLS) in STATA to explain social capital. The aim is to determine the significance and difference between level (per capita) and network indicators for migration and commuting as well as their overall effects. To complete this part, commuting time is incorporated in the analysis. The production of social capital can be represented by a general model given by, SK i, t1 0 1X i, t0 2Mi, t0 3Ci, t0 4CTi, t0 5CM i, t0 6CCi, t0 i where X represents a matrix of explanatory variables previously explained in the section Data and variables, M is migration per capita including in-migration and out-migration, C is commuting per capita including in-commuting and out-commuting, CT is commuting time in minutes, CM is closeness migration including in-closeness and out-closeness, and CC is closeness commuting including in-closeness and out-closeness. The coefficients 2, 3, and 4 will test the first hypothesis proposed. And, the coefficients 5 and 6 will test the second hypothesis proposed. Finally, this model will be estimated to explain social capital index and each of its components. 116

130 Second model: The spatial regression model Recent studies have highlighted the importance of spatial analysis in regional science to incorporate geographical characteristics as a main factor to explain public policy s consequences, migration patterns, and poverty rates (Lacombe, 2004; Rupasingha & Goetz, 2007). Social capital cannot be excluded from this new technique i.e., spatial analysis due to its locational nature and the association between social capital and economic growth. Figure 4.1 depicts the spatial distribution for social capital index in 2009 which shows clusters with a higher value of social capital index in the upper Midwest counties. Spatial econometrics has become an important technique to control for spatial dependence. The weaker expression of spatial dependence is spatial autocorrelation and spatial heterogeneity. In the literature, spatial dependence and spatial autocorrelation are interchangeably used. Spatial dependence arises when values observed in the data are locally related, e.g. the poverty level in area A depends on the poverty level in area B. On the other hand, spatial heterogeneity refers to the instability in model coefficients or error variances caused by geographical factors. In that sense, the relationship between the social capital production and the independent variables proposed (especially movement population variables) may change over space. Based on LeSage & Pace (2009), I propose to use a generalized spatial model (SAC) using Bayesian Markov Chain Monte Carlo (MCMC) methods 31 which consider spatial dependence in the dependent variable and disturbance term. This model is specified as: SK W1 SK X1 31 The Bayesian MCM is employed to avoid heteroskedasticity problems. 117

131 1 W 2 ~ N 0, where W 1 is the traditional first-order contiguity matrix which is row standardized (row sums are one), this matrix collects distance information among neighboring counties by using latitude and longitude coordinates of the internal point 32. X is a vector of demographic and socio-economic characteristics including migration, commuting and network variables which affect social capital production, and SK is social capital. To test if the specification model is correct, I follow LeSage & Pace (2009): if 0 then it is better to use a spatial error model (SEM) which considers that the spatial dependence affects only the disturbance term; whereas with 0 it is better to use a spatial autoregressive (SAR) model which considers that the spatial dependence affects only the dependent variable (social capital production) Third model: Panel data The third model is based on a panel data approach for two periods of independent (t0=1990 and t1=2000) and dependent (t0=1997 and t1=2009) variables with random-effects. The fixed-effect model is not employed because of the presence of time invariant variables, such as rural areas codification or race. However, since the population movement variables are not time-constant, the fixed-effect estimation results are shown in the Appendix Table 4.G to complement the analysis. The key advantage of the fixed-effect panel estimation is that corrects for timeconstant unobserved heterogeneity (i.e. omitted variables). 32 The internal point is an area-weighted average point calculated by using the latitude and longitude of all bounding features

132 Panel data estimation is more informative (more observations are included in the econometric analysis), it captures inter-temporal changes (behavior across time) and the randomeffects model allows for possible correlations of the error between two periods. Likewise, panel data estimation reduces multicollinearity and any bias produced by endogeneity. The coefficient associated with migration (commuting) gives us the impact of changes of migration (commuting) on social capital across time and between counties Results First model The first model considered in the econometric analysis is a linear regression model for social capital at each year, 1997 and that was estimated using a robust variance estimator to address heteroskedasticity problems 34. According to the VIF 35 analysis there is a multicollinearity problem caused by age, age squared, in-closeness and out-closeness migration (VIF>10) in both years 2009 and The average for closeness centrality is included instead of in- and out- closeness to avoid the multicollinearity problem. Age and age squared are needed to control for demographics characteristics. The results in the first column for 1997 and 2009 in Table 4.6 and Table 4.7 confirm hypothesis 1 proposed in this research 36 that population movement variables have a statistically significant impact on social capital. However, out-migration per capita does not have the expected sign, and the commuting variables do not have a clear effect over time. 33 An additional regression for 2005 is presented in the Appendix, Table 4.C. The results are similar to 2009 s results. The decision of not including social capital for 2005 responds to a limitation in the data for the independent variables. 34 The Huber-White sandwich estimators were employed. 35 See Appendix Table 4.D. 36 Similar results are found for social capital index in 2005 (see Appendix Table 4.C). 119

133 The results based on the model without closeness indicators show that a unit increase in out-migration per capita will increase by 5.1 (4.6) units the social capital index in 1997 (2009). The positive effect of out-migration may reflect the searching for comfort by joining or creating associations by those left behind after relatives or friends have left the origin county. The analysis of each component (Tables 4.8 and 4.9) confirms that out-migration has a positive effect on the number of associational and non-profit organizations in 1997 and On the other hand, a unit increase in in-migration per capita will decrease by 6.8 (5.6) units the social capital index in 1997 (2009). The negative effect of in-migrants per capita reflects the distrust on the newcomers and their low participation in the social structure of the receiving county. Tables 4.8 and 4.9 show that in-migration per capita has a negative effect on each component of social capital, that is, the higher number of in-migrants per capita reduce the social participation (associations and non-profit organizations) as well as the civic participation (voter turnout and response rate). The effects of migration remain similar when commuting variables are incorporated 37. Unlike migration, commuting variables do not show the same behavior over time. Incommuting per capita is not significant for 1997 but has a positive and significant effect on social capital in Meanwhile, out-commuting per capita has a negative and significant effect on social capital 1997 but is not significant for social capital Although, commuters (in- and out-) have less time to invest in social capital, the positive effect of in-commuters and the negative effect of out-commuters suggest that commuters are more involved in the social life of the job county rather than in the social life of their residence county, especially participating in associational organizations as Tables 4.8 and 4.9 shown. 37 The results without commuting are presented in the Appendix Table 4.E. 120

