Are Social Networks Exclusive? The Case of Immigrant Economic Assimilation

Transcription

1 Are Social Networks Exclusive? The Case of Immigrant Economic Assimilation Jingjing Ye Southern Methodist University January 1, 2013 Abstract Previous research has highlighted the importance of informal job search via social networks, however, the impacts of dierent types of networks have always been examined in isolation. This paper incorporates multiple social networks into one framework, which enables not only the valuation of labor market impacts from social networks, the comparison of their relative eectiveness, but also the identication of their interactions. In particular, the paper models the job search process of immigrants, who may utilize the \native network" accessed through inter-ethnic marriage and the \immigrant network" formed by neighboring immigrants. Theoretically, the impact of either network is contingent on whether the use of social networks is exclusive. The empirical evidence accords well with the exclusive case where being a member of one network limits one's access to information from other sources, and concludes that the relative eectiveness of the native network depends crucially on the quality of the immigrant network. JEL: C26, J15, J64, R23 Keywords: social network, interaction, job search, immigrant assimilation, exogamy I am grateful to Dr. Daniel Millimet for his guidance, support, and suggestions. I also thank participants at the 2011 Southern Economic Association Annual Conference and at the 2012 Brown Bag seminar series at SMU. All remaining errors are mine. Corresponding author: Jingjing Ye, Department of Economics, Box 0496, Southern Methodist University, Dallas, TX Tel: (214) Fax: (214) jingjingy@mail.smu.edu.

2 1 Introduction In 2001, around 44 percent of the workers in the U.S. have found their jobs via social networks. 1 This pervasive use of social networks raises important research questions about their impact on labor market outcomes. Dierent types of networks have been studied, for instance, network consisting of friends and relatives (Bentolila et al., 2010; Goel and Lang, 2012; etc.), network based on clustered immigrants (Andersson et al., 2009; Tu, 2010; etc.), and network takes the form of inter-ethnic marriage (Furtado and Theodoropoulosv 2009; Kantarevicv 2004; etc.). However, existing studies have always examined the impact of dierent networks in isolation and the empirical evidence remains mixed. Motivated by the fact that individuals tend to search from several sources of information simultaneously, this paper seeks to answer the following questions. What is the impact of each network, if multiple networks are available, on job search outcomes? Does it matter whether one social network correlates with another? If yes, what is the relationship between dierent networks? And how does this relationship aect the productivity of social networks in job search? To answer these questions, the paper introduces multiple networks into the job search model, which generates predictions about how the productivity of social networks varies, according to their possible interactions. Then those predictions are tested empirically to yield evidence towards the relationship between social networks. This approach is appealing not only because it helps to understand the impact from multiple sources of information, but also provides an indirect way to identify the relationship between networks, when direct solutions are hindered by stringent data requirement. The model takes the role of social networks as a manner of obtaining job-related information, which depends on each agent's memberships in dierent social networks as well as the information quality transmitted by each network. Each agent's job search outcomes depend on his information relative to competitors in the labor market. Intuitively, the membership in one social network correlates with that in another in two possible ways. If individuals opt out of other networks when joining a new one, the use of social networks is exclusive. This exclusivity of social networks may emerge from either the demand-side eects (i.e., the individual interacts less with previous network members because of social preference or time constraint) or from the supply-side eects (i.e., the individual is marginalized by previous network contacts as a result of social sanctions or network resources protection). It is also possible that people are able to maintain multiple memberships when social networks are non-exclusive, if none of those eects holds. Specically, this paper models the job search process of immigrants who utilize two networks to collect job-related information: the \immigrant network" formed by clustered immigrants and the \native 1 According to the study by Franzen and Hangartner (2006) based on the 2001 International Social Survey. 1

