The Effects of Social Networks on Wages: A Cross-City Analysis Gabriel Movsesyan∗ December 22, 2014

Abstract I evaluate the wage impacts from exogenous and endogenous peer effects within different metropolitan areas. I construct comparison groups for workers within the same industry and PUMA residence based on support from existing literature on social networks, and estimate peer and network effects in a spatial auto-regressive model with a network structure that incorporates group fixed effects. Inclusion of these effects reveals that both observed and unobserved factors within each network have significant effects on individual outcomes. This is the first attempt to estimate the social effects on wages using the spatial econometric approach proposed by Lee (Journal of Econometrics 2007; 140(2), 333-374). The industry-PUMA combination is a statistical association that merges geographic and social space for a new perspective on the realm of spatial effects. The model includes spatial autoregressive processes in the dependent variable. I find limited evidence of spillover (endogenous) effects, which vary significantly across the cities in the analysis.

JEL-Classification: J31, C31 Keywords: Social interactions, Spatial dependence, Peer effects.

∗ Ph.D. Program in Economics, The Graduate Center, The City University of New York, New York, United States, email: [email protected].

1

1

Introduction

Social networks are a primary determinant of labor market outcomes. Explicit and implicit interactions within these networks comprise the ties and resources that individuals use to learn about and search for jobs. These interactions are significantly influenced by an individual’s surrounding spatial and social structures. A rapidly growing literature has emerged at the intersection of urban and labor economics – it seeks to measure and distinguish the ways in which neighbors and neighborhoods shape an individual’s labor market outcomes.1 We know, however, that many individual decisions are determined concurrently with social networks and the outcomes of one’s neighbors. Individuals sort themselves into social groups, neighborhoods, and industries, and the resulting endogeneity of social structures present a major challenge in identifying effects that originate from different sources. Does one’s network transmit any spillover effects that improve one’s labor market outcomes? This study contributes to the literature by examining the effects of social networks on wages in various United States metropolitan areas. Using a spatial econometrics framework, correlated effects at the group level can be controlled for and the reflection problem that Manski (1993) highlighted can be partially addressed. The Manski model categorizes the social effects a group can have on an individual’s outcomes, and the challenges in uniquely identifying them in conventional econometric frameworks. Endogenous or social effects stem from the decisions or outcomes of one’s peers. Exogenous or contextual effects arise from the characteristics of one’s peers. Additionally, if individuals have comparable outcomes due to any similar observable or unobservable attributes or face similar environmental factors, then correlated effects are present. Given an appropriate model specification and data, these can be uniquely identified. The separate identification is crucial because the endogenous effect (the impact of peers’ outcomes on one’s own) is the source of the social multiplier – the spillover effects from a neighbor’s improved labor market outcomes in our example.2 This paper is the first to use the framework introduced by Lee (2007) and further developed by Lee, Liu and Lin (2010) and Lin (2010) where peer effects are estimated for wages. Determining causal inference of social interaction effects is also dependent on the data availability and selection of appropriate reference groups. In general, econometric approaches to networks assume that the individual is influenced by peers in his or her group but not by any outside of it. Lee’s (2007) model assumes that an individual is equally affected by all the other members in one’s reference group. Studies using data that specifies the actual friendship network or relationship structure can customize their spatial weights matrix (e.g. (Lin, 2010)).3 Other innovative studies create reasonable reference groups on some aggregated level (such as grade or school) when there is no information on the intra-group correlation. My proposed reference group is the intersection created by the overlap of industry classification and one’s neighborhood.4 There is extensive empirical evidence of significant social interactions operating on the level of local neighborhoods. These interactions occur more frequently in larger cities and among individuals who share demographic and socio-economic characteristics. Policy interven1

The evaluation of social interactions and network effects in the context of labor and urban economics thus far are comprehensively reviewed by Ioannides (2013) and Zenou (2009). 2 The social multiplier is defined by Ioannides (2013) as the sum of the direct effect on an individual and “the indirect effects through the feedback from the effects of others in the social group” (p18). 3 The Add Health dataset has been especially popular for such empirical studies on network effects as its student respondents list their friends who are also respondents in the survey (so long as they are in the same school). Prominent examples include (Fryer and Torelli, 2010), (Bramoull´e, Djebbari and Fortin, 2009), and (Calvo-Armengol, Patacchini and Zenou, 2009). 4 This study uses the Public Use Microdata Area (PUMA), which is the smallest identifiable geographic unit within the publicly-available Census and American Community Survey Data. These datasets are described in Section 3.

2

tions may look to improve labor market outcomes for specific communities; a strong endogenous (spillover) effect indicates that targeting a sample of group members would directly benefit them and have multiplicative effects on their peers. If only contextual effects are present, however, then labor market outcomes might be more contingent on one’s neighborhoods or workplaces, and so more extensive transportation, training, or residential subsidizations would be necessary. By using the industry-PUMA nexus as the reference group, I look to evaluate the existence of social and spillover effects, conditional on residence in a particular neighborhood and employment in a particular industry. The peers in these networks are those that live in the same PUMA and work in the same industry. Demographic differentials are an important inquiry here as well: empirical evidence suggests greater proportional effects from social and neighborhood interactions for Latinos (Weinberg, Reagan and Yankow, 2004). Immigrants may have significant wage effects in the population as well, depending on the degree to which they are substitutes for native workers. Once foreign-born and natives are in the same industry-PUMA reference group, other determinants of wage differentials may become apparent. Bayer, Ross and Topa (2008) use restricted-access residential microdata and find that individuals, matched on economic and demographic characteristics, who live on the same block are significantly more likely to work in the same location. This suggests strong information transmission and informal referrals about job opportunities. Our analysis highlights a more specific form of neighborhood effect, albeit at a more aggregated group level. Borjas (2014) emphasizes that estimates of spatial correlation are not robust to specification changes. The definition and structure of peer networks may lead to similar concerns in this study.

2 2.1

Previous Literature Spillover Effects in Networks and Neighborhoods

Within groups and neighborhoods, individual decisions and social networks are simultaneously determined. The confluence of individual characteristics, behaviors, and decisions has significant impacts on the outcomes of each member. Social scientists have tried to observe and define social environments and networks in order to estimate their influence. Becker (1974) was among the first to formalize an economic analysis of social interactions, and made the case for incorporating one’s social environment into a welfare function. This allows social interactions to be interpreted as externalities, as these relationships and connections produce potential spillover effects for all network members. One member’s behavior or decisions makes another member more likely to exhibit similar behavior and make similar decisions. The measurement of such peer effects is essential for better understanding of social networks and interactions, as well as potential urban and labor policy initiatives. The Manski model (1993,2000) has been the foundation for much empirical investigation since. Can the effects of social interactions be identified using econometric models to start with, and if so, what are the sources of the spillover effects then observed? Manski’s challenge, deemed the reflection problem, was that inference can not typically be determined because of the nature of social effects: peers acting simultaneously prevent identification of exogenous effects (the influence of other group members’ characteristics) from endogenous effects (the influence of other group members’ outcomes). Furthermore, any observed or unobserved structural or environmental factors that are shared among members, or lead to their network forming in an endogenous manner, result in correlated effects that will prevent estimation in conventional linear models. The extent to which individuals choose their peers and networks determines how they also ”choose their neighborhood effects” (Ioannides, 2013). 3

The initial hurdle in many analyses of networks is to isolate the correlated effects from the social effects. Several studies handle the correlated effects as group-specific fixed effects. Recent methodological innovations have proposed ways that then uniquely distinguish the remaining social effects into the two components, endogenous and exogenous/contextual, that researchers are primarily interested in. The Lee (2007) model addresses the reflection problem and allows separate identification of the endogenous and exogenous peer effects by virtue of differing group sizes. Typically fixed effects are assigned to a level higher than the unit of observation: industry fixed effects for companies or school fixed effects for students, for example. Following the methodology of Lee (2007), Lin (2010) and Boucher, Bramoull´e, Djebbari and Fortin (2014) include fixed effects at the reference group level for their cross-sectional data. The spatial auto-regressive model with group fixed effects excludes the individual from his own reference group and as a result allows the peer effects to be separated from any unmeasured correlated effects. Thus far, many studies have focused on networks within school settings, due to the obvious nature of interest in peers’ effects on academic outcomes as well as how amenable certain datasets are to these models. The Add Health dataset has been a boon for studies of social networks’ effects because of the way it details the composition of social networks – namely, which fellow students any individual student considers to be his or her friends.5 Lin (2010) creates a spatial weights matrix, Wr , where the reference group, r, is the student’s school-grade level. All students in the group who are listed as respondent i’s n friends received a spatial weight of 1/n if listed as a friend and zero otherwise. This narrows the peer group traditionally used in such studies from the school-grade level to the circle of friends as designated by the individual student. The group fixed effects term will control for the factors encountered by all students in the same grade of the same school. Of all possible specifications, her full model with fixed effects has the greatest log likelihood value and presents clear evidence for a social multiplier via a significant estimated coefficient on the endogenous effect. Specifically, the social multiplier indicates an increase of one standard deviation in peers’ GPA raises own GPA by 0.221 (the mean GPA is 2.817). Calvo-Armengol et al. (2009) use several designs of network connectedness among students who list each other as friends and estimate a parameter which combines peer effects and correlated effects. They find that a one standard deviation increase in their primary index of network centrality correlates to approximately 7 percent of a standard deviation in academic performance (for the benefit of comparison, the same effect from parental education is 17 percent. Beyond academic results, the Add Health dataset has also been utilized for several other interesting outcomes, only sometimes related to the students themselves. Weinberg (2007) examines the sorting that takes place within reference groups and models the way individuals react in response to changes in group behaviors and attributes as evidence of assortative matching. He emphasizes the role of the feedback effects that social interactions have – how associating with people of similar behaviors reinforces one’s own tendency to exhibit such behaviors. Using Add Health data, he explores how students create and choose their social spaces within groups of different sizes. There is a rich literature on the role of peer effects in location choices and housing markets, consistently reinforcing strong externalities based on spatial models. Individuals prefer to live near others like themselves, and housing prices and investments also follow clear spatial patterns. Ioannides and Zabel (2003,2008) show strong evidence of a social multiplier on housing structure improvements, transmitted via peer effects consisting of both psychological and financial motivations. Rossi-Hansberg, Sarte and Owens III (2010) use an urban revitalization program to show marked positive spatial externalities on land values from investment in existing housing. Patacchini 5