134 Similarly, the results in the second column show that commuting time from the residence to the job area will reduce social capital index by 0.07 in 1997 and by 0.04 in 2009 for one minute increased in commuting. However, in- and out- commuting variables are no longer significant when commuting time is included in the analysis. The next step includes closeness centrality indicators for commuting and migration in the econometric analysis for 1997 and 2009 to test hypothesis 2 but commuting time is dropped. The results for migration and commuting variables remain very similar to the previous ones but out-commuting becomes significant with a negative effect on social capital in Additionally, log of median income in 1999 and agricultural employees in 2000 are no longer significant; whereas urban 2003 and manufacturing employees in 2000 become significant. Closeness indicators for commuting have a negative effect on social capital for both years 1997 and This suggests that counties in a better location within the commuting flow experience a negative effect for social capital production. It can be argued that those counties with a better location are not actually receiving commuters in a permanent way but rather they are playing a role of intermediary, and thus the more central counties do not capture the transmission of ideas but rather the attitudes and distrust toward the commuters (non natives) which is reflected in its negative effect on each component of social capital (see Tables 4.8 and 4.9). Similar to closeness commuting, the closeness centrality indicator for migration also has a negative effect. Those counties with a better location within the closeness migration network have a lower social capital index in 1997 and This result is in the same vein as the negative effect of in-migrant per capita flow, which suggests that the main effect of networks on social capital may be through the attitudes towards in-migrants. 121

135 The results for closeness indicators suggest that the negative impact of negative attitudes passed through migration and commuting flows is more important than the potential positive effect of information passed through migration and commuting flows. In that sense, the analysis for each component of social capital shows that closeness measures have a negative impact on associational and non-profit organizations which suggests that information passed on through migration and commuting is not important for social capital formation. Besides, a higher index for closeness reflects a broader connection within the migration and commuting networks that implies a higher diversity of contacts. Hence, it is possible to explain the negative effect as a response to diversity similar to the effect of the variable ethnic diversity. However, the negative effect is very low around 0 for both measures. A unit increase in the closeness migration decreases the social capital index by in 1997 and by in And, a unit increase in the closeness commuting decreases social capital by in 1997, and by in Based on the absolute value of the standardized coefficients (Appendix, Table 4.F 38 ) without considering population movement variables and their network centrality measures, median age and family characteristics (ratio of families with children and ratio of family households) have the largest impact on social capital index for both years 1997 and After including migration and commuting variables, the strength of the variables decreases, and inmigrants (-0.33 for 1997 and for 2009) and out-migrants (0.23 for 1997 and 0.19 for 2009) become important determinants of social capital. Finally, closeness centrality measures were included in the analysis, closeness migration and closeness commuting have a large impact on social capital in 2009 but their impact is less important in These results confirm the relevance of including population movement variables to explain social capital. 38 Appendix Table 4.F reports the standardized (beta) coefficients to compare the relative strength of the predictors included in the analysis. 122

136 In general, the results are similar to those found in the literature but there are also some contrasting and new results to discuss. The results for income inequality and educational attainment have the expected signs according to the literature, and they are highly significant for the base (without network measures) and full models. The absolute size of those coefficients increases in 2009 compared to Based on the full model, counties with a high income inequality also show a lower social capital index; a unit increase in income inequality in 1997 decreases the social capital index by 2.17 in 1997 and by 2.36 in In the case of educational attainment, the positive relationship means that counties with a high population holding a bachelor or higher degree have a higher social capital index. A percent increase of population with bachelor s degree increases social capital index by 0.05 in 1997 and by 0.06 in An important difference with respect to the previous literature is the effect of urban and rural areas. Both variables have a positive and significant effect on social capital in 2009 when the closeness centrality indicators are included in the analysis. These results suggest that after controlling for connections (closeness) there is little difference between the impact of urban and rural areas on social capital in Second model The results employing a spatial econometric approach are showed in Table 4.10 and The estimation of the SAC model reveals that is not statistically significant when closeness indicators are included in the estimation of social capital in 2009 which means that SEM is the best model for the estimation in that case. For social capital in 1997, both spatial parameters ( and ) are statistically significant. Therefore, the most appropriate model to estimate is SAC. 123

137 The results show that the spatial parameter for estimations in 1997 and 2009 is positive and significant which confirms hypothesis 3 proposed in this research. There exists a spatial relationship in the production of social capital, that is, the social capital index in a given county is not independent of the social capital index in the neighboring counties. The lag effect of social capital ( coefficient) changes through time. The value of the coefficient in 2009 indicates that a 10 percentage point reduction in the index of social capital in county j reduces the social capital index in a nearby county by 4.6%. Meanwhile, there was a positive effect in 1997, 4.7% for the base model and 3.6% for the model with closeness indicators. On the other hand, the spatial error coefficient is significant for estimation of the social capital index in 1997 and in 2009 without closeness indicators which indicates that spatially omitted variables in a county have a significant and positive effect not only on the social capital index of that given county but also on the social capital index of nearby counties. The results in Table 4.10 and 4.11 show slight differences in comparison to the results in Tables 4.6 and 4.7. Based on the model with closeness indicators (third column), the significance of most variables remains with no change but most of the coefficients have worsened from the OLS estimation. For the estimation of social capital in 1997, the variables 1990 median age and 1990 in-commuting per capita become significant which confirms the previous results obtained by OLS that commuters are more involved in the community life of their destination county. In the case of social capital 2009, only the ethnic diversity variable becomes negatively significant compared to OLS results. Finally, when controlling for spatial dependence, commuting variables (in- and out-) do remain significant even when time commuting is included in the analysis unlike the results found by the OLS estimation. 124