3 network" accessed through inter-ethnic marriage. The model reveals that, if social networks are exclusive, a stronger immigrant network will aect the in-group members' job search outcomes positively, but negatively for the outsiders (people in the native network). While in the non-exclusive case, a stronger immigrant network is benecial for all job seekers. Prior to continuing, two notes concerning the reason to focus on the group of immigrants are worth mentioning. Firstly, immigrants may rely more heavily on informal job search networks due to cultural or language barriers (Andersson et al., 2009; Alcobendas et al., 2010; etc.). Within the U.S., the immigration system is "...one of paradox: relatively welcoming to immigrants, yet laissez-faire toward immigrant integration," (Duncan et al. 2008, p. 6) while immigrants and their descendants have become increasingly responsible for future U.S. population growth. Not surprisingly, the lack of either institutional or public assistance gives social networks a crucial role in helping immigrants' cultural assimilation, and eventually their labor market outcome convergence. 2 Another reason arises from the relative ease of conceptualizing dierent social networks for immigrants. In the literature, the immigrant network has been considered as the clustering of co-ethnic people, namely the immigrant enclaves. Most empirical studies favor positive eects of enclave participation on immigrants' economic outcomes (e.g., Edin et al., 2003; Tu, 2010), while others nd the opposite (e.g., Borjas, 2000). Recent studies revealed that the productivity of the enclave can be better captured by the density of neighboring employed immigrants (professional acquaintances) rather than all the immigrants (random acquaintance) in the economic assimilation process (Hellerstain, 2008; Andersson et al., 2009). On the other hand, relationships with natives could provide access to the host labor market for immigrants by facilitating social and cultural assimilation, and then lead to broader job choices (Drever et al., 2008). Inter-ethnic marriage is commonly understood to be the most direct and ecient way to be accepted by the other ethnic network. However, the eectiveness of this search channel remains empirically ambiguous. According to Kantarevic (2004), the intermarriage premium on immigrants' income obtained by OLS completely vanishes once the self-selection issue is taken into consideration. Other studies (Meng and Gregory, 2005; Furtado and Theodoropoulos, 2009) found the positive casual eect of exogamy (interethnic marriage) on other labor market outcomes after taking account of human capital and endogeneity for immigrants. Proceeding forth, I test the hypotheses of the model using a married male immigrant subsample from 2 Although the most straightforward and traditional indicator for the integration process might be naturalization, the acquisition of citizenship is usually costly in time, and lags behind immigrants' integration status. To this end, the economic assimilation performance of immigrants has been measured in terms of their labor market outcomes in the literature, and in this paper the outcomes are immigrant employment status and their realized wage rates. 2

4 the 5 percent 2000 U.S. Census (IPUMS, Ruggles et al., 2008). The potential endogeneity of network usage is addressed by utilizing an instrumental variable (IV) approach. I use the same data to construct the measures of social networks. Based on previous studies, the immigrant network quality is dened by the average employment rate of similar immigrants living in the neighborhood; and the citizenship of the immigrant's spouse is the indicator for native network membership. Those measurements will be further explained in Section 3.2. The results are summarized as following to answer the previously raised questions. First, I nd robust evidence that the usage of social networks is exclusive; namely, being a member of a certain network limits one's access to information from other sources. This nding justies the necessity to consider multiple search methods and their interaction in the framework simultaneously, otherwise, the potential omitted variable bias and the heterogeneous network eects due to the network interaction need to be addressed. Second, as the theoretical model predicts, the relationship between networks does matter as it aects the productivity of social networks in job search. The empirical nding is consistent with the theoretical prediction. Given the identied relationship of exclusivity, the impact of an increase in information provided by one network is contingent on which network one possesses. For immigrants aliated to the immigrant network, positive networking premium is found with a stronger immigrant network. However, immigrants who are members of the native network become worse o since their rival network now gets stronger. Lastly, consonant with the mechanism underlying the theoretical model, the impact of joining one network depends on its relative information quality. If the quality of the immigrant network exceeds a certain value - in another word, enough information is transmitted by that network - the immigrant network is more eective in enhancing labor market performance. Such critical value is estimated to be about 80% average employment ratio in the immigrant network for the employment outcome; once the immigrant network average employment ratio drops below 80%, the native network membership becomes more benecial. This research contributes to the literature along several dimensions. First, the paper develops a simple job search model for the purpose of understanding the mechanism through which social networks aect job search outcomes. Particularly, the model generates testable hypotheses concerning the interaction between networks. Inspired by the theoretical model, the empirical model overcomes the imperfect observation of social network memberships in the data under certain assumptions. Furthermore, by incorporating two types of social networks into one framework, this approach allows one to compare the relative eectiveness of each type and then shed some light into the \black box" of why people use one channel over the other, which has to date received little attention in the literature. 3 3 A related strand of literature introduced formal job search methods along with the informal one in the job search models 3

5 Besides, the paper oers insights on the current mixed empirical ndings on network eectiveness. Inspired by the seminal work of Calvo-Armengol and Jackson (2004), the model presented in this paper newly introduces the role of competitors in the one-sided job search model, which drives the possible negative return of networks for outsiders. Such ambiguity of the network eectiveness is veried by the empirical results. Moreover, this paper oers a preferred empirical model by including multiple job search methods. Based on literature, the choices of specic search channels are driven by cost-benet considerations for alternative search methods. Thus, including variables of multiple search methods addresses the omitted variable bias. In addition, including interaction terms of dierent informal job search methods reduces the bias resulting from mis-specied functional form. Such attempts were hindered by selection problems (since use of each job search channel is endogenous), small sample sizes, and multicollinearity in the literature (Holzer, 1988; Weber and Mahringer, 2008). Finally, this analysis has larger implications beyond just understanding the eects of networks and their interplay on job search for immigrants, but rather it sheds some new light on the bigger questions on immigrants assimilation and multiculturalism. The inability to search from multiple networks simultaneously for jobs reects the possible intergroup conict, in response to increased job competition among people from dierent ethnic, race, and culture backgrounds. To protect valuable resources and preserve their use for in-group members, assimilation is resisted by groups by forming relatively closed ethnic boundaries. Indeed, it is not rare to observe that the succession of various groups of immigrants in previously nativedominated neighborhoods instigates ethnic awareness and solidarity, and leads to conict, particularly at the time of economic recession. The paper proceeds as follows. Section 2 models the job search process utilizing multiple information sources. Section 3 discusses the empirical model specication, the data, and the identication strategies. Section 4 presents the empirical results across the dierent outcome variables and estimation techniques. Discussion and conclusion are given in Section 5. (see e.g. Holzer, 1988; van den Berg and van der Klaauw, 2006; and Weber and Mahringer, 2008). Their work, primarily theoretical, found unambiguous search eort redistribution from the high cost search channel to the relatively low cost ones, regardless of the cross-methods relationships. It seems quite plausible that the eectiveness of most formal search methods will be independent in the production of job oers, particularly for the various institutional methods (Holzer, 1988). However, the productivity of one social network is very likely to depend on its relationship with other networks, as what this paper predicts. 4