The National Longitudinal Study of Adolescent Health combines survey data on students’ physical, psychological, social, and economic status with contextual data on family, neighborhood, school, friendships, and peer groups.

4

and Venanzoni (2014) use the home follow-up survey component of Add Health and the assumption that adolescents’ social contacts are a reasonable approximation for their parents’ social contacts to estimate the role of peer effects on the demand for housing quality. Ajilore (2014) shows the existence of peer effects and family structure on sexual activity and risky sexual behavior. The indirect effects generally match the existing literature’s findings on adolescent sexual behavior. In their work on obesity and overweight status, Alijore, Amialchuk, Xiong and Ye (1974) interpret the contextual effects as social norms, and find these to be more significant than any social multiplier in influencing obesity. The study of social network interactions is particularly salient to obesity research, as there has been a rapid increase in work trying to identify how obesity prevalence is related to both reactions to the behaviors of others and in adaptations to social norms. The use of the Add Health and the Framingham Heart Study datasets have allowed researchers to integrate not only social interactions in their studies but also hypotheses on physical and social proximities.6 Boucher et al. (2014) use a large adminstrative dataset of standardized subject test results from fourth- and fifth-grade students in Quebec. Lacking information on social interactions between students, they propose a student’s reference group to be fellow students in the same school who have all taken the same courses granting eligibility to take these subject tests necessary for graduation. The average size of these school cohort groups is relatively small, which poses a potential problem for identification, but the dispersion is wide, which helps the identification cause. Boucher and his co-authors find evidence of some peer effects. Primarily, the social multiplier is significant for math test scores, but not in the other subjects analyzed in the study (science, history, and French). The policy takeaway here is that increased resources devoted to math learning have the greatest potential for a spillover effect among fellow students. In some cases, academic settings have also been a source of natural experiments that allow identification of peer effects. If individuals are randomly sorted into groups, then some measure of exogeneity is secured. Random assignment of roommates in college has been used to test for the existence of network effects at various levels such as room, floor, and dormitory. Sacerdote (2001) and Zimmerman (2003) study the role of peer effects on roommates and dormmates, estimating various academic and social outcomes as functions of the characteristics and behaviors of themselves and their co-residents. Some significant results emerge, particularly in terms of peers’ effects on grades. The findings on other student outcomes (such as joining a fraternity or choice of major) are mixed, and vary depending on the outcome and reference level the observed student is compared to. Spatial autoregressive models accounting for fixed effects among groups is not the only way to identify peer effects. Some studies have also relied either on datasets where some ad hoc group is created based on observational data where some group features can be reasonably deduced.7 This latter type of social interactions literature inspires the manner of group selection in this study. For example, Aizer and Currie (2004) look at the utilization of publicly-funded prenatal care among women in California between 1989 to 2000, and define a pregnant woman’s reference group as new mothers in the same zip code and race/ethnicity group. This choice of network implicitly assumes that pregnant women are most likely to be influenced by new mothers from the same geographic and ethnic group regarding their takeup of publicly-funded prenatal care. After demonstrating 6

The vigorous debate surrounding the results, data, and econometric specifications of health outcomes and social networks is briefly reviewed in section 2.7.2.3 of Ioannides (2013). 7 In addition, controlled experiments are another profilic source for studies measuring peer effects. Randomized field experiments provide opportunities to better understand social interactions and neighborhood effects. Also, laboratory experiments intersect with expanding studies within behavioral economics, advancing economic understanding of social interactions as well.

5

the significant correlations for takeup within these groups, their detailed data lets them theorize on the source of the network effects. They conclude that, at least among their sample, network effects spread not from information sharing among local co-ethnics as much as from the way the hospitals and health institutions treat members of different groups (or potentially how individuals from different groups respond to equivalent treatment). Section 2.2 reviews more examples of the transmission of peer effects through ethnic networks in terms of job accessibility and labor market outcomes. The motivating interest is in the social multiplier and the endogenous effect that is exclusively responsible for producing it. After accounting for non-spatial factors such as ethnicity, age, sex, education, income, immigration status, and other exogenous characteristics, what social effects, in the form of externalities from interactions, remain? If present, how might the interactions produce a social multiplier?

2.2

Ethnic Concentration and Spatial Distribution of Networks in Urban and Suburban Areas of Cities

Peer effects manifest themselves through networks that exist over geographic and social spaces. Within geographic networks, social interactions are typically highly localized. This can have positive and negative effects: frequency of interactions become more common and information sharing less costly, but if segregation, discrimination, or environmental constraints exist, then network and neighborhood effects may in fact be counter-productive or restrictive. On the other hand, such effects may provide resiliency and flexibility for workers in the face of disadvantageous employment circumstances. Racial/ethnic and residential sorting implies the existence of job market networks, through which we can investigate potential social and neighborhood effects. These effects and other social phenomena such as discrimination can influence labor market outcomes in both urban and suburban areas. The spatial mismatch hypothesis has long been considered a contributing factor to adverse labor market outcomes among minorities. This argument holds that the unequal spatial distribution between job availability and residential locations leads to higher unemployment rates and lower wages for minority workers. Combined with segregation and discrimination in housing and employment, these labor market outcomes are sustained and may be further exacerbated by ongoing trends in employment shifts from urban to suburban regions. Distance to jobs, however, is not the only explanation for these disequilibria, since they exist both in cities where minority workers live close to jobs and in cities where they live far away from jobs. To examine this further, various hypotheses have been proposed to explain urban labor market matches and outcomes. Zenou (2009) models social networks in three configurations that outline the transmission of residential location and social interactions into positive and negative labor market outcomes. Though these models differ in their mechanics, they all emphasize the role of strong and weak ties that allow individuals to potentially exploit their social relationships for improved resources and work opportunities. These social network descriptions are complementary to spatial mismatch. Other studies have proposed additional, sometimes overlapping, explanations for poor labor market outcomes, particularly among minority or immigrant communities: Hellerstein, Neumark and McInerney (2008) propose the notion of racial mismatch – that job availability in a particular area does not matter as much as the availability of jobs into which racial minorities are often employed. Car ownership and commuting are significant determining factors in differential labor market outcomes as well. Ong and Miller (2005) focus on Los Angeles and discuss the notions of “auto-mismatch” and “transport mis-match,” where workers’ lack of access to automobiles is a major obstacle to better labor market prospects. 6

In further research, Hellerstein, Kutzbach and Neumark (2014) call attention to the mixed record of findings regarding networks stratified by race/ethnicity. They conclude that labor market networks have persistent spatial features that go beyond common race or ethnic group measurements. When testing for general residence-based networks, significant positive effects are found for both job turnover and earnings. However, once the groupings of workers who are more connected to neighbors of the same race/ethnic strata are included, the effect on earnings becomes negative (and intermittently significant). These results suggest that workers are preferring non-wage amenities or are constrained by institutional or structural circumstances faced collectively. A large body of research supports the existence of labor market networks within PUMAs and other smaller neighborhood-levels, stratified to some extent by race and ethnicity; the industryPUMA networks created here are an alternative benchmark to some of the networks discussed above or to examples of ethnic enclaves (e.g., Edin, Fredriksson and Aslund (2003), Zhu, Liu and Painter (2014)).