138 Third model Table 4.12 shows the results for a panel data estimation including both sets of independent and dependent variables for 1999 and 2000 to explore inter-temporal changes. The Breusch-Pagan Lagrangian Multiplier test (LM) indicates that the model with random effects is adequate. The results are similar to those obtained for the estimation of social capital in 2009 and 1997 by OLS and the spatial approach. However, urban area, median age and log median income, and percentage of employees (manufacturing, agricultural, and professional) become significant in this more robust model. The main result is that in-commuting is no longer significant, that is, over time changes in in-commuting per capita are not important for the formation of social capital. Results for the other variables related to population movement remain the same as the results obtained using OLS and spatial methods. In comparison with the fixed-effect model (presented in the Appendix Table 4.G), the estimates for the variables in-migrants per capita and closeness centrality for commuting are no longer significant which suggest that the random-estimates for those variables were inflated by unobserved heterogeneity. However, the variables of out-migration per capita, out-commuting per capita and closeness centrality for migration remain significant. Likewise, commuting time is significant in both models and has a negative effect on social capital Conclusions This paper investigates the importance of population movement variables in the formation of social capital at the county level in the U.S. This study expands the literature on social capital by 125

139 including several measures of population movement beyond the migration flow; commuting per capita, commuting time, and closeness centrality indicators were included in the econometric analysis to explain the total effect of population movement on social capital. Three econometric techniques were employed to obtain a clear effect of each movement variable on social capital. Migration variables (in- and out-) have the same effect regardless of the econometric technique employed. Out-migrants per capita have a positive effect on social capital in 1997 and 2009, whereas in-migrants per capita have a negative effect (which disappears only for the fixed-effect model). The effect of migration variables remains the same after including closeness centrality indicators. The results suggest that higher out-migration per capita will lead to a search for comfort among those left behind, which can be an opportunity to build tight knitting not only for recovering but also increasing social capital. On the contrary, inmigration per capita has a negative effect which likely reflects distrust in the newcomers. On the other hand, the findings for commuting are sensitive to the econometric technique employed but commuting time which has a negative effect on social capital regardless the econometric technique. OLS estimates revealed that out-commuting per capita has a negative impact on social capital in 1997 but not in 2009, and in-commuting per capita is not significant for social capital in 1997 but it has a significant and positive effect in The same results for commuting variables are found if closeness centrality indicators are included in the econometric estimation for social capital in 1997, but both indicators are significant in Based on the panel data estimates, in-commuting per capita does not have a significant effect over time. Meanwhile, the spatial effect model shows that in- and out-commuting per capita have a negative and positive effect on social capital index 1997 and 2009, respectively. These results suggest that commuters are more involved in the social and civic life on their job area instead of their 126

140 residence area. Time commuting remains significant in each model and has a negative effect which suggests as it is expected that time used in commuting has a detrimental effect on social capital because it reduces the spare time that people has to invest in social capital. Finally, the closeness centrality indicators for migration and commuting remain statistically significant regardless of the econometric technique employed with the only exception of closeness for commuting which is no longer significant under the fixed-effect panel model. Although closeness centrality indicators are significant and have a negative impact on social capital index, this impact is negligible, with a coefficient estimated around zero. But, their negative impact on social capital index reflects the negative attitude toward migrants and commuters rather than the information passed through the commuting and migration networks. In sum, the three econometric models confirm the relevance of including population movement variables in the analysis of social capital. However, as was summarized not all the population variables remain significant. Out-commuting and out-migration are significant under the three econometric models whereas in-commuting and in-migration per capita variables are sensitive to the econometric model employed. Policy makers may consider the relevance of population movement on social capital especially from the results of out-migration and out-commuting. First, the positive effect of outmigration per capita cannot be understood as a tool to increase social capital by fostering outmigration but rather should be taken as an opportunity to re-build and build new social capital since the results suggest that people remaining in the original county are more likely to participate in the creation of social capital. Second, the negative effect of out-commuting should draw attention on the commuting behavior, such as the time travelled which has a negative effect on social capital, and the county to where people commute. Besides considering the positive 127

141 effect of in-commuting, there is an opportunity to create ties among those counties connected by commuters to enhance the production of social capital, to learn and share experiences. Finally, all the population movement variables show a significant effect at least in one of the econometric models proposed. Hence, it is relevant to include these variables in the social capital analysis and to take them into account when policy measures are delineated, such as measures to deter the negative effect of in-migration on social capital References Becker, Gary (1965). A theory of the allocation of time. The Economic Journal, 75 (299), pp Becker, Gary (1974). A theory of social interactions. The Journal of Political Economy, 82 (6), pp Bonacich, P. & P. Lloyd (2001). Eigenvector-like measures of centrality for asymmetric relations. Social Networks, 23, pp Borgatti, Stephen (2005). Centrality and network flow. Social Networks, 27(1), pp Borgatti, S., A. Mehra, D. Brass & G. Labianca (2009). Network analysis in the social sciences. Science, 323(5916), pp Bourdieu, Pierre (1986). The forms of capital. In: Richardson, J.G. (Ed.), Handbook of Theory and Research for the Sociology of Education. Greenwood Press, New York, pp Butts, Carter (2009). Revisiting the foundations of network analysis. Science, 325 (414) Callois, J. & B. Schmitt (2009). The role of social capital components on local economic growth: Local cohesion and openness in French rural areas. Review of Agricultural and Environmental Studies, 90 (3), pp Caragliu, A. & P. Nijkamp (2008). The impact of regional absorptive capacity on spatial knowledge spillovers. Tinbergen Institute Discussion Papers /3. Coffe, H. & B. Geys (2006). Community Heterogeneity: A burden for the creation of social capital? Social Science Quarterly, 87 (5), pp

142 Coleman, James (1988). Social capital in the creation of human capital. The American Journal of Sociology, 94, pp. S95-S120. Coleman, James (1990). Foundations of social theory. Cambridge: Harvard University Press. Dinda, Soumyananda (2008). Social capital in the creation of human capital and economic growth: A productive consumption approach. The Journal of Socio-Economics, 37, pp Durlauf, Steven (2002). On the empirics of social capital, mimeo, University of Wisconsin. Eagle, N., M. Macy & R. Claxton (2010). Network diversity and economic development. Science, 238, pp Fidrmuc, J. & K. Gerxhani (2008). Mind the gap! Social capital, East and West. Journal of Comparative Economics, 36, pp Flora, C. & J. Flora (2008). Rural communities: legacy and change. Westview Press: Boulder, CO., 402p. Gertler, M. S. (2000). Social capital. In R. Johnston, D. Gregory, G. Pratt, D. Smith & M. Watt (eds.). The Dictionary of Human Geography (4th ed.). Cambridge, MA: Blackwell Publishers, Glaeser, E. & C. Redlick (2009). Social capital and urban growth. International Regional Science Review, 32, pp Glaeser, E., D. Laibson & B. Sacerdote (2002). An economic approach to social capital. The Economic Journal, 112, pp. F437-F458. Goetz, S., Y. Han, J. Findeis & K. Brasier (2010). U.S. commuting networks and economic growth: Measurement and implications for spatial policy. Growth and Change, 41 (2), pp Hanifan, L. J. (1916). The rural school community center. Annals of the American Academy of Political and Social Science, 67, pp Hanneman, R. & M. Riddle (2005). Introduction to Social Network Methods. Riverside, CA: University of California, Riverside (published in digital form at Haug, Sonja (2008). Migration networks and migration decision-making. Journal of Ethnic and Migration Studies, 34 (4), pp Hoang, H. & B. Antoncic (2003). Network based research in entrepreneurship: A critical review. Journal of Business Venturing, 18 (2), pp