6 2 Theoretical Framework The aim of the theoretical model is threefold: rst, it incorporates the role of information into the immigrant job search process; second, it analyzes the interplay between two social networks by formalizing how each social network contributes to the immigrant's information advantage in the labor market; third, the theoretical predictions of networking eect aid in the interpretation of the empirical results. 2.1 Job Search and Information This section introduces the role of information explicitly in the job search model based on Mortensen et al. (1994) and van den Berg (1994). This model allows information to play two roles in the job search process: rst, more information leads to more job oers heard by each applicant; second, more information improves the job oer quality by increasing the highest possible wage. This one-sided stationary job search model is set up as follows. Assume job oers w arrive according to a Poisson process with parameter, which is referred to as the job oer arrival rate in the literature. Such job oers are random draws from a bell-shaped distribution F (w), dened over non-negative real number line with nite mean Ew. w r is the reservation wage, and w is the highest wage that the immigrant would be oered by a rm. The subjective rate of discount is denoted by. The individual is interested in choosing a policy (i.e., the optimal reservation wage) that tells him whether or not to accept any particular job oer. Job oers must be answered as they are discovered. Otherwise, the worker rejects the oer and searches for a better one. During the spell of unemployment, non-labor income b is received. The reservation wage maximizing the individual's own expected present value of income is the unique solution to: Z w w r = b + (w w r w r )df (w): The instantaneous rate at which a searcher escapes from unemployment, also called the hazard rate is given by the product of the job oer arrival rate and the conditional probability of accepting a qualied job oer: = Pr(w r w w) = [F (w) F (w r )]: The expected value of the relative wage oer received given that the participant is qualied, and given that the oer is acceptable is E(wjw r w w) = R w w r wdf (w) R w w r df (w) : 5

7 Next, the assumption that the arrival rate and the highest possible oer depend in no way on the behavior of the agent is relaxed. The relative information gained by the agent will aect the job search results via both the frequency and the quality of job oers the participant is going to hear. Empirical studies support the assertion that informal job search is productive in generating job oers (Edin et al 2003, Andersson et al 2009, Tu 2010), and that each job seeker is more likely to receive a job oer when there is more information available to exchange in the network. Thus the job oer arrival rate,, is a non-decreasing function of information advantage 0, when is dierentiable on I. Information advantage I is a relative measure of information for individual i comparing to other job seekers in the economy. More information can help reduce uncertainties, search friction and skill depreciation during the job search. The maximum wage this unemployed individual could hope to receive, w, is non-decreasing on information advantage I has a non-negative impact on reservation the highest possible wage oer. 4 0, when w is dierentiable on I. It has to be noted that information 0; through its eect on job oer arrival rate and Since both w r and w (weakly) increase with one's information advantage I, the impacts of I on labor market outcomes, and E(wjw r w w), are ambiguous. One way to resolve this ambiguity is to impose the restriction that job oers come from a log concave probability distribution function f(w). 5 The log-concavity property of the empirical wage distribution is justied in Appendix A.4. The intuition behind a log concave wage distribution function states that if the reservation wage slightly increases, then the expected wage in employment increases by a smaller or equal amount. Log-concavity provides a simple restriction on f(w) that guarantees an increase in the frequency of oers increase and lowers the expected duration of unemployment, which is summarized 0, when f(w) is log concave. (1) It is worth pointing out that although theoretically ambiguous, most empirical literature on immigrants favors the positive eect from the job oer arrival rate on (Andersson et al., 2009; Tu, 2010; Meng and Gregory, 2005; Furtado and Theodoropoulos, 2009, etc.). This result probably originates from the fact that immigrants have low and rather xed reservation wage. The rst generation immigrants are found 4 See proof in Appendix A.1. 5 See a formal proof in the Appendix A.2. f(w) is a log-concave function is only one of the sucient to be nonnegative. The class of distribution satisfying log concavity is quiet extensive. It includes the exponential and beta families, the families of logistic and extreme value distributions that are truncated from below at zero, and the whole family of normal distributions truncated at zero. See Bagnoli and Bergstrom (2005) for lists of common distribution function satisfying log concavity. 6