3

Data

I use the public use microdata from the 2000 Census 5% sample and the 2009-2011 American Community Survey (ACS) 3-year sample. These are made available by the IPUMS-USA database (Ruggles, Alexander, Genadek, Goeken, Schroeder and Sobek, 2010).8 These datasets provide detailed demographic and socio-economic data at the individual and household levels, and can be used for estimates of actual population figures. The Census Bureau compiles these microdata for Public Use Microdata Areas (PUMAs), which comprise geographic areas of at least 100,000 people. This is the smallest geographic unit for which microdata is available. The 2009-2011 ACS data is the most recent data that maintains identical PUMA boundaries with the 2000 Census geography. Within each social network, it is assumed that individuals influence each other equally. These peer effects are endogenous or contextual, with Zenou (2009) describing them as ”an average intragroup externality that identically affects all members of a given group” (p377). A crucial part of this literature is then how the boundaries of the social network are defined and how the social structure that comprises it is determined. Building on the findings from the literature discussed above, I create groups based on the intersection of industry and place of residence.9 This is how the social network is defined for this study. There is an expectation of significant social interaction effects within these networks, or reference groups. The nature of the empirical framework assumes that members of each group interact within groups, but not between. That is, the labor market outcome of individual A in a given industry-PUMA group influences the labor market outcome of individual B in the same industry-PUMA group, and vice versa, but not the outcomes of any member of a different industryPUMA group. After excluding agriculture, mining, and those in active-duty military, industries were organized into 45 categories as outlined in Table A1. The industry-PUMA group incorporates a measure of both geographic and social spaces, via the individual’s PUMA of residence and the individual’s industry. Since each individual labor market outcome is affected contemporaneously by the labor market outcomes of their peers but by none 8

The remainder of this paper refers to the pooled 2009-2011 ACS data as the 2010 Census. After multiplying each industry and PUMA group, a small number of groups are dropped if they were singletons or if such groups included individuals who all belonged to the same qualitative measure, i.e., each was female or Latino. The deviations from group mean transformation for such groups would generate columns of zeros in the explanatory variable matrix. 195 groups are dropped out of 4950 total for Los Angeles, 2000, 305 out of 4950 for Los Angeles, 2010; 81 out of 1530 for Houston, 2000, 97 out of 1530 for Houston, 2010; 77 out of 1485 for Miami, 2000; 154 out of 1485 for Miami, 2010. 9

7

outside of it, the goal is to identify the social effects due to peers as well as the correlated effects stemming from the similar environments that group members face. As discussed in the next section, these can be treated as component fixed effects. This consideration is also important in regards to the nature of ethnic networks within neighborhoods, as the fixed effects can address some of the endogeneity of location choice. This study analyzes three metropolitan areas of different size and nature of race and ethnicity networks. Los Angeles, Houston, and Miami all experienced steep and sustained growth in their Latino and foreign-born populations over the last 50 years. Singer (2004) categorizes United States metropolitan areas along a continuum of their status as immigrant gateways. Los Angeles, Houston, and Miami are among the primary examples of “post-World War II” gateways.

4 4.1

Empirical Framework The Reflection Problem

The linear-in-means model has been widely used in the social interaction literature since Manski (1993). The reflection problem occurs when endogenous (behavioral) effects and exogenous (contextual) effects can not be separately identified in a linear-in-means model, and can be formally stated as Yir = λ0 E(Yr |r) + β10 Xir + β20 E(Xr |r) + ir .

(4.1)

Here, Yir denotes the outcome of individual i from group r with characteristics Xir , E(Yr |r) is the group mean outcome, and E(Xr |r) is the group mean characteristics. Furthermore, λ0 measures how an individual’s outcome is affected by her peers’ outcomes, whereas β20 captures the influences of the peers’ exogenous characteristics. The reduced form of (4.1) is given by   λ0 β10 + β20 Yir = β10 Xir + E(Xr |r) + ir . (4.2) 1 − λ0

In (4.2), only β10 and λ0 β10 +β20 / 1−λ0 can be identified. One cannot separate the endogenous effects coefficient λ0 from the coefficient of the contextual effects β20 . Another difficulty in a social interaction estimation is the separation of social effects from other confounding effects. Formation of groups is not typically random, after all. In this setting, individuals select their neighbors, neighborhoods, and industries to various degrees. Similarities in the outcomes of group members may be due to the effects of unobservables or common variables faced by the same group members. These are called “correlated effects” in Manski (1993). In this model, the errors are likely to consist of two components: ir = αr + εir ,

(4.3)

where αr represents characteristics that are common to all group members, usually unobserved, and the εir s are innovations. If the true error terms are as in (4.3), ignoring αr will result in omitted variable bias; that is, if αr and the outcome are positively correlated, then the parameters representing social effects will be overestimated. Since identification is not possible in (4.2), Lee (2007) proposes a spatial autoregressive framework in which identification is feasible under certain conditions. Integrating social interactions and spatial models has become increasingly prevalent in empirical social interaction models. Lee (2007) 8

considers the following SAR model in a social group interaction setting: Yr = λ0 Wr Yr + X1r β10 + Wr X2r β20 + lmr αr + εr , r = 1, 2, . . . , R,

(4.4)

where X1r and X2r are, respectively, mr × k1 and mr × k2 matrices of exogenous variables, which   0 1 may or may not be the same in both matrices. Wr = mr −1 lmr lmr − Imr is the mr × mr nonstochastic network matrix, mr is the number of observations in the rth group, and R is the total number of groups. By the definition of Wr , an individual i in group r is equally affected by the members of the group. Defining Wr in this way is preferred when there is no information on how members are linked in a group. However, when data is available on the formation of groups, Wr may be asymmetric and may have zero off-diagonal elements. Note also from Wr that individual i is not included in the calculation of the group mean outcome and group mean characteristics for himself. This is not the case in Manski (1993) as reproduced in (4.1). In (4.4), Wr Yr represents the average outcome of group r, and it is the source of the endogenous effect in the model. The parameter λ0 captures the endogenous social interaction effect (in this study, this is the impact of peers’ wages on one’s own wages) and αr captures the group specific fixed effects. Moreover, Wr X2r stands for the contextual (exogenous) effects. In this study, these contextual effects include socio-economic and demographic variables such as number of vehicles in the household, experience, years of education, and indicators for sex, married, having at least one child, race/ethnicity categories, years in the United States, and English ability. The estimation of (4.4) is plagued by the fact that the number of group specific fixed effects increases as R (and therefore n) increases.10 The common approach to address this problem is to transform the model in such a way so that the updated model does not involve the group specific fixed effects. Similar to the panel econometric literature, Lee (2007) derives two different specifications from (4.4): (i) the between specification, and (ii) the within specification. The between specification for group r is given by 1 0 1 0 1 0 1 0 1 0 1 0 lmr Yr = λ0 lmr Wr Yr + lmr X1r β10 + lmr Wr X2r β20 + lmr lmr αr + l εr . mr mr mr mr mr mr mr (4.5) 0

0

Since lmr Wr = lmr , (4.5) can be further written as Yr = X1r

β10 β20 αr εr + X2r + + , r = 1, 2, . . . , R. (1 − λ0 ) (1 − λ0 ) (1 − λ0 ) (1 − λ0 )

(4.6)

Note that in (4.6), αr and λ0 cannot be separately identified in the between specification. Estimates β10 of the parameters, such as (1−λ , could be obtained, but β10 , β20 , and λ0 could not isolated, so 0) the between specification must be discarded. 0 To derive the within specification, let Jr = Imr − m1r lmr lmr be the matrix that creates deviations
0 from the group mean. Note that Jr lmr = Imr − m1r lmr lmr lmr = lmr − lmr = 0. This is the process by which group specific fixed effects will be netted out. Premultiplying (4.4) with Jr yields the 10

This is also known as the incidental parameter problem.

9

following within specification Jr Yr = λ0 Jr Wr Yr + Jr X1r β10 + Jr Wr X2r β20 + Jr lmr αr + Jr εr λ0 β20 =− Jr Yr + Jr X1r β10 − Jr X2r + Jr εr . mr (0) mr (0)

(4.7)

Since αr is eliminated in (4.7), the identification of λ0 , β10 , and β20 may be available from the within specification. However, OLS would provide inconsistent estimates at this point because of the presence of the endogenous variable Yr on the right-hand side. A nonlinear estimator must be used, such as the maximum likelihood estimator. For observation i in the rth group, (4.7) can be written as   λ0 β20 (Yri − Yr ) = (X1ri − X1r )β10 − (X2ri − Xr2 ) 1+ + (εri − εr ). (4.8) mr (0) mr (0) P Then, with mr (λ0 ) = mr − 1 + λ0 and m = R1 r=1 mr (m represents the average group size), (4.8) can be re-written as: mr (λ0 ) m β20 (Yri − Yr ) = (X1ri − Xr1 )β10 − (X2ri − Xr2 ) + (εri − εr ). mr (0) mr (0) m

(4.9)

   0 0 0 Following Lee (2007), let Zri = X1ri , − mm X and δ = β , β /m . Then, the within 2ri m 1 2 r (0) specification can be further represented as mr (λ) (Yri − Yr ) = (Zri − Zr )δm + (εri − εr ). mr (0) Denoting Yri − Yr by Yri∗ , it follows from (4.10) that

(4.10) mr (λ) ∗ mr (0) Yri

∗ δ + ε∗ and Y ∗ = = Zri m ri ri

mr (0) ∗ mr (λ) Zri δm0 +

mr (0) ∗ mr (λ) εri ,

where the terms with stars are defined similarly. Transforming the error terms of the model, Pmr ∗ the model with Jr induces linear dependence among ∗ because i=1 εri = 0. Therefore the components of the vector (εr1 , . . . , ε∗r,mr ) are linearly dependent. Thus, only mr − 1 linearly independent components are needed to write the likelihood. Thus, Lee (2007) writes the last term as a linear function of the first mr − 1 terms, and delineates the sample log-likelihood as ln(L) = C +

R X

R(m − 1) ln(σ 2 ) 2 r=1 0   R  1 X mr (λ) ∗ mr (λ) ∗ ∗ ∗ − 2 Y − Zr δm Y − Zr δm . 2σ mr (0) r mr (0) r mr (0) ln(mr (λ)) −

(4.11)

r=1

For computational simplicity, the concentrated log-likelihood is obtained by concentrating out δ and σ 2 from the log-likelihood, with the resulting concentrated log-likelihood expression11 : 11

Where C ≡ − ln(2π) R(m − 1) + 2

1 2

PR

r=1

ln(mr ) −

PR

r=1

mr (0) ln(mr (0)).