143 Israel, G. & L. Beaulieu (2004). Investing in communities: Social capital s role in keeping youth in school. Community Development Society Journal, 34(2), pp Iyer, S., M. Kitson & B. Toh (2005). Social capital, economic growth and regional development. Regional Studies, 39 (8), pp Keefer, P. & S. Knack (1997). Does social capital have an economic payoff? A cross-country investigation. Quarterly Journal of Economics, 112 (4), pp Krishna, Anirudh (2001). Moving from the stock of social capital to the flow of benefits: The role of agency. World Development, 29 (6), pp LaCombe, Donald (2004). Does econometric methodology matter? An analysis of public policy using spatial econometric techniques. Geographical Analysis, 36 (2), pp LeSage, J. & R. Pace (2009). Introduction to spatial econometrics. CRC Press/Taylor & Francis, 354p. Levitte, Yael (2004). Bonding social capital in entrepreneurial developing communities Survival networks or barriers? Community Development Society Journal, 35(1), pp Massey, D., J. Arrango, G. Hugo, A. Kouaouci, A. Pellegrino & J.E. Taylor (1993). Theories of international migration: A review and appraisal. Population and Development Review, 19(3), pp Maya-Jariego, I. & N. Armitage (2007). Multiple senses of community in migration and commuting. The Interplay between time, space and relations. International Sociology 22(6), pp McLaren, Lauren (2010). Cause for concern? The impact of immigration on political trust. London: Policy Network Paper, 40p. Olfert, Margaret & Mark Partridge (2011). Creating the cultural community: Ethnic diversity vs. agglomeration. Spatial Economic Analysis, 6 (1), pp Paldam, Martin (2000). Social capital: One or many? Definition and measurement. Journal of Economic Surveys, 14 (5), pp Palloni, A. D. Massey, M. Ceballos, K. Espinosa & M. Spittel (2001). Social capital and international migration: A test using information on family networks. American Journal of Sociology, 106 (5), pp Portes, Alejandro (1998). Social capital: Its origins and applications in modern sociology. Annual Review of Sociology, 24, pp

144 Portes, A. & P. Landolt (2000). Social capital: Promise and pitfalls of its role in development. Journal of Latin American Studies, 32(2), pp Portes, A. & J. Sensenbrennerr (1993). Embeddedness and immigration: Notes on the social determinants of economic action. American Journal of Sociology, 98(6), pp Putnam, R., R. Leonardi & R. Nanetti (1993). Making democracy work: civic traditions in modern Italy. Princeton University Press, Princeton, N.J. Putnam, Robert (1995). Turning in, turning out: The strange disappearance of social capital in America. Political Science & Politics, 28 (4), pp Putnam, Robert (2000). Bowling alone: The collapse and revival of American community. New York: Simon Schuster. Putnam, Robert (2007). E Pluribus Unum: Diversity and community in the twenty-first century. The 2006 Johan Skytte prize lecture. Scandinavian Political Studies, 30(2), pp Rupasingha, A. & S. Goetz (2007). Social and political forces as determinants of poverty: A spatial analysis. The Journal of Socio-Economics, 36, pp Rupasingha, A., S. Goetz & D. Freshwater (2006). The production of social capital in US counties. The Journal of Socio-Economics, 35, pp Schaeffer, Lucas (2011). Self-employment income and U.S. migration networks: Is there a relationship? Master Thesis at The Pennsylvania State University. Schafft, K. & D. Brown (2003). Social capital, social networks, and social power. Social Epistemology, 17(4), pp Schiff, Maurice (2000). Love thy neighbor: Trade, migration, and social capital. Washington: Development Research Group, 29p. Sobel, Joel (2002). Can we trust social capital? Journal of Economic Literature, 40 (1), pp Teachman, J., K. Paasch & K. Carver (1996). Social capital and dropping out of school early. Journal of Marriage and the Family, 58, pp Tutzauer, Frank (2007). Entropy as a measure of centrality in networks characterized by pathtransfer flow. Social Networks, 29, pp Wallis, J., P. Killerby & B. Dollery (2004). Social economics and social capital. International Journal of Social Economics, 31 (3), pp

145 Wasserman, S. & K. Faust (1994). Social Network Analysis: Methods and Applications. Cambridge, ENG and New York: Cambridge University Press. Woolcock, Michael (1998). Social capital and economic development: Toward a theoretical synthesis and policy framework. Theory and Society, 27(2), pp

146 Table 4.1. Descriptive statistics Social capital 1997 Social capital 2009 Mean S.D. Min. Max. Mean S.D. Min. Max. Social capital index Income inequality % bachelor's degree or higher Rate of female labor participation Urban (=1) Rural (=1) % owner-occupied hh Median age Median age squared Ratio family hh/total hh Ethnic diversity Log median hh income Ratio family w/ children Manufacturing employees/10, Agricultural employees/10, Profesional employees/10, Out-migrants percapita In-migrants percapita Out-commuting percapita In-commuting percapita Commuting time (minutes) Average closeness migration Average closeness commuting