8 to have lower reservation wages than both native workers and second generation workers (Constant et al. 2010). The model can also be used to study the eect from information on the expected wage rates. Under the assumption that more information has non-negative eect on both the maximum wage and the reservation wage, then realized (expected) wage rate conditional on employment can be shown to be (weakly) increasing in the amount of information collected from social networks: r w 0: (2) To sum up, in the immigrant job search process, information is benecial in terms of both higher probability to nd a job match, and higher expected wage rate. If more information helps immigrants to achieve better performance in the labor market, then the question of what contributes to larger amount of information become interesting. 2.2 Information and Social Networks It has been shown that information facilitate job search and therefore lead to better labor market outcomes for immigrants. In the interchange of job-related information, the role of multiple networks and their possible interplay will be formalized in this section. Denition 1 For each immigrant i, his information advantage I is a function of i's information mix I i, all his competitors' information mix I i, depending on his socioeconomic characteristics X i. I = I(I i ; I i ; X i ): (3) The information advantage I describes the amount of information possessed by a representative immigrant i to outperform other job seekers in the labor market. This immigrant i gains information advantage when his own information mix increases, but loses if his competitors' information > 0; < 0: i Each immigrant i collects his information from all the social networks he can access. The amount of information obtained from a certain network s depends on whether he has access to it (membership) as well as the quality of that network. 6 See formal proof in Appendix A.3. 7

9 In this setting, conditional on their socioeconomic characteristics X i, workers are heterogeneous only in their access to information through dierent informal information sources. It has to be noted that the formal job search eect is considered to be approximately homogeneous for all job seekers given X i. This is not an unreasonable assumption, especially in the context of immigrants job search since immigrants rely heavily on informal job search methods and the productivity of formal job search channel can be mostly explained by the observed individual characteristics, such as race, education, gender, etc. Mahringer, 2008). (Weber and Denition 2 The information mix obtained from S networks is denoted as I i = P S s=1 Ds i Qs, where D s i is a dichotomous variable (0; 1) indicates i's membership to the network s, Q s is the quality of the network s. 7 Similarly, the competitors' information mix I i depends on the average access to all social networks by other immigrants. Suppose there are K immigrants, then the average information obtained by competitors 1 P for individual i is K P S K 1 k6=i s=1 Ds k Qs. It can be rewritten as I i = P S s=1 Ds i Qs, letting D s i = P K k6=i Ds k ), which is always less than 1, since Ds i is a dichotomous variable.8 ( 1 K 1 An interesting remark from the denition of the information mix function can be summarized as s D s i i > 0; which means if the quality of a certain network increases, individual i will benet at least as much as his average competitors, as long as i is a member of that network. For the sake of simplicity and without loss of generality, I consider a two network case below where s = N; M. Suppose immigrant i faces two sources of information, the immigrant network source M and the native network source N. For example, he can hear oers from his friends with the same ethnical/cultural background about where to nd the job suits him the best, or he can learn from his native friends who know the local labor market better and reduce the search frictions. Following the denition of information mix, the amount of information obtained through native network N and immigrant network M becomes: I i = Di N Q N + Di M Q M (6) I i = D N iq N + D M iq M ; 7 The simplication of assuming equal weight for every network s can be relax without changing the theoretical conclusion and the empirical model setting. 8 As long as the population exhibits some variation in the choice of social networks. 8

10 which depends on i's membership of each network, Di s, and as well as the quality of each network Qs, s = N; M. Consider a newly arrived immigrant i who is looking for a job, his initial social network endowment only contains his immigrant contacts. He is not able to access information from the native network due to either cultural and language barriers, or the lack of native friends. As such, this immigrant is belonged to the immigrant network, Di M = 1, but not to the native network, so Di N = 0. After spending some time in the receiving country, the immigrant can accumulate his native social capital by interacting with native contacts as well. He can choose to join the native network by setting D N i = 1. This action invokes the question of how his immigrant network membership is going to change. If the use of immigrant and native network is exclusive, one can only choose one network to join. One possible explanation comes from the constraint of time to be spent on networks. In order to get useful information, people need to cultivate links in each social network. However, the amount of time is limited. Immigrants may choose to distribute all his time in the more productive network. Another way to interpret the exclusivity of social networks arises from the concept of membership itself. The notion of membership implies a coherent sense of self that depends on stable group values and a sense of "belongingness". For people belong to multiple social networks, their conguration of loyalties and identication is constantly in ux and their social boundaries are never secure. The literature of crosscultural psychology is rich with examples of all kinds of problems encountered when people are intensely exposed to other cultures (Triandis, 1984). Due to these potential tension and anxieties that may be encountered, immigrants may choose only one social network to be belonged to. Meanwhile, immigrants may be marginalized by other in-group members even if they desire to maintain connections outside of the network. In those groups, multiple memberships can be considered as a violation of trust, and a threat of valuable resources spillover to out-group members. The strong sense of interdependence among in-group members engenders trust and solidarity, and forms a relatively closed social (cultural/ethnic) boundaries and social sanction (Jackson and Schneider, 2011). These relatively closed network boundaries can preserve the use of valuable job information for in-group members, and at the same time, exclude the possibility of multiple network memberships. If the immigrant and native networks are non-exclusive, one is able to maintain both memberships at same time. When immigrants have a strong desire to retain their identity of origin or when pluralism is encourage or accepted in the receiving country, it is likely for immigrants to be a member of immigrant and native network at the same time. There are rich literature in sociology studying the trans-nationalism of international migrants and their osprings, whose identications and loyalties transcend the ethnic/cultural boundaries (Sanders 2002). 9