10

ln(Lλ ) = C +

R X r=1

mr (0) ln(mr (λ)) −

R(m − 1) ln(ˆ σ 2 (λ)), 2

(4.12)

which depends on only one parameter, λ.12 Variation in group size is the crucial component of identification. Lee (2007)
notes that when 2 ∂ ln σ ˆ (λ) 2 = m+λ−1 . Thus, mr = m (meaning there is no variation in group size) for r = 1, 2, . . . , R, ∂λ PR m(0) R(m−1) ∂ ln(Lλ ) 2 = r=1 m(λ) − ∂λ 2 m+λ−1 = 0 and the concentrated maximum likelihood estimator breaks down. Therefore, identification is contingent on the variation in mr across R groups. The identification conditions outlined in Lee (2007) are not easy to check in empirical applications. Bramoull´e et al. (2009) provide conditions for the identification of the model in (4.4), with and without the correlated effects. In the absence of correlated effects, if there is variation in group sizes and if λ0 β10 + β20 6= 0, then endogenous effects are identified.13 To see
why this condition works, let Wn = D W1 , . . . , WR , where D(·) is the matrix operator that 0 0
0 generates a block-diagonal matrix from a given list of matrices. Also, let Yn = Y1 , . . . , YR ,

0 0
0 X1n = D X11 , . . . , X1R , X2n = D X21 , . . . , X2R , and εn = ε1 , . . . , εR . Bramoull´e et al.
(2009) show that when X1n = X2n = Xn and (In , Wn , W2n ) are linearly dependent, E Wn Yn |Xn is perfectly collinear with (ln , Xn , Wn Xn ) and no instrument is available for the Wn Yn term. However, if In , Wn , and W2n are linearly independent, one can see from the reduced form of the model that (W2n Xn , W3n Xn , . . .) can be used as instruments for Wn Yn , i.e., the restrictions imposed by the network structure yields identification. If some individuals in a group are not linked to their neighbors’ neighbors (if there exists intransitive triads in the network), then In , Wn , and W2n would be linearly independent. In that case, neighbors’ neighbors’ characteristics can be used to instrument for the outcomes of an individual’s neighbors. Note that if there is variation in the group sizes in Wn , then the second condition (In , Wn , and W2n are linearly independent) and identification follows if λ0 β10 + β20 6= 0. In the presence of correlated effects, identification requires both that λ0 β10 + β20 6= 0 and (In , Wn , W2n , W3n ) are linearly independent.14 Due to the elimination of correlated effects first, the identification condition is more stringent and requires existence of an intransitive tetrad so that neighbors’ neighbors’ neighbors’ characteristics can be used to instrument for an individual’s outcomes.

4.2

Results

I begin by estimating models with composite peer effects, with and without the group fixed effects. These models first restrict the endogenous effect coefficient to zero and measures only the exogenous (or contextual) effects, and then restricts the exogenous effects to zero, measuring only the endogenous effect. The final estimations, again with and without the group fixed effects, integrates both types of social effects. The inclusion of industry-PUMA group (or correlated) fixed effects should capture information otherwise specific to this intersection that affects workers’ wages. In particular, I expect these controls should reduce the size of any social effects, since a substantial network effect is channeled through the industry-PUMA group. The own and contextual effects are ˆ See (Lee, 2007) for the expressions for δ(λ) and σ ˆ 2 (λ). See Proposition 2 in Bramoull´e et al. (2009). 14 See Propositions 4 and 5 in Bramoull´e et al. (2009). 12 13

11

drawn from assorted socioeconomic and demographic variables available through the Census/ACS that are useful for wage estimations. Why should integrating correlated effects result in such drastic changes? If information sharing takes place within these industry-PUMA networks and influence labor market outcomes, they should have a significant effect. Moreover, if information sharing is characterized by some of the types of networks described in the literature review and individuals are strongly linked to those similar to themselves, then variables for race/ethnic groups and foreign-born should be prominently affected. 4.2.1

Los Angeles

Table 7 presents the 2000 Census cross-sectional results for Los Angeles. Columns (1) and (2) are the models that include an individual’s own characteristics and his or her peer group’s characteristics. Column (1) is estimated without the correlated effects, and column (2) includes them. The initial model shows significant contextual effects on one’s wage. These exogenous effects notably change throughout after inclusion of the correlated effects. In certain cases, the industry-PUMA grouping used here may be a wide net, capturing many unobserved factors its members have in common. In fact, many contextual regressors lose their significance, have notable magnitude shifts, or even switch signs. The household number of vehicles, experience, female, married, black, and Asian, and arriving to the U.S. between 10 and 20 years ago variables all become insignificant. Having at least one child remains significant, but changes from positive to negative. Having arrived in the country less than 10 years ago remains positive and significant (from a 1% to 10%level), but decreases in magnitude. White is the omitted race/ethnicity category – Latino retains its significant impact though decreasing in magnitude. Columns (3) and (4) display the results with the exogenous/contextual effects restricted to zero, so the model uses only the endogenous factor and the individual’s own characteristics. In the transition from column (3) to (4), the correlated effects sweep away a large amount of omitted information. For one, the social effect that is measured here via the endogenous effect (or the mean outcomes of the industry-PUMA group) declines from 0.803 to 0.237 (and losing significance). Some of the individual race/ethnicity variables have changed as well: black becomes negative and significant and Latino switches from positive and significant to negative and significant. Since columns (1)-(4) represent the mixed effects of contextual and endogenous factors, no clear inference of the peer effects can be determined. This is an illustration of the “reflection problem.” Overall, moving from the odd to the even column within each specification demonstrates the considerable impact of the correlated effects. Furthermore, there is a significant increase in the log-likelihood within each of these pairs, indicating the noteworthy improvement of the models integrating for group correlated effects over those without. Again, the correlated effects are accounting for the unobserved variables that are occuring within the local industry-PUMA networks. Moving to the full model in columns (5) and (6), it is crucial to note the changes that occur in sign and significance among several factors between the two specifications. It is clear that estimation of peer effects can be severely biased when group fixed effects are not offset. Again, the log-likelihood value shows the column (6) model provides the best fit. The endogenous coefficient (the spillover effect) declines by about three-quarters and is no longer significant. The coefficients on individual covariates are all resistant to the inclusion of the correlated fixed effects; this was not the case in the prior specifications, lending support to the classification of peer and correlated effects used here overall. Among the contextual effects, most lose their significance after the inclusion of correlated effects. The important group determinants now are having at least one child, years of education, and speaking no English (the omitted category for language status is English as a native language). 12

The English ability categories are open to interpretation – the coefficient demonstrates that the industry-PUMA reference group’s low English language skills have a positive and significant impact on one’s wage, but that one’s own English deficiencies have a negative and significant repercussions on one’s wage. For those in the average industry-PUMA group, the native English speaker has a clear wage premium over others in his group. For this cross-section of Los Angeles, 2000, black workers face an own wage penalty of 11.2%. The contextual coefficient on having black peers in the industry-PUMA group is -0.362. After accounting for the group fixed effects, the network effects for black workers are decidely unfavorable within the analysis for this city and year. Table 8 displays the results for Los Angeles using the 2010 data. Isolating the preferred model in column (6), the results for the spillover effect, individual coefficients, and most contextual coefficients are similar. The most prominent changes between the 2000 and 2010 results appear for one own effect (speaking no English) and a small number of group contextual effects: Latino status and the English-speaking categories. Why might the peer effects from Latinos in one’s reference group shift from insignificant in 2000 to positive and significant in 2010? The industry-PUMA definition remains the same for both cross-sections. Prior to the inclusion of group correlated effects (column (5) within Tables 7 and 8), this factor was positive and highly significant both in 2000 and 2010, so the even after the unobserved characteristics are absorbed by the industry-PUMA group fixed effects, this provides some evidence for Latino network support. The positive and significant effects from the contextual English categories (none, poorly, or well, as opposed to native speakers) imply that having nonnative English speakers in one’s industry-PUMA group augment one’s wages. 4.2.2