147 Table 4.2. Migration per capita in 1990 and 2000 In-migration percapita 1990 In-migration percapita 2000 Out-migration percapita 1990 Out-migration percapita 2000 Highest counties Highest counties Highest counties Highest counties Flagler, FL Chesapeake, VA San Juan, CO Geary, KS Douglas, CO Arlington, VA Colonial Heights, VA Chesapeake, VA Storey, NV Gilpin, CO Geary, KS Arlington, VA Esmeralda, NV Douglas, CO Hinsdale, CO Harding, NM Nye, NV Montgomery, VA Hooker, NE Prairie, MT Riley, KS Albemarle, VA Boise, ID Modoc, CA Coryell, TX Riley, KS Wheeler, OR Dinwiddie, VA Camden, GA James City, VA Summit, CO Swisher, TX Onslow, NC Park, CO Grand, CO St. Louis city, MO James City, VA Crowley, CO Converse, WY Clear Creek, CO Clay, SD Prince George, VA Terrell, TX Somerset, MD York, VA Storey, NV Sublette, WY Prince George, VA Gunnison, CO Elbert, CO Park, CO San Juan, CO Liberty, GA Madison, ID Lewis, ID Liberty, GA Summit, CO Nye, NV Duchesne, UT Daggett, UT Park, CO Concho, TX Golden Valley, MT Roanoke, VA Fayette, GA Whitman, WA Wheeler, NE Pitkin, CO Teller, CO Sumter, FL Stanton, KS Albemarle, VA Hernando, FL Coryell, TX Moffat, CO Whitman, WA Pike, PA Forsyth, GA Fallon, MT Riley, KS Lowest counties Lowest counties Lowest counties Lowest counties Elk, PA Erie, NY Campbell, TN Assumption, LA Magoffin, KY Cook, IL Page, VA Luzerne, PA Kings, NY Kings, NY Cocke, TN Lincoln, WV McIntosh, ND Elk, PA Cherokee, AL Rutherford, NC Logan, WV Los Angeles, CA Wayne, KY Robeson, NC Harlan, KY Webb, TX Wilkes, NC Wayne, KY Wyoming, WV McDowell, WV Greene, TN Treutlen, GA Buchanan, VA St. James, LA Lancaster, PA Noble, OH McDowell, WV Harlan, KY Schuylkill, PA Scott, TN St. James, LA Starr, TX Luzerne, PA Page, VA

148 Table 4.3. Commuting per capita in 1990 and 2000 In-commuting percapita 1990 In-commuting percapita 2000 Out-commuting percapita 1990 Out-commuting percapita 2000 Highest counties Highest counties Highest counties Highest counties New York, NY Tunica, MS James City, VA Arlington, VA District of Columbia New York, NY Arlington, VA Storey, NV Arlington, VA District of Columbia Bedford, VA New Kent, VA Somervell, TX Gilpin, CO New Kent, VA Lee, GA Butte, ID Arlington, VA Stafford, VA Spencer, KY Fulton, GA Fulton, GA Jones, GA Stanley, SD St. Louis city, MO Union, SD Douglas, CO Lincoln, SD Martin, IN St. Louis city, MO Charles City, VA York, VA Denver, CO Los Alamos, NM Colonial Heights, VA Bedford, VA Suffolk, MA Pitkin, CO Storey, NV Elbert, CO Plaquemines, LA Suffolk, MA King and Queen, VA Greene, VA Boone, KY Boone, KY Oconee, GA Harris, GA Pitkin, CO Denver, CO Lee, GA Paulding, GA King George, VA Martin, IN Howard, MD Charles City, VA Los Alamos, NM Scott, KY Roanoke, VA Douglas, CO Prince George, VA Plaquemines, LA Putnam, NY Gilpin, CO San Francisco, CA Montour, PA Gilpin, CO Stafford, VA Montour, PA Butte, ID Greene, VA Oglethorpe, GA Union, OH Durham, NC Paulding, GA Catoosa, GA Hancock, KY Union, OH Camden, NC Cheatham, TN Lowest counties Lowest counties Lowest counties Lowest counties San Diego, CA Maricopa, AZ Blaine, ID Box Butte, NE Hidalgo, TX Webb, TX Sweetwater, WY Cascade, MT Lincoln, NV Lemhi, ID Harney, OR Lemhi, ID Maricopa, AZ Clallam, WA Campbell, WY Pima, AZ Mesa, CO Pima, AZ Grant, NM Harney, OR Pima, AZ Mesa, CO Rosebud, MT Webb, TX Fremont, WY Humboldt, CA Maricopa, AZ Nantucket, MA Aroostook, ME Lincoln, MT Humboldt, CA Humboldt, CA Maverick, TX Fremont, WY Webb, TX Maricopa, AZ Camas, ID Aroostook, ME Glacier, MT Clark, NV

149 Table 4.4. Closeness centrality measure for commuting flows in 1990 and 2000 In-closeness 2000 In-closeness 1990 Out-closeness 2000 Highest counties Highest counties Highest counties Highest counties Out-closeness 1990 Cook, IL Cook, IL Harris, TX Cook, IL Harris, TX Harris, TX Maricopa, AZ Harris, TX Dallas, TX Los Angeles, CA Los Angeles, CA Los Angeles, CA Fulton, GA Dallas, TX Cook, IL Dallas, TX Los Angeles, CA Fulton, GA Tarrant, TX Tarrant, TX Shelby, TN District of Columbia Dallas, TX Maricopa, AZ Maricopa, AZ San Diego, CA San Diego, CA St. Louis, MO Wayne, MI Shelby, TN Bexar, TX Fairfax, VA District of Columbia Wayne, MI Hennepin, MN Hennepin, MN Hennepin, MN Fairfax, VA Allegheny, PA Bexar, TX New York, NY Marion, IN St. Louis, MO San Diego, CA St. Louis, MO New York, NY Fulton, GA Oklahoma, OK Mecklenburg, NC Maricopa, AZ Wake, NC Allegheny, PA Fairfax, VA Hennepin, MN Oklahoma, OK Shelby, TN Denver, CO Mecklenburg, NC Jefferson, KY Franklin, OH Marion, IN Bexar, TX Orange, CA King, WA Jefferson, KY St. Louis, MO Denton, TX Orange, CA Allegheny, PA Cuyahoga, OH Franklin, OH Wayne, MI Cuyahoga, OH Orange, FL New York, NY Hamilton, OH Davidson, TN Tarrant, TX Collin, TX DuPage, IL Lowest counties Lowest counties Lowest counties Lowest counties Lewis, ID Thomas, NE Hodgeman, KS Garden, NE Jerauld, SD Arthur, NE Lane, KS Kiowa, CO Garfield, MT Hanson, SD Greeley, KS Logan, NE Faulk, SD Wheeler, NE Wheeler, NE Hooker, NE Powder River, MT Prairie, MT Haakon, SD Dundy, NE Adams, ID McIntosh, ND Logan, NE Greeley, KS McIntosh, ND Gregory, SD Renville, ND Arthur, NE Sheridan, KS Sheridan, MT Wichita, KS Daniels, MT McPherson, SD Douglas, SD Buffalo, SD Stanton, KS Prairie, MT Hooker, NE Arthur, NE Grant, NE