11 The possible interactions between social networks could be summarized as follows. If immigrant and native network are exclusive, then either immigrant i keeps searching for jobs only through the immigrant network and stays out of the native network, then Di M = 1; Di N = 0; or, immigrant opts out of the immigrant network and become a member of the native network, then Di M = 0; Di N = 1. If social networks are non-exclusive, immigrant i can access information through both immigrant and native networks at the same time, hence Di M = 1; Di N = 1. For the sake of simplicity in further analysis, the notion of exclusivity can be captured by a single parameter by letting D M i = 1 + D N i ; (7) where = 0 refers to the case when native and immigrant networks are non-exclusive, and = the two networks are exclusive. 1 when 2.3 Eect of Social Network Proceeding forth, I investigate the eect of each network s, s = M; N; in terms of labor market outcomes, and E(w) under dierent scenarios explained in the previous section. The equations (1) and (2) imply that a larger informational advantage increases the probability of employment and earning payos. As such, the eectiveness of the of each @Q = s s, has the same s, since Utilizing equations (4), (5), and (6), the eect of the immigrant network M and native network N to the hazard rate and expected earning payos E(wjw r w w) can be summarized as following. 9 When social networks are exclusive ( = 1) and Di M = 1; Di = 0; > < 0 @I M (1 D i) > 0; N 0 D N i < i 9 It would also require a sucient but not necessary condition on the relative measure of i = i : It means that for the same amount of change in the information mix for both individual i and his average competitors, the eect (in magnitude) from his own information mix I i is as same as the eect (in magnitude) from his competitors' information mix I i. Therefore, if both I i and I i increases the same amount, the information advantage of individual i will be as same as before. : 10

12 When social networks are exclusive ( = 1) and D M i = 0; D N i M N > @I i D M i < 0, and 0 i N (1 D i) > 0; i When social networks are non-exclusive ( = 0), Di M = 1; Di N = 1; > > 0 N Therefore, the s relationship between M i D M i) > N i D N i) > along with data on D s s i, can be used to ascertain the The intuition runs as follows. If social networks are exclusive, joining in any one social network will prevent the immigrant from obtaining information from other sources. For immigrant i connects with the immigrant network M, a stronger immigrant network will result in a better informational advantage for i and therefore better labor market outcomes, while a stronger native network N will only benet their competitors, i, leads to a loss in i's informational advantage. If the immigrant i connects with the native network N, a stronger immigrant network M now deteriorates the informational advantage of the outsider i since he can not access the high quality information anymore, while his competitors - the immigrants remaining in the immigrant network - are now better o in the labor market. Thus a higher Q M results in worsening job search outcomes while a higher Q N leads to better information and also better job search outcomes. If the immigrant and native networks are non-exclusive, the more network connections the immigrant has, the bigger informational advantage he has, and also the better labor market outcomes will be. Then no matter which network is stronger, labor market performance will be improved. 3 Empirical Methodology This section establishes the empirical model to test the immigrant network eectiveness on labor market outcomes, addressing concerns over potential endogeneity and sample selection issues. 11

13 3.1 Empirical Model Following the theoretical model, the labor market outcomes Y i for immigrant i depend on i's information mix I i, all his competitors' information mix I i, as well as i's socioeconomic characteristics X i. Utilizing equations (1), (2), and (3), the empirical model is specied as following: Y i = I i + 3 I i + X i 4 + e i ; (8) where Y i is the job acquisition probability and log wage rate w for immigrant i, and e i is the error term contains the unobserved heterogeneity. Substituting for I i - and similarly I i - using (6) and allowing heterogenous network quality yield the structural model: Y i = (Di N Q N i + Di M Q M i ) + 3 (D N iq N i + D M iq M i ) + X i 4 + e i : (9) However, I only have empirical measures of Di N and Q M i, but not Di M and Q N i. To proceed, I substitute in (7) for the unobserved Di M and to assume that Q N i is constant across individuals. Following the convention in the literature, the native network membership Di N uses the citizenship of the spouse. The assumption of constant Q N i implies that all native spouses oer the same quality information regarding the local job market. Consider native spouses as bridges to the local culture and social conventions for immigrants: they can teach immigrants how to appropriately ll job applications, how to present themselves during interview and how to speak the native language more uently. This information may not vary a lot across dierent native individuals. 10 Then equation (10) is what the reduced empirical model looks like by substituting (7) into (9) and assuming Q N i = Q N for all i: Letting Y i = D N i + 3 Q M i + 4 (D N i Q M i ) + 5 D N i + 6 (D N iq M i ) + X i + e i (10) 1 = 1 (11) 2 = 2 Q N 3 = = 2 5 = 3 Q N 10 When native spouses help their husbands/wives very dierently, there exists heterogenous treatment eect. With the identication stategy of using IV, the results will be interpreted cautiously as LATE estimators. 12