Houston

Table 9 displays the estimation results for the cross-section of Houston 2000. As with Los Angeles, most individual characeristics are resistant to the inclusion of the correlated effects. Some intriguing exceptions within the model excluding contextual effects, are the indicators for Latino and for having arrived between 10 and 20 years ago: in the transition from columns (3) to (4), the former maintains significance but switches signs to negative upon accounting for the group fixed effects, while arrival between 10 and 20 years ago becomes negative and significant. The full model in column (6) is again the preferred model, as it distinguishes between the endogenous and exogenous social effects and also has the greatest log-likelihood. Overall, there are some notable differences from the Los Angeles results. In Los Angeles, there is only one instance where integrating the correlated effects changes a contextual coefficient from being insignificant to significant (this does not occur for Miami at all, as discussed in the next subsection). For Houston in 2000, this does occur for years of education (in 2010, the coefficient is strongly significant in both columns (5) and (6) but its magnitude rises nearly three-fold). Education is the predominant peer factor for wages within Houston in both years – any factors absorbed by the correlated effects still leave it as a strong peer effect. Additionally, the spillover, or endogenous, coefficient is negative and significant (having switched sign after correlated effects were used). This suggests that the average wage of the workers in one’s reference group has a negative impact on another’s observed wage. Negative spillover effects might be caused by a labor supply shock. However, the spillover effect undergoes the same transition in the 2010 cross-sectional results (Table 10), although the final result is insignificant. It is not evident that Houston experienced any sort of positive labor supply shock prior to 2000 or during the years surrounding the ACS surveys around 2010. A possible explanation for the negative endogenous effect, which can also be interpreted as the social multiplier, is the trade-off some workers make for non-wage amenities in their job search 13

and wage outcomes. If, after accounting for all unobserved factors within the industry and PUMA intersection, a worker’s peers’ wages drives his own wages lower, then this might be evidence for such sorting. Table 10 also reports interesting contextual effects for English-speaking ability. As opposed to 2000, where the coefficients for speaking poor or no English were negative or insignificant, they are now each positive and significant. Why should English exogenous effects become so much more important in 2010, while the endogenous factor diminishes? The own coefficients for no or poor English also lose standard significance in 2010. The interpretation for Latinos is not necessarily clear, as their own wage penalty in this model doubles from 2000 to 2010 (from 6.4% to 12.8%), but the influence of having Latino peers within the industry-PUMA group is insignificant. 4.2.3

Miami

For Miami, there are again marked differences between the full models as estimated in columns (5) and (6) when using the group fixed effects. The social multiplier changes from positive and significant to insignificant (positive in 2000 and negative in 2010). Most contextual effects lose their impact – only black and education maintain a positive effect in both cross-sections (in 2010, female has a negative effect and no English has a positive effect). The own variables are completely resistant to the inclusion of the correlated effects. Most of the coefficients have expected signs, with race/ethnicity, poor or no English, and recent arrival all being negative race. The range of results for the peer effects among the three cities highlighted here illustrate that the effects of social interactions vary widely among metroplitan areas and along different demographic and socio-economic characteristic group characteristics. Like Houston and Los Angeles, Miami’s full model features several variables that have opposite signs between the individual cluster and the contextual cluster. If the race/ethnicity, native status, or English ability of an individual and his peers influence his wage in opposing directions, then perhaps workers in the same industry-PUMA group are more likely to be complementary rather than substitutes for one another. The strong influence consistently seen for education, after netting out the correlated effects, might be suggestive evidence for this as well. Borjas (2014) surveys the most recent findings on the existence of complementarity between immigrants and natives, for example, and describes the evidence on elasticity between workers within the same age and education grouping. Within an industry-PUMA group, the range of potential education and English-speaking ability is widely dispersed. Nothing here reveals how the different workers within the same industry and residential PUMA interact, share information, or potentially complement or substitute for one another in the workplace.

5

Conclusion

Social interactions within networks and neighborhoods, and the resulting peer effects on individuals, have been an ascendant area of study. It is well-known that, to different extents, individual outcomes have spillover effects on their peers’ outcomes, and that some outcomes are products of conforming to group characteristics. Recent research has used novel data and methods to address the innate difficulties in estimating and identifying these peer effects. This study is an attempt to measure such influences within an overlap of geographic and social space. I use an intersection of two spaces recognized for their social interactions among peers with like characteristics - industry and residential PUMA - and apply a spatial autoregressive model extended to allow for group fixed effects. This model has been developed and refined in work by Lee (2007), Lin (2010), and Bramoull´e et al. (2009). 14

The formation of industry-PUMA groups is a heavily endogenous process – individuals select their industry, residence, and peers, and thus their peers’ characteristics and outcomes contemporaneously with their neighbors. They then interact with each other and influence each other’s wages and other potential labor market outcomes. Wage outcomes within industry-PUMA groupings are likely to be highly correlated. Treating these intra-group correlations as group fixed effects (analogous to a global transformation in a panel model) allow, given the limited source of data used here, a narrow look at how peer effects potentially operate in different metropolitan areas. The estimation results from the models before and after the inclusion of the group fixed effects are markedly different, demonstrating how much information is missing due to omitted variable bias or endogeneity. Positive spillover effects are not observed, and in some cases, negative ones are. Returns from the exogenous, or contextual, factors are also very limited. In fact, after netting out correlated effects, the impact from peers is so insubstantial that it may be the case that the group correlated effects are subsuming too much. In this case, it contains too much of the intra-group correlations, which may stem from information-sharing interactions that occur between individuals or any unobserved variables that the group members share. Richer data that relates more details about the relationships between individuals would allow one to link them in a more specified reference group (and thus weight matrix). Alternatively, the introduction of correlated effects could change the interpretation. In another paper that creates ad hoc network groups, Aizer and Currie (2004) estimate significant network effects for a group of pregnant women based on ethnicity and location but who were already aware of a treatment program, so information sharing or social effects could not be a cause. Their suggestion is that instead of any spillover effects, the network effects should be interpreted as an exogenous force on these groups of women (specifically, the way in which they are treated by or receive communications from the hospitals they deliver in). If the introduction of fixed effects for industryPUMA groups have such a significant impact, for the cities where there are no endogenous effects, some structural or institutional factor may be involved. For example, employers within an industry may discriminate based on place of residence or using what Manski (2000) called “expectations interactions,” where employers statistically discriminate against prospective employees if they share demographic characteristics with existing employees. I find limited evidence for externalities within social networks, insofar as the industry-PUMA is a useful reference group. If indeed social networks have substantial spillover effects, different policy initiatives may prove advantageous. Employment subsidies or affirmative action policies can be effective in promoting labor market outcomes for minority workers, but are highly dependent on the structure of existing spatial mismatch, so are only welfare-enhancing for cities of certain structures (Zenou, 2009). Knowing more about how information is shared within social networks, and how, if any, spillover effects are present, would help with policy proposals such as regional-based visas for areas experiencing net outmigration, or relocation vouchers to promote employment-based mobility.

15

Table 1: Los Angeles Variable

Obs.

Mean

Std.

Min

Max

274047 37807.26 44690.98

4

325000

vehicles

274047 2.592533 1.832603

1

9

black

274047 .0646312 .2458745

0

1

Asian

274047 .1196693

.324575

0

1

Latino

274047 .3551508 .4785599

0

1

experience

274047 19.33537 11.90554

0

59

2000 wage income

female

274047 .4569034 .4981401

0

1

married

274047 .5470376 .4977835

0

1

child

274047 .4671097

.498918

0

1

arr.<10

274047 .1006178 .3008225

0

1

11≤arr.<20

274047 .1096001 .3465723

0

1

Eng.none

274047 .0314034 .1744056

0

1

Eng.well

274047 .3275788 .4693312

0

1

Eng.poor

274047 .0795375 .2705763

0

1

education

274047 12.95206 3.254613

Groups

4755

# of pumas

110

57.63344 60.55165

0

18

2

564

2010 wage income

186246 50242.24 56305.46 4 410627

vehicles

186246 2.669539 1.633793

1

9

black

186246 .0533273

.224686

0

1

Asian

186246

.3651148

0

1

Latino

186246 .3919977 .4881975

0

1

experience

186246 15.66556 14.18326

0

59

.158398

female

186246 .4770196

.499473

0

1

married

186246 .5413969 .4982847

0

1

child

186246 .4638006 .4986892

0

1

arr.<10

186246 .0751962 .2637084

0

1

11≤arr.<20

186246

.3107094

0

1

Eng.none

186246 .0290906 .1680608

0

1

Eng.well

186246 .3796323 .4852967

0

1

Eng.poor

186246 .0846569 .2783711

0

1

education

186246

Groups

4645

# of pumas

110

.10826

3.196233

0

18

40.09602 43.93844

13.3835

2

410

16

Table 2: Houston Variable

Obs.

Mean

Std.