150 Table 4.5. Closeness centrality measure for migration flows in 1990 and 2000 In-closeness 2000 In-closeness 1990 Out-closeness 2000 Highest counties Highest counties Highest counties Highest counties Out-closeness 1990 Maricopa, AZ San Diego, CA Los Angeles, CA Los Angeles, CA San Diego, CA Maricopa, AZ Maricopa, AZ Harris, TX Los Angeles, CA Los Angeles, CA San Diego, CA San Diego, CA Harris, TX Harris, TX Cook, IL Cook, IL Cook, IL Dallas, TX Harris, TX Dallas, TX Clark, NV Cook, IL Dallas, TX Maricopa, AZ Tarrant, TX Tarrant, TX Clark, NV Tarrant, TX Dallas, TX Orange, FL San Bernardino, CA Orange, CA El Paso, CO El Paso, CO Orange, CA San Bernardino, CA Bexar, TX Bexar, TX Tarrant, TX El Paso, CO Onslow, NC Onslow, NC El Paso, CO Bexar, TX Pima, AZ Orange, CA Bexar, TX Hillsborough, FL Chesapeake, VA Clark, NV King, WA Wayne, MI Hillsborough, FL Cumberland, NC Riverside, CA Denver, CO Davidson, TN Duval, FL Orange, FL Pinellas, FL San Bernardino, CA San Bernardino, CA Hillsborough, FL Broward, FL Pinellas, FL Pierce, WA Broward, FL Travis, TX King, WA Pinellas, FL Duval, FL Orange, FL Wake, NC Hillsborough, FL Pima, AZ Oklahoma, OK Hennepin, MN King, WA Pinellas, FL Clark, NV Lowest counties Lowest counties Lowest counties Lowest counties Owsley, KY Jim Hogg, TX Pickett, TN Grant, NE Thomas, NE St. James, LA Logan, ND Buffalo, SD Webster, GA Kidder, ND Thomas, NE Stonewall, TX McMullen, TX Hooker, NE Rock, NE Boyd, NE Jones, SD Logan, NE Webster, GA Baker, GA Robertson, KY Daniels, MT Sanborn, SD Jerauld, SD Taliaferro, GA Pierce, ND Greeley, KS Arthur, NE Harding, SD Treasure, MT Jerauld, SD Taliaferro, GA Hayes, NE Arthur, NE Logan, NE Treasure, MT Wheeler, NE Thomas, NE Grant, NE Garfield, MT

151 Table 4.6. OLS results for social capital, 1997 Dep. Var. Social Capital Income inequality 1990 % bachelor's degree or higher 1990 rate of female labor participation 1993 Urban (=1) 1993 Rural (=1) 1990 % owner-occupied hh 1990 Median age 1990 Median age squared 1990 Ratio family hh / total hh 1990 ethnic diversity 1989 Log median hh income 1990 Ratio family w/ children 1990 Manufacturing employees/ Agricultural employees/ Professional employees/ Out-migrants percapita 1990 In-migrants percapita 1990 Out-commuting percapita 1990 In-commuting percapita 1990 Commuting time 1990 Average closeness migration 1990 Average closeness commuting Constant Observations R-squared (adjusted) Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< *** (-11.56) *** (-8.30) *** (-9.69) 0.045*** (9.11) 0.042*** (9.51) 0.048*** (9.88) 1.467* (1.75) 2.203*** (2.66) 4.624*** (5.70) (0.80) 0.232*** (4.79) 0.249*** (5.01) 0.205*** (4.38) (1.06) (0.27) 0.020*** (4.69) 0.023*** (6.08) 0.018*** (4.78) (1.64) 0.105* (1.90) (1.60) (0.01) (-0.39) (0.08) *** (-11.57) *** (-11.78) *** (-12.88) *** (-11.83) *** (-12.96) *** (-12.73) (-1.53) 0.452*** (3.05) 0.461*** (2.92) 6.262*** (5.81) 4.317*** (4.41) 6.094*** (6.14) (-1.41) (-0.90) (0.31) (-0.57) (0.05) (0.37) (-1.02) 0.041** (1.98) (-0.68) 5.087*** (10.36) 3.525*** (7.80) 3.979*** (8.83) *** (-20.24) *** (-16.80) *** (-18.61) *** (-5.17) (-0.06) *** (-8.97) (-0.08) (-1.14) (0.56) *** (-12.45) ** (-2.13) ** (-2.10) *** (-5.29) *** (-7.90) 8.104*** (4.15) 5.771*** (3.16) 6.481*** (3.40) 3,045 3,045 3,

152 Table 4.7. OLS results for social capital, 2009 Dep. Var. Social Capital Income inequality 2000 % bachelor's degree or higher 2000 Rate of female labor participation 2003 Urban (=1) 2003 Rural (=1) 2000 % owner-occupied hh 2000 Median age 2000 Median age squared 2000 Ratio family hh / total hh 2000 Ethnic diversity 1999 Log median hh income 2000 Ratio family w/ children 2000 Manufacturing employees / Agricultural employees / Professional employees / Out_migrants percapita 2000 In_migrants percapita 2000 Out-commuting percapita 2000 In-commuting percapita 2000 Commuting time 2000 Average closeness migration 2000 Average closeness commuting Constant Observations R-squared (adjusted) Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p< *** (-9.84) *** (-7.66) *** (-9.03) 0.056*** (10.07) 0.058*** (11.54) 0.060*** (11.50) 5.030*** (4.62) 5.485*** (5.66) 7.340*** (7.64) (-0.43) 0.183*** (3.73) 0.163*** (3.25) 0.374*** (8.36) 0.204*** (4.98) 0.188*** (4.46) 0.023*** (4.79) 0.023*** (5.33) 0.019*** (4.53) (0.29) (0.43) (0.46) 0.002* (1.85) 0.001* (1.77) 0.001* (1.80) *** (-7.94) *** (-8.37) *** (-9.30) (-0.85) (0.56) (0.04) *** (-6.47) (-0.94) (-0.91) 6.590*** (5.98) 3.539*** (3.59) 5.042*** (5.13) (-1.25) 0.037* (1.91) 0.054*** (2.96) ** (-2.43) (0.68) (0.64) (-0.77) (1.34) (-0.21) 4.609*** (9.06) 3.122*** (6.54) 3.585*** (7.46) *** (-15.65) *** (-12.97) *** (-14.28) (-0.36) (-0.78) *** (-6.64) 1.131*** (3.72) (1.63) 1.001*** (3.21) *** (-9.26) *** (-5.63) *** (-5.60) *** (-4.82) *** (-5.82) *** (5.84) 9.475*** (4.48) 9.427*** (4.38) 3,045 3,045 3,