14 6 = 3 : I can now restate some of the theoretical predictions in light of the empirical model. First, the value of, which is = (12) identies the relationship between immigrant and native social network from equation (7). 11 In addition, the impact of immigrant social network in this linear model setting M i = D N i + 6 D N i (13) = 2 D M i + 3 D M i, (14) which depends on whether one is a member of that network, D M i, as well as the marginal returns to i's and all his competitors' information sets, holding competitor's membership xed. According to the theoretical prediction in Section 2.3, the impact of increasing immigrant network quality is negative, D N i = 1; if social networks are exclusive; such impact becomes positive, networks are M M i < 0; when > 0; when D N i = 1, if social Furthermore, the benets of the native network membership in terms of labor market outcome improvement under the model specication is Y i j D N i =1 Y i j D N i =0 = Q M i (15) = 2 (Q N + Q M i ); (16) which depends on information quality transmitted through both networks, Q N and Q M i, and the impact on membership in other network,. In case two networks are exclusive, the formula (16) becomes 2 (Q N Q M i ). The impact of joining one network is determined by 2, the marginal return to i's information mix, and (Q N Q M i ), the relative information quality of social networks. 3.2 Data and Descriptive Statistics This study employs the ve percent sample of the 2000 U.S. Census as reported by the Integrated Public Use Microdata Series (IPUMS, Ruggles et al., 2008). Because of the diculties in interpreting labor market 11 In fact, the transformation of model from equation (9) to (10) imposes a non-linear restriction that 2= 5 = 4= 6 on coecients. Native network quality Q N can also be identied from Q N = :The estimation of the structual coecients will be discussed in Section 4. 13

15 outcomes of females and students, the sample is restricted to married, foreign born males between the ages of 18 and 60 who are not in school. Also, only immigrants who arrived in the US before the age of 18 are included since they are most likely to have been exposed to the U.S. marriage market. An immigrant is dened to be anyone who is born outside of the 50 states and not born to U.S. parents. A native is dened as a citizen born within the U.S. Observations who are U.S. citizens but not born in the U.S. are not included due to their possibly lack of local labor market knowledge. It has to be noted that marriage with a native spouse could grant the other party in the marriage citizenship automatically. Thus the estimated native network eect may combine both of the benets from citizenship as well as the help from native spouses on local labor market. However, from my empirical evidence, citizenship acquisition does not drive the conclusion since results remain robust using immigrants without citizenship. 12 Also, Furtado et al. (2011) found that those in inter-ethnic marriages do not obtain any additional benet from acquiring citizenship. To minimize sampling error, the sample is restricted to immigrants residing in Metropolitan Statistical Areas (MSAs) with more than 10 observations and belonging to ethnic groups with more than 30 observations. 13 In order to deal with singleton dummies and measurement error, observations whose spouse is younger than the age of 18 or older than 56 are excluded. The nal sample consists of 84,485 observations. The dependent variables are employment status and log hourly wages. The employment dummy is set equal to one if the immigrant is at work; it takes value zero if the observation is unemployed or not in the labor force. Unpaid family workers are also dropped from the sample. Wage rates are calculated as each respondent's total pre-tax wage and salary income divided by number of hours worked during the previous year. The list of control variables X i includes immigrant i's age, education level, veteran status, race, disability dummy, birth place, whether lives inside/outside central city in metro area, English uency, natural log of years in the US, whether he is a recent immigrant, census region; spouse's age, education level, birthplace, English uency; and the fraction of the MSA population made up by immigrants with the same country of origin, the unemployment ratio of native laborers of the same age in that MSA. The individual disturbance term is assumed to be independent across geographical units, but is allowed to be correlated across observations within the same location. 12 The author analyzed the sample of individuals without U.S. citizenship. Results are summarized in Table A7. The restricted sample exhibits limited variation in inter-ethnic marriage: about 15% of them marry to natives. However, the empirical results from this restricted sample remain robust. Thus we can exclude the possibility that citizenship, comes with exogamy, drives the conclusion. 13 Metropolitan statistical areas are counties or combinations of counties centering on a substantial urban area. 14