Min

Max

2000 wage income

78299 37741.27 42944.88

4

324000

vehicles

78299

1

9

2.36533

1.628898

black

78299 .1500785 .3571507

0

1

Asian

78299 .0482637 .2143243

0

1

Latino

78299 .2379979 .4258606

0

1

experience

78299 19.35597 11.66979

0

59

female

78299 .4520747

.497701

0

1

married

78299 .6016169 .4895682

0

1

child

78299 .4864685 .4998201

0

1

arr.<10

78299 .0799372 .2711976

0

1

11≤arr.<20

78299 .0677148 .2512574

0

1

Eng.none

78299 .0203962 .1413521

0

1

Eng.well

78299 .2141151 .4102097

0

1

Eng.poor

78299 .0501667 .2182901

0

1

education

78299 13.03224 2.985278

0

18

Groups

1449

54.03658 61.40861

2

477

wage income

58125 52066.42 57435.52

4

398430

vehicles

58125 2.393325 1.433305

1

9

# of pumas

34

2010

black

58125 .1467355 .3538451

0

1

Asian

58125 .0770409 .2666586

0

1

Latino

58125 .2892215 .4534049

0

1

experience

58125 15.28084

14.1355

0

59

female

58125 .4652559 .4987957

0

1

married

58125 .5915011 .4915605

0

1

child

58125 .4669419 .4989103

0

1

arr.<10

58125 .0782968 .2686403

0

1

11≤arr.<20

58125 .0820989 .2745177

0

1

Eng.none

58125 .0233462 .1510019

0

1

Eng.well

58125 .2622624 .4398683

0

1

Eng.poor

58125 .0579097 .2335746

0

1

education

58125 13.43142 3.060242

0

18

Groups

1433

2

485

# of pumas

40.56176 50.41229

34

17

Table 3: Miami Variable

Obs.

Mean

Std.

Min

Max

2000 wage income

66359 34970.94 41625.98

4

320000

vehicles

66359 2.380581 1.741004

1

9

black

66359 .1803825 .3845086

0

1

Asian

66359 .0207206 .1424485

0

1

Latino

66359 .3905876 .4878858

0

1

experience

66359 20.23725 12.01386

0

59

female

66359 .4806432 .4996289

0

1

married

66359 .5462108 .4978637

0

1

child

66359 .4605705 .4984466

0

1

arr.<10

66359 .1475309

.354637

0

1

11≤arr.<20

66359 .1532573

.360238

0

1

Eng.none

66359 .0299884 .1705565

0

1

Eng.well

66359 .3877696 .4872452

0

1

Eng.poor

66359 .0681445 .2519956

0

1

education

66359 13.26875 2.743561

0

18

Groups

1408

47.12997 47.86338

3

290

wage income

45281 46208.04 52554.21

10

381151

vehicles

45281

1

9

# of pumas

33

2010

2.34891

1.529903

black

45281 .1829907 .3866631

0

1

Asian

45281 .0295488 .1693408

0

1

Latino

45281 .4480687 .4973013

0

1

experience

45281

16.0309

14.26878

0

59

female

45281

.503147

.4999956

0

1

married

45281 .5271306 .4992689

0

1

child

45281 .4484662 .4973427

0

1

arr.<10

45281 .1374307 .3443053

0

1

11≤arr.<20

45281

.3440959

0

1

Eng.none

45281 .0301893 .1711097

0

1

Eng.well

45281 .4407146 .4964783

0

1

Eng.poor

45281

.2636764

0

1

education

45281 13.75906 2.741061

0

18

Groups

1331

2

263

# of pumas

.137232

.075175

34.02029 36.80349

33

18

Table 4: Results for LA 2000 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.010∗∗∗ (0.001) 0.076∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.332∗∗∗ (0.004) 0.125∗∗∗ (0.004) 0.045∗∗∗ (0.004) -0.100∗∗∗ (0.008) -0.078∗∗∗ (0.007) -0.040∗∗∗ (0.006) 0.097∗∗∗ (0.001) -0.190∗∗∗ (0.007) -0.074∗∗∗ (0.006) -0.142∗∗∗ (0.012) -0.059∗∗∗ (0.005) -0.156∗∗∗ (0.009)

-0.017∗∗∗ (0.001) 0.072∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.316∗∗∗ (0.005) 0.136∗∗∗ (0.005) 0.033∗∗∗ (0.005) -0.108∗∗∗ (0.010) -0.087∗∗∗ (0.009) -0.051∗∗∗ (0.007) 0.098∗∗∗ (0.001) -0.182∗∗∗ (0.009) -0.085∗∗∗ (0.008) -0.111∗∗∗ (0.015) -0.045∗∗∗ (0.007) -0.153∗∗∗ (0.011)

0.399∗∗∗ (0.005) 0.212∗∗∗ (0.003) -0.004∗∗∗ (0.000) -0.702∗∗∗ (0.011) -0.548∗∗∗ (0.026) 0.327∗∗∗ (0.026) 0.282∗∗∗ (0.019) 0.077∗∗∗ (0.026) 1.563∗∗∗ (0.025) 0.400∗∗∗ (0.002) 0.166∗∗∗ (0.039) 0.840∗∗∗ (0.036) 0.296∗∗∗ (0.064) -0.688∗∗∗ (0.030) 0.567∗∗∗ (0.054) No -361762.797

0.013 (0.042) 0.020 (0.024) 0.001 (0.001) 0.156 (0.162) 0.094 (0.193) -0.346∗ (0.194) -0.138 (0.320) -0.425 (0.321) 0.902∗∗∗ (0.255) 0.478∗∗∗ (0.020) 0.613∗ (0.332) 0.228 (0.276) 2.036∗∗∗ (0.550) 0.035 (0.250) 0.765∗ (0.401) Yes -335840.178

(3) 0.803∗∗∗ (0.001)

(4) 0.237 (0.173)

(5) 0.703∗∗∗ (0.003)

(6) 0.151 (0.344)

-0.003∗∗∗ (0.001) 0.075∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.299∗∗∗ (0.003) 0.110∗∗∗ (0.004) 0.038∗∗∗ (0.004) 0.007 (0.007) -0.048∗∗∗ (0.006) 0.071∗∗∗ (0.005) 0.097∗∗∗ (0.001) -0.129∗∗∗ (0.007) -0.041∗∗∗ (0.006) 0.006 (0.011) -0.014∗∗∗ (0.005) -0.056∗∗∗ (0.008)

-0.017∗∗∗ (0.001) 0.072∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.320∗∗∗ (0.004) 0.135∗∗∗ (0.004) 0.039∗∗∗ (0.004) -0.105∗∗∗ (0.007) -0.080∗∗∗ (0.007) -0.068∗∗∗ (0.005) 0.090∗∗∗ (0.001) -0.194∗∗∗ (0.007) -0.090∗∗∗ (0.006) -0.147∗∗∗ (0.009) -0.046∗∗∗ (0.005) -0.167∗∗∗ (0.007)

-0.015∗∗∗ (0.001) 0.073∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.323∗∗∗ (0.004) 0.131∗∗∗ (0.004) 0.041∗∗∗ (0.004) -0.103∗∗∗ (0.008) -0.079∗∗∗ (0.007) -0.060∗∗∗ (0.005) 0.091∗∗∗ (0.001) -0.192∗∗∗ (0.007) -0.085∗∗∗ (0.006) -0.145∗∗∗ (0.011) -0.050∗∗∗ (0.005) -0.163∗∗∗ (0.008)

-0.018∗∗∗ (0.001) 0.071∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.322∗∗∗ (0.005) 0.135∗∗∗ (0.005) 0.033∗∗∗ (0.005) -0.112∗∗∗ (0.009) -0.085∗∗∗ (0.009) -0.061∗∗∗ (0.007) 0.095∗∗∗ (0.001) -0.185∗∗∗ (0.009) -0.088∗∗∗ (0.008) -0.118∗∗∗ (0.013) -0.048∗∗∗ (0.007) -0.156∗∗∗ (0.010)

Yes -323700.650

0.130∗∗∗ (0.005) 0.012∗∗∗ (0.003) -0.000∗∗∗ (0.000) 0.014 (0.011) -0.255∗∗∗ (0.024) 0.067∗∗∗ (0.025) 0.158∗∗∗ (0.018) 0.080∗∗∗ (0.025) 0.514∗∗∗ (0.024) 0.056∗∗∗ (0.002) 0.188∗∗∗ (0.037) 0.311∗∗∗ (0.034) 0.197∗∗∗ (0.059) -0.173∗∗∗ (0.028) 0.286∗∗∗ (0.050) No -344144.138

-0.035 (0.044) -0.029 (0.036) 0.001 (0.001) -0.129 (0.188) 0.037 (0.201) -0.379∗ (0.196) -0.362 (0.281) -0.333 (0.319) 0.374 (0.270) 0.301∗∗∗ (0.034) 0.524 (0.339) 0.122 (0.277) 1.660∗∗∗ (0.482) -0.081 (0.243) 0.599 (0.370) Yes -323647.588

No -348805.695

Standard errors are in parentheses.