153 Table 4.8. OLS results for components of social capital, Income inequality 1990 % bachelor's degree or higher 1990 rate of female labor participation 1993 Urban (=1) 1993 Rural (=1) 1990 % owner-occupied hh 1990 Median age 1990 Median age squared 1990 Ratio family hh / total hh 1990 ethnic diversity 1989 Log median hh income 1990 Ratio family w/ children 1990 Manufacturing employees/ Agricultural employees/ Professional employees/ Out-migrants percapita 1990 In-migrants percapita 1990 Out-commuting percapita 1990 In-commuting percapita 1990 Average closeness migration 1990 Average closeness commuting Associational organizations Non-profit organizations 1997 Voter turnout Response rate Census *** (-7.07) *** (-9.23) *** (-2.77) *** (-5.79) (0.30) 0.752*** (7.53) 0.006*** (15.15) (0.89) 2.737*** (5.43) ** (2.41) (0.75) 0.384*** (5.78) (-0.72) 4.883*** (3.96) 0.030*** (6.27) (0.68) (0.07) (1.37) 0.008* (1.88) *** (-4.87) 0.007*** (3.07) ** (-1.96) 0.004*** (11.41) *** (-5.99) 0.108*** (3.13) 2.669** (2.47) (-1.39) (-1.60) * (-1.79) (-0.43) 0.000* (1.70) 0.000** (2.12) *** (-8.00) *** (-15.25) *** (-5.42) *** (-4.79) ** (-2.55) *** (-5.82) *** (-12.22) *** (-14.09) (0.22) (0.58) 0.049*** (4.03) 0.052*** (3.93) 2.295*** (4.16) *** (7.99) (-0.93) 0.624*** (7.70) (-0.16) (-0.74) (1.57) (0.66) (-0.57) (-0.34) (0.95) 0.041** (2.19) (-1.58) (1.32) (-0.27) ** (-2.51) 0.846*** (3.19) *** (8.63) 0.372*** (9.95) *** (-5.47) *** (-11.29) *** (-12.19) *** (-9.93) *** (-11.88) *** (-8.05) *** (-10.99) *** (-4.39) 0.146*** (5.50) 0.486** (2.42) (0.61) *** (-3.32) 0.066*** (2.68) (-1.46) (-1.55) *** (-2.72) 0.004** (2.40) (-1.21) *** (-8.23) *** (-7.62) * (-1.80) Constant 2.446** (2.05) *** (7.76) 0.720*** (4.30) 0.465*** (2.90) Observations R-squared (adjusted) Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 3, ,045 3,045 3,

154 Table 4.9. OLS results for components of social capital, Income inequality 2000 % bachelor's degree or higher 2000 Rate of female labor participation 2003 Urban (=1) 2003 Rural (=1) 2000 % owner-occupied hh 2000 Median age 2000 Median age squared 2000 Ratio family hh / total hh 2000 Ethnic diversity 1999 Log median hh income 2000 Ratio family w/ children 2000 Manufacturing employees / Agricultural employees / Professional employees / Out_migrants percapita 2000 In_migrants percapita 2000 Out-commuting percapita 2000 In-commuting percapita 2000 Average closeness migration 2000 Average closeness commuting Constant Associational organizations Non-profit organizations 2009 Voter turnout Response rate Census *** (-7.21) *** (-7.04) *** (-5.15) * (-1.67) (0.54) 1.442*** (8.32) 0.007*** (17.15) *** (-6.49) 3.434*** (4.92) *** (3.31) 0.740*** (9.48) 1.030*** (9.22) (-1.11) 6.567*** (2.93) 0.016*** (3.84) *** (-2.79) 0.063** (2.43) 5.422*** (3.46) 0.008** (2.17) *** (-4.58) (1.32) *** (-2.95) 0.004*** (11.86) *** (-3.76) (0.78) (0.13) (-0.30) (-0.71) (0.97) 0.049** (2.07) (0.93) (0.20) *** (-10.25) *** (-8.17) (-0.33) (-1.53) 0.176*** (2.84) (-0.90) *** (-3.98) *** (-14.03) (-0.87) ** (-2.53) 0.055*** (3.65) 0.196*** (9.57) 5.353*** (7.97) *** (6.26) *** (-5.04) 0.195** (2.14) (1.49) (0.51) 0.007*** (4.71) (-0.45) (-0.82) ** (2.31) (-0.42) (0.62) (-1.41) 2.402** (2.45) *** (-2.83) *** (-2.83) 1.110*** (3.01) *** (7.16) 0.210*** (5.27) *** (-6.82) *** (-10.16) *** (-9.63) *** (-10.65) *** (-6.13) *** (-5.48) *** (-9.19) 0.045** (1.98) 0.127*** (4.14) 0.932*** (4.18) (0.80) (1.52) 0.128** (2.55) ** (-2.54) *** (-5.29) *** (-3.84) 0.007*** (4.67) *** (-3.27) *** (-4.34) *** (-4.45) (-1.53) 6.353*** (4.10) *** (7.82) (-0.53) *** (-6.91) Observations R-squared (adjusted) Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 3, , , ,