16 The native network membership(exogamy) Di N takes the value of one if the immigrant is married to a native who is present in the household and zero if he is married to another immigrant. Marriage to another immigrant does not necessarily imply marrying someone with the same ethnic background, and in an exogamous marriage the husband and wife may come from the same ancestry. However, based on the data, neither issue seems to be a major concern. The immigrant network quality for each individual i, denoted by Q M i ; is the share of employed immigrants who reside in the same spatial unit, fall into his age range, and have the same ethnicity background (referred to as an age-ethnic-geographical cell hereafter). There are two geographical levels, MSA and PUMA (Public Use Microdata Area) used for this quality measurement. 14 Intuitively, the age-ethnic-geographic criterion groups similar immigrants together, who are more likely to communicate with each other, and more importantly, who know more about local labor markets. This quality measurement includes both the direct and indirect connections within the network as information can be transmitted through word-of-mouth communication. The share of employed similar immigrants measures the quality of information being exchanged within these groups, since the more employed group members an individual has, the more likely the individual will hear about new job openings. Comparing to the most widely used geographically or ethnically based \enclave" measurements in the literature, this immigrant network measurement captures not only the distance of spatial and ethnic proximity among the immigrant group, but also the \quality" of information ow in that network. Table 1 summarizes the demographic and economic descriptive statistics for the married male immigrants subsample from the 2000 US census 5% data. About 78 percent of immigrants who marry other immigrants share the same self-reported rst ancestry and 87 percent share the same birth place. But only 32 percent of the immigrants who marry natives reported the same ancestry and none have the same birth place. As Table 1 reports, these immigrants perform well in the labor market since nearly 80 percent of them are employed. This may not be surprising, given that most of the immigrants have already spent over twenty years in the U.S. In order to see whether the native network is a potential contributor in helping immigrants in the labor market, it is helpful to compare the immigrants marrying native spouses with the immigrants marrying immigrant spouses. We nd that the immigrants with native spouses tend to have a higher employment probability. While the results may be due to their higher education attainments, better English skills, and higher spouse employment probability, the facilitation of job information ow in the network arising from 14 PUMA identies the Public Use Microdata Area (PUMA) where the housing unit was located. PUMAs generally follow the boundaries of county groups, single counties, or census-dened "places". 15

17 exogamy could also inuence the results. 3.3 Identication One empirical diculty in isolating the causal social network eect results from the endogenous network measurements. If immigrants self-select into intermarriage, or they choose their residency based on potential labor market outcomes, the estimation will merely show the correlation between individuals' labor market outcomes and their social network membership or quality. In addition, wage rates are only available for those who are employed, and the possible sample selection bias needs to be addressed. These two issues and their solutions will be discussed in this section Endogeneity The rst endogeneity issue arises with the variable Di N, the native network membership. As pointed out by Kantarevic (2004), intermarried immigrants may possess characteristics that are valued in both labor and marriage markets, such as physical appearance, high IQ, language prociency, and so on. Although a fairly extensive list of controls are included in the model (such as age, education attainment, race, birthplace, and unemployment rates of natives living in the neighborhood) the marriage choice can still be correlated with unobserved traits of immigrants. The estimates may also be biased if immigrants choose their residency based on unobservables aecting their labor market outcomes, aside from those due to the networking eect. Besides, the quality of the immigrant network, Q M i, becomes endogenous due to simultaneity, since the outcome of one immigrant also aects other immigrants in the enclave. People may choose to live with neighbors sharing similar features such as educational degrees, income levels, or personality traits. For instance, more outgoing immigrants may want to live in more hospitable communities while those who like to keep to themselves more will choose to reside in a more private neighborhoods. If these characteristics are aecting the job nding process independent of their eects on the quality of information ows in the social network, the estimation will be biased. To address these endogeneity issues, instrumental variables for both native network membership and the quality of the immigrant network are used. Following the literature, sex ratios (i.e., relative numbers of men and women) are often used as an instrument for marriage prospects. 15 In this paper, two sex ratios are constructed: the ratio between immigrant females and males and the share of immigrant females versus all 15 According to Angrist(2002), the empirical results are broadly consistent with theories where higher sex ratios increase female bargaining power in the marriage market. 16

18 females. 16 The two instruments vary on ethnicity, age group, and/or geographical unit to better describe the exogamy prospects in details. The rst instrument Z 1 is the share of females in the immigrant's age group residing in his MSA that is foreign-born. It captures the probability of running into an immigrant female conditional on randomly meeting a female: Z 1 = imm f age MSA native f age MSA + immf age MSA ; where imm f age MSA is the number of female immigrants in the age-msa cell, and nativef age MSA is the number of native-born females in the age-msa cell. The second set of instruments reects the probability of meeting a woman instead of a man with similar ethnic background: Z 2, dened as the sex ratio between immigrant females over immigrant males in an observations' age-ethnic-msa cell. The denition of age groupings remains the same. The higher this sex ratio is, the smaller the female bargaining power in the marriage market will be. Z 2 = immf age ethnic MSA imm m ; age ethnic MSA where imm m age ethnic MSA is the number of male immigrants in the age-ethnic-msa cell, and imm f age ethnic MSA is the number of female immigrants in the age-ethnic-msa cell. The age group is created within an eight year interval. 17 For an immigrant i, his age group contains females who are not younger by more than six years and not older by more than two years than him, and males who are not more than four years younger/older than him. It diers for genders to take into account the fact males tend to marry females who are younger than they are, and the mean age dierence in our sample between husbands and wives is two years. In the sample, more than 70% of couples fall into this range of age dierences. There are several potential pitfalls of using these instruments for native network membership. Both female and male immigrants may be attracted to a city for higher predicted labor market rewards. However, if female and male respond to this incentive dierently, the sex ratio exclusion restriction may not be valid since the instrument captures the gender-specic dierential reaction to labor market outcome aside from the probability of nding a spouse. 16 One more instrument was considered in the estimation: the ratio between immigrants' fraction of unmarried females and the natives' fraction of unmarried females in the ethnic-age-msa group. However, this instrument was dropped for the native network as it did not pass the redundancy test. 17 The robustness of the interval of age group has been tested by using smaller and larger age-range groupings. The empirical results are not sensitive to the choice of age interval in the ethnic-age-msa cell. 17