19

Table 5: Results for LA 2010 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.018∗∗∗ (0.001) 0.029∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.310∗∗∗ (0.005) 0.252∗∗∗ (0.006) 0.175∗∗∗ (0.006) -0.068∗∗∗ (0.012) -0.084∗∗∗ (0.009) -0.071∗∗∗ (0.007) 0.088∗∗∗ (0.001) -0.207∗∗∗ (0.010) -0.074∗∗∗ (0.008) 0.003 (0.016) -0.077∗∗∗ (0.007) -0.063∗∗∗ (0.011)

-0.027∗∗∗ (0.002) 0.031∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.280∗∗∗ (0.006) 0.250∗∗∗ (0.007) 0.145∗∗∗ (0.007) -0.086∗∗∗ (0.014) -0.073∗∗∗ (0.010) -0.077∗∗∗ (0.009) 0.087∗∗∗ (0.001) -0.212∗∗∗ (0.012) -0.090∗∗∗ (0.010) 0.054∗∗∗ (0.020) -0.061∗∗∗ (0.008) -0.048∗∗∗ (0.014)

0.408∗∗∗ (0.006) -0.082∗∗∗ (0.003) 0.002∗∗∗ (0.000) -1.042∗∗∗ (0.014) 0.125∗∗∗ (0.029) 0.981∗∗∗ (0.029) 1.061∗∗∗ (0.028) 0.014 (0.028) 1.325∗∗∗ (0.027) 0.542∗∗∗ (0.002) 0.086∗ (0.047) 0.921∗∗∗ (0.044) 1.720∗∗∗ (0.072) -0.154∗∗∗ (0.030) 1.603∗∗∗ (0.049) No -263331.954

0.055 (0.042) -0.001 (0.019) 0.001 (0.000) 0.040 (0.147) 0.049 (0.170) -0.152 (0.172) 0.472 (0.331) 0.411 (0.261) 1.105∗∗∗ (0.211) 0.519∗∗∗ (0.019) -0.074 (0.316) 0.343 (0.257) 3.637∗∗∗ (0.506) 0.468∗∗ (0.207) 2.125∗∗∗ (0.339) Yes -240881.187

(3) 0.840∗∗∗ (0.001)

(4) 0.350∗∗ (0.175)

(5) 0.686∗∗∗ (0.003)

(6) 0.275 (0.344)

-0.005∗∗∗ (0.001) 0.024∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.263∗∗∗ (0.004) 0.235∗∗∗ (0.005) 0.168∗∗∗ (0.005) 0.080∗∗∗ (0.010) -0.049∗∗∗ (0.008) 0.081∗∗∗ (0.006) 0.098∗∗∗ (0.001) -0.168∗∗∗ (0.009) -0.054∗∗∗ (0.008) 0.227∗∗∗ (0.015) -0.012∗∗ (0.006) 0.101∗∗∗ (0.010)

-0.029∗∗∗ (0.001) 0.031∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.284∗∗∗ (0.005) 0.252∗∗∗ (0.005) 0.150∗∗∗ (0.005) -0.097∗∗∗ (0.009) -0.083∗∗∗ (0.008) -0.107∗∗∗ (0.007) 0.074∗∗∗ (0.001) -0.212∗∗∗ (0.009) -0.100∗∗∗ (0.008) -0.044∗∗∗ (0.012) -0.074∗∗∗ (0.006) -0.107∗∗∗ (0.009)

-0.025∗∗∗ (0.001) 0.030∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.290∗∗∗ (0.005) 0.250∗∗∗ (0.005) 0.157∗∗∗ (0.005) -0.088∗∗∗ (0.011) -0.083∗∗∗ (0.008) -0.095∗∗∗ (0.007) 0.078∗∗∗ (0.001) -0.209∗∗∗ (0.009) -0.092∗∗∗ (0.008) -0.031∗∗ (0.015) -0.075∗∗∗ (0.006) -0.094∗∗∗ (0.010)

-0.030∗∗∗ (0.002) 0.032∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.287∗∗∗ (0.006) 0.248∗∗∗ (0.007) 0.142∗∗∗ (0.007) -0.096∗∗∗ (0.013) -0.078∗∗∗ (0.010) -0.097∗∗∗ (0.009) 0.081∗∗∗ (0.001) -0.209∗∗∗ (0.012) -0.087∗∗∗ (0.010) 0.038∗∗ (0.017) -0.062∗∗∗ (0.008) -0.061∗∗∗ (0.013)

Yes -229552.117

0.146∗∗∗ (0.006) -0.047∗∗∗ (0.003) 0.001∗∗∗ (0.000) -0.136∗∗∗ (0.013) -0.133∗∗∗ (0.027) 0.204∗∗∗ (0.027) 0.399∗∗∗ (0.026) 0.062∗∗ (0.025) 0.489∗∗∗ (0.025) 0.121∗∗∗ (0.003) 0.167∗∗∗ (0.043) 0.363∗∗∗ (0.040) 0.587∗∗∗ (0.066) 0.004 (0.027) 0.581∗∗∗ (0.045) No -248741.521

-0.023 (0.044) 0.007 (0.022) 0.000 (0.001) -0.153 (0.163) -0.127 (0.192) -0.280 (0.181) 0.104 (0.301) 0.232 (0.261) 0.383∗ (0.231) 0.292∗∗∗ (0.032) 0.090 (0.323) 0.482∗ (0.265) 3.092∗∗∗ (0.445) 0.450∗∗ (0.203) 1.701∗∗∗ (0.333) Yes -229474.908

No -253704.237

Standard errors are in parentheses.

20

Table 6: Results for Houston 2000 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.008∗∗∗ (0.002) 0.072∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.373∗∗∗ (0.007) 0.132∗∗∗ (0.008) 0.058∗∗∗ (0.007) -0.190∗∗∗ (0.010) -0.151∗∗∗ (0.019) -0.032∗∗ (0.013) 0.112∗∗∗ (0.001) -0.134∗∗∗ (0.015) -0.028∗ (0.015) -0.119∗∗∗ (0.028) -0.064∗∗∗ (0.012) -0.098∗∗∗ (0.020)

-0.017∗∗∗ (0.002) 0.067∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.355∗∗∗ (0.008) 0.142∗∗∗ (0.009) 0.054∗∗∗ (0.009) -0.183∗∗∗ (0.012) -0.143∗∗∗ (0.021) -0.047∗∗∗ (0.014) 0.115∗∗∗ (0.001) -0.121∗∗∗ (0.017) -0.020 (0.016) -0.097∗∗∗ (0.031) -0.067∗∗∗ (0.014) -0.092∗∗∗ (0.022)

0.734∗∗∗ (0.010) 0.212∗∗∗ (0.005) -0.004∗∗∗ (0.000) -0.813∗∗∗ (0.019) -0.478∗∗∗ (0.046) 0.076∗ (0.046) -0.100∗∗∗ (0.027) 0.380∗∗∗ (0.101) 1.903∗∗∗ (0.069) 0.354∗∗∗ (0.003) 0.145∗ (0.082) 0.071 (0.093) -1.473∗∗∗ (0.166) -0.653∗∗∗ (0.080) 0.157 (0.126) No -101229.031

0.296∗∗∗ (0.070) -0.031 (0.036) 0.001 (0.001) 0.059 (0.251) 0.015 (0.285) -0.170 (0.285) 0.167 (0.378) 0.696 (0.595) 1.205∗∗∗ (0.415) 0.509∗∗∗ (0.022) 0.761 (0.541) 0.501 (0.461) -0.298 (1.012) -0.821∗∗ (0.399) 0.384 (0.647) Yes -92182.568

(3) 0.791∗∗∗ (0.002)

(4) -0.668∗∗∗ (0.166)

(5) 0.722∗∗∗ (0.005)

(6) -0.750∗∗∗ (0.166)

-0.009∗∗∗ (0.002) 0.069∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.331∗∗∗ (0.006) 0.107∗∗∗ (0.007) 0.048∗∗∗ (0.007) -0.102∗∗∗ (0.009) -0.140∗∗∗ (0.017) 0.051∗∗∗ (0.011) 0.112∗∗∗ (0.001) -0.085∗∗∗ (0.014) 0.005 (0.013) 0.021 (0.025) -0.031∗∗∗ (0.011) -0.007 (0.018)

-0.023∗∗∗ (0.002) 0.066∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.351∗∗∗ (0.006) 0.140∗∗∗ (0.007) 0.056∗∗∗ (0.007) -0.186∗∗∗ (0.008) -0.154∗∗∗ (0.016) -0.071∗∗∗ (0.011) 0.103∗∗∗ (0.001) -0.133∗∗∗ (0.013) -0.030∗∗ (0.013) -0.093∗∗∗ (0.019) -0.049∗∗∗ (0.010) -0.097∗∗∗ (0.016)

-0.019∗∗∗ (0.002) 0.068∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.361∗∗∗ (0.007) 0.139∗∗∗ (0.007) 0.058∗∗∗ (0.007) -0.188∗∗∗ (0.009) -0.156∗∗∗ (0.017) -0.061∗∗∗ (0.012) 0.106∗∗∗ (0.001) -0.135∗∗∗ (0.014) -0.030∗∗ (0.013) -0.100∗∗∗ (0.025) -0.053∗∗∗ (0.011) -0.098∗∗∗ (0.018)

-0.020∗∗∗ (0.002) 0.065∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.360∗∗∗ (0.008) 0.142∗∗∗ (0.009) 0.052∗∗∗ (0.008) -0.188∗∗∗ (0.011) -0.143∗∗∗ (0.020) -0.064∗∗∗ (0.014) 0.109∗∗∗ (0.002) -0.128∗∗∗ (0.017) -0.026∗ (0.016) -0.103∗∗∗ (0.024) -0.061∗∗∗ (0.013) -0.093∗∗∗ (0.019)

Yes -88614.167

0.223∗∗∗ (0.010) 0.010∗∗ (0.005) -0.000∗∗∗ (0.000) 0.028 (0.018) -0.235∗∗∗ (0.042) -0.025 (0.042) 0.108∗∗∗ (0.025) 0.245∗∗∗ (0.092) 0.599∗∗∗ (0.064) 0.024∗∗∗ (0.004) 0.139∗ (0.075) 0.057 (0.084) -0.370∗∗ (0.152) -0.175∗∗ (0.073) 0.095 (0.115) No -94897.776

0.152∗∗ (0.071) -0.049 (0.039) 0.001 (0.001) -0.421∗ (0.250) 0.152 (0.291) -0.215 (0.284) -0.180 (0.352) 0.529 (0.582) 0.316 (0.426) 0.304∗∗∗ (0.046) 0.293 (0.534) 0.178 (0.453) -0.666 (0.834) -0.577 (0.387) 0.211 (0.599) Yes -88580.241

No -96205.529

Standard errors are in parentheses.