155 Table Spatial econometric estimation for social capital, 1997 Dep. Var. Social Capital Income inequality 1990 % bachelor's degree or higher 1990 rate of female labor participation 1993 Urban (=1) 1993 Rural (=1) 1990 % owner-occupied hh 1990 Median age 1990 Median age squared 1990 Ratio family hh / total hh 1990 ethnic diversity 1989 Log median hh income 1990 Ratio family w/ children 1990 Manufacturing employees/ Agricultural employees/ Professional employees/ Out-migrants percapita 1990 In-migrants percapita 1990 Out-commuting percapita 1990 In-commuting percapita *** (-8.42) *** (-6.62) *** (-7.36) 0.048*** (11.63) 0.049*** (12.38) 0.052*** (12.47) (-0.98) (0.63) 1.577** (2.47) (-0.26) 0.120** (2.25) 0.124** (2.34) 0.094* (1.84) (0.61) (0.21) 0.020*** (6.40) 0.019*** (5.85) 0.016*** (4.99) 0.140*** (3.55) 0.130*** (3.37) 0.132*** (3.37) (-1.36) (-1.27) (-1.16) *** (-9.62) *** (-10.44) *** (-10.87) *** (-5.89) *** (-8.27) *** (-7.44) (-0.62) 0.471*** (3.61) 0.457*** (3.42) 2.637*** (4.15) 1.954*** (3.05) 2.787*** (4.38) ** (-2.09) (-0.84) (-0.41) (-0.48) (0.37) (0.56) (-0.66) (1.33) (0.13) 2.738*** (9.29) 2.385*** (8.37) 2.451*** (8.26) *** (-15.25) *** (-14.37) *** (-14.58) *** (-6.73) *** (-4.09) *** (-10.39) 0.891*** (3.52) (1.29) 0.668*** (2.58) 1990 Commuting time (minutes) *** (-8.23) 1990 Average closeness migration 1990 Average closeness commuting *** (-4.06) *** (-4.19) *** (-4.08) *** (-5.49) Rho 0.464*** (11.79) 0.300*** (7.36) 0.357*** (8.37) Lambda 0.185*** (2.63) 0.350*** (6.05) 0.305*** (4.67) Constant (0.76) (0.83) (0.77) Observations R-squared (adjusted) 3, , , Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<

156 Table Spatial econometric estimation for social capital, 2009 Dep. Var. Social Capital Income inequality 2000 % bachelor's degree or higher 2000 Rate of female labor participation 2003 Urban (=1) 2003 Rural (=1) 2000 % owner-occupied hh 2000 Median age 2000 Median age squared 2000 Ratio family hh / total hh 2000 Ethnic diversity 1999 Log median hh income 2000 Ratio family w/ children 2000 Manufacturing employees / Agricultural employees / Professional employees / Out_migrants percapita 2000 In_migrants percapita 2000 Out-commuting percapita 2000 In-commuting percapita 2000 Commuting time (minutes) 2000 Average closeness migration 2000 Average closeness commuting *** (-6.74) *** (-7.39) *** (-7.96) 0.048*** (12.48) 0.049*** (12.64) 0.051*** (12.89) (-0.80) 3.241*** (3.97) 3.922*** (4.76) (0.68) 0.103** (2.35) 0.094** (2.15) 0.148*** (3.45) 0.150*** (3.73) 0.145*** (3.65) 0.010** (2.36) 0.014*** (3.70) 0.012*** (2.99) (1.50) (0.49) (0.61) 0.001** (1.95) 0.001** (2.27) 0.001** (2.10) *** (-6.17) *** (-9.32) *** (-9.45) *** (-6.16) *** (-2.84) *** (-3.37) *** (-6.06) (-0.80) (-0.68) 3.295*** (4.30) 3.346*** (4.33) 3.810*** (4.88) ** (-2.30) (1.58) 0.028* (1.72) (-0.69) (0.55) (0.61) (0.80) (0.65) (0.16) 2.540*** (8.01) 3.203*** (9.06) 3.304*** (9.06) *** (-10.95) *** (-13.51) *** (-14.08) ** (-2.13) *** (-2.60) *** (-5.72) 1.107*** (4.90) 0.551** (2.35) 0.775*** (3.34) *** (-4.72) *** (-5.29) *** (-5.11) *** (-4.67) *** (-5.26) Rho *** (-9.34) Lambda 0.867*** (54.43) 0.509*** (21.49) 0.525*** (21.52) Constant *** (5.90) (3.93) 6.486*** (3.49) Observations R-squared (adjusted) Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 3, , ,

157 Table Panel data estimation results (Random effects) Social Capital Index Income inequality % bachelor's degree or higher Rate of female labor participation Urban (=1) Rural (=1) % owner-occupied hh Median age Median age squared Ratio family hh / total hh Ethnic diversity Log median hh income Ratio family w/ children Manufacturing employees /10000 Agricultural employees /10000 Professional employees /10000 Out_migrants percapita In_migrants percapita Out-commuting percapita In-commuting percapita *** (-13.66) *** (-5.77) *** (-6.89) 0.034*** (9.37) 0.029*** (7.74) 0.032*** (8.51) (-0.80) (0.83) 1.006* (1.74) 0.093** (2.56) 0.123*** (3.43) 0.120*** (3.36) 0.171*** (5.03) 0.119*** (3.51) 0.117*** (3.45) 0.014*** (3.96) 0.021*** (6.21) 0.019*** (5.59) 0.116*** (2.84) 0.102*** (2.62) 0.118*** (3.05) (-0.24) (-0.06) (-0.44) *** (-11.09) *** (-12.95) *** (-13.56) *** (-15.70) *** (-10.66) *** (-11.09) *** (-11.16) 1.114*** (9.65) 1.060*** (9.09) 4.241*** (5.85) 2.324*** (3.26) 3.010*** (4.26) *** (-2.94) *** (-2.73) ** (-2.35) 0.465** (2.46) ** (-2.14) *** (-2.64) (-0.40) 0.050*** (3.11) 0.032** (2.07) 3.233*** (11.55) 3.000*** (10.70) 3.184*** (11.33) *** (-11.22) *** (-14.03) *** (-14.79) *** (-9.62) *** (-8.01) *** (-13.88) (0.86) (-1.23) (-0.53) Commuting time (minutes) *** (-7.33) Average closeness migration Average closeness commuting Constant *** (-4.20) *** (-4.33) *** (-9.61) *** (-10.53) 9.500*** (8.45) *** (-3.46) *** (-3.17) Observations Number of fips R-squared (adjusted) LM Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<

158 Figure 4.1. Social capital index,