19 The second pitfall may arise from the possibility of a lag from the time the marriage decision was made to the time the information was collected. If this is the case then the marriage market conditions may have changed in the interim. However, since the results from the new immigrants sample (who arrived within ve years) appear to be very similar, the author does not overly concern over this issue. Besides, Z 1 and Z 2 seem to be quite stable over time. The mean dierences of Z 1 and Z 2 from 1980 to 2000 are less than 0:15, and the standard deviations are 0:07 and 0:85, respectively. As such, the instruments are quite stable over time. To address the possible endogeneity of the immigrant network quality, Q M i, three instruments are considered. They are the average of education attainment, language uency and log year in the U.S. in that immigrant's age-ethnic-msa group, excluding the immigrant's observational information. Only the average education attainment of residents in the age-ethnic-msa cell is kept after the instrument redundancy test. To test the robustness of the results, alternative sets of instruments were also used. Using the Census data in 1980, a set of instrumental variables similarly dened as in the 2000 Census were constructed. 18 Also, two other instruments are considered: the exogamy rate in the 1980 Census sample for the native network membership Di N and the employment rate in the 1980 Census sample for the immigrant network quality Q M i Sample Selection Bias The sample selection bias arises because wages are only selectively observed for those whom are employed. This paper uses the Heckman selection correction model (1979) with two identication strategies. Identication comes from either the non-linearity of the IMR (Inverse Mill's Ratio), which is due to the assumption of joint normality of the errors, or exclusion restrictions that are especially useful in the presence of possible substantial collinearity. The exclusion restriction arises from including an extra variable in the rst step to predict the employment probability. The extra variable is a dummy variable indicating whether the immigrant has any children in the household. However, there are mixed empirical ndings about the eect of presence of child(ren) in the household on father's wage. Angrist and Evans (1998) use instrumental variables to estimate the eect of the birth of a third child on the labor supply of men and women and nd no signicant eect of this birth on men's labor supply. In Millimet (2000), children are found to aect male wages signicantly in a sample of older 18 However, the PUMA level data is not available in the 1980 Census, so only the MSA level estimation will be done. 18

20 households, where the wife is between the age of 35 and 49. Using the same PSID data, Lundberg and Rose (2002) found that fatherhood signicantly increases the hourly wage rates and annual hours of work for men. Immigrants are found to be highly responsive to their earning potential and, unlike that of US-born natives, are less constrained by their family roles (Stier and Tienda, 1992). In Cobb-Clark (1993), the presence of children and the number of children younger than 5 years old do not signicantly aect the wage rates of immigrants. 19 In case that the exclusion restriction of using presence of child(ren) in the household is invalid, I rely on the non-linearity of the IMR for identication for comparison. 4 Empirical Results 4.1 Baseline Employment Table 2 presents the baseline results on employment using equation (10), with the left panel containing the results treating social network measurements as exogenous and the right panel displaying the two stage least square (2SLS) results treating those variables as endogenous. Unlike the structural coecients in equation (9), parameters estimated directly from the reduced model, specied in equation (10), do not have intuitive interpretation. Therefore, I additionally present the derived intuitive results in the middle panel within each set of results, utilizing formulas in (12), (13), and (15). First, I test whether the social networks are exclusive by their relationship coecient () and marginal immigrant network eectiveness M i ). Second, I report the estimated critical value of immigrant network quality (Q M i ) by comparing the relative eectiveness of social networks. Also, I present the results of several diagnostic tests at the bottom panel in Table 2. Finally, in both the OLS and 2SLS models, I provide results from dierent geographical scales, MSA and PUMA. However, for the sake of brevity, I discuss the results at MSA level as the baseline and leave the results at PUMA level in the robustness section. Turning to the OLS results, I obtain three main ndings. whether social networks are exclusive. The rst salient point to make is that There are two ways to identify the relationship between social networks. The rst one is using the theoretical prediction on the sign of marginal networks' M i. For immigrant i in the native network, the marginal eect of immigrant network should be positive, if two networks are non-exclusive; in the case of exclusive networks, the marginal eect of immigrant network becomes negative. In fact, according to (13), the marginal eect of immigrant network lies within 19 Cobb-Clark was using the sample of wives, who are believed to be more likely aected by fertility decisions. See Table 2 on page 991 in Cobb-Clark (1993). 19