21

Table 7: Results for Houston 2010 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.017∗∗∗ (0.003) 0.022∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.341∗∗∗ (0.009) 0.290∗∗∗ (0.009) 0.151∗∗∗ (0.009) -0.220∗∗∗ (0.013) -0.131∗∗∗ (0.019) -0.093∗∗∗ (0.014) 0.101∗∗∗ (0.002) -0.186∗∗∗ (0.017) -0.051∗∗∗ (0.017) -0.036 (0.032) -0.085∗∗∗ (0.014) 0.006 (0.023)

-0.031∗∗∗ (0.003) 0.023∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.317∗∗∗ (0.010) 0.282∗∗∗ (0.011) 0.124∗∗∗ (0.010) -0.224∗∗∗ (0.015) -0.123∗∗∗ (0.022) -0.108∗∗∗ (0.016) 0.102∗∗∗ (0.002) -0.161∗∗∗ (0.020) -0.065∗∗∗ (0.018) 0.010 (0.035) -0.072∗∗∗ (0.016) 0.027 (0.026)

0.670∗∗∗ (0.013) -0.103∗∗∗ (0.005) 0.003∗∗∗ (0.000) -1.070∗∗∗ (0.022) 0.420∗∗∗ (0.049) 0.768∗∗∗ (0.051) 0.633∗∗∗ (0.036) 0.273∗∗∗ (0.087) 1.683∗∗∗ (0.070) 0.511∗∗∗ (0.004) -0.853∗∗∗ (0.088) -0.024 (0.090) 1.990∗∗∗ (0.161) -0.292∗∗∗ (0.077) 1.725∗∗∗ (0.122) No -80021.087

0.164∗∗ (0.070) -0.060∗∗ (0.027) 0.002∗∗ (0.001) -0.216 (0.226) 0.162 (0.244) -0.299 (0.257) 0.453 (0.330) 0.608 (0.520) 1.144∗∗∗ (0.356) 0.551∗∗∗ (0.019) 0.199 (0.477) -0.539 (0.435) 3.493∗∗∗ (0.710) 0.138 (0.366) 2.437∗∗∗ (0.565) Yes -72670.009

(3) 0.835∗∗∗ (0.002)

(4) -0.505∗∗ (0.204)

(5) 0.692∗∗∗ (0.006)

(6) -0.580 (0.431)

-0.010∗∗∗ (0.002) 0.017∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.291∗∗∗ (0.007) 0.256∗∗∗ (0.009) 0.144∗∗∗ (0.008) -0.104∗∗∗ (0.011) -0.116∗∗∗ (0.018) 0.019 (0.013) 0.109∗∗∗ (0.001) -0.159∗∗∗ (0.016) -0.031∗∗ (0.015) 0.156∗∗∗ (0.029) -0.042∗∗∗ (0.013) 0.155∗∗∗ (0.021)

-0.035∗∗∗ (0.003) 0.024∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.309∗∗∗ (0.008) 0.274∗∗∗ (0.009) 0.130∗∗∗ (0.008) -0.232∗∗∗ (0.010) -0.132∗∗∗ (0.017) -0.137∗∗∗ (0.013) 0.086∗∗∗ (0.002) -0.162∗∗∗ (0.016) -0.050∗∗∗ (0.015) -0.078∗∗∗ (0.022) -0.074∗∗∗ (0.012) -0.037∗∗ (0.018)

-0.030∗∗∗ (0.003) 0.024∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.321∗∗∗ (0.008) 0.282∗∗∗ (0.009) 0.138∗∗∗ (0.008) -0.230∗∗∗ (0.012) -0.134∗∗∗ (0.018) -0.125∗∗∗ (0.013) 0.091∗∗∗ (0.002) -0.172∗∗∗ (0.016) -0.051∗∗∗ (0.015) -0.071∗∗ (0.029) -0.078∗∗∗ (0.013) -0.028 (0.021)

-0.036∗∗∗ (0.003) 0.024∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.322∗∗∗ (0.010) 0.274∗∗∗ (0.011) 0.119∗∗∗ (0.010) -0.235∗∗∗ (0.013) -0.124∗∗∗ (0.021) -0.128∗∗∗ (0.016) 0.093∗∗∗ (0.002) -0.159∗∗∗ (0.019) -0.060∗∗∗ (0.018) -0.033 (0.028) -0.072∗∗∗ (0.015) 0.003 (0.023)

Yes -69260.516

0.228∗∗∗ (0.012) -0.049∗∗∗ (0.005) 0.001∗∗∗ (0.000) -0.115∗∗∗ (0.022) -0.061 (0.044) 0.140∗∗∗ (0.046) 0.356∗∗∗ (0.033) 0.185∗∗ (0.079) 0.611∗∗∗ (0.065) 0.098∗∗∗ (0.005) -0.157∗ (0.081) 0.002 (0.082) 0.710∗∗∗ (0.147) -0.033 (0.070) 0.590∗∗∗ (0.111) No -75203.946

-0.030 (0.075) -0.017 (0.032) 0.000 (0.001) -0.525∗∗ (0.232) 0.046 (0.283) -0.396 (0.272) -0.118 (0.304) 0.377 (0.515) 0.332 (0.368) 0.273∗∗∗ (0.043) 0.154 (0.487) -0.371 (0.428) 1.745∗∗∗ (0.640) 0.078 (0.352) 1.521∗∗∗ (0.521) Yes -69230.268

No -76773.491

Standard errors are in parentheses.

22

Table 8: Results for Miami 2000 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.007∗∗∗ (0.002) 0.064∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.330∗∗∗ (0.007) 0.134∗∗∗ (0.008) 0.039∗∗∗ (0.008) -0.182∗∗∗ (0.011) -0.054∗∗ (0.025) -0.114∗∗∗ (0.012) 0.100∗∗∗ (0.001) -0.179∗∗∗ (0.012) -0.055∗∗∗ (0.011) -0.259∗∗∗ (0.025) -0.021∗ (0.011) -0.229∗∗∗ (0.018)

-0.017∗∗∗ (0.002) 0.059∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.328∗∗∗ (0.009) 0.142∗∗∗ (0.010) 0.040∗∗∗ (0.010) -0.167∗∗∗ (0.013) -0.086∗∗∗ (0.030) -0.122∗∗∗ (0.014) 0.104∗∗∗ (0.001) -0.205∗∗∗ (0.014) -0.062∗∗∗ (0.013) -0.203∗∗∗ (0.030) -0.006 (0.013) -0.209∗∗∗ (0.022)

0.450∗∗∗ (0.009) 0.245∗∗∗ (0.005) -0.004∗∗∗ (0.000) -0.502∗∗∗ (0.022) -0.281∗∗∗ (0.047) -0.123∗∗∗ (0.046) 0.536∗∗∗ (0.028) 2.175∗∗∗ (0.140) 0.318∗∗∗ (0.049) 0.354∗∗∗ (0.004) 0.817∗∗∗ (0.059) 0.755∗∗∗ (0.060) 0.146 (0.125) -0.265∗∗∗ (0.052) 0.395∗∗∗ (0.095) No -85692.000

-0.003 (0.067) 0.035 (0.037) -0.000 (0.001) -0.432 (0.273) 0.062 (0.286) -0.073 (0.297) 1.160∗∗∗ (0.382) 0.724 (0.907) 0.003 (0.425) 0.533∗∗∗ (0.028) -0.302 (0.416) 0.425 (0.387) 2.536∗∗∗ (0.868) 0.397 (0.395) 1.263∗∗ (0.620) Yes -78566.655

(3) 0.805∗∗∗ (0.002)

(4) 0.252 (0.285)

(5) 0.693∗∗∗ (0.006)

(6) 0.168 (0.936)

-0.003∗ (0.002) 0.062∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.295∗∗∗ (0.006) 0.119∗∗∗ (0.007) 0.031∗∗∗ (0.007) -0.070∗∗∗ (0.009) -0.055∗∗ (0.023) -0.041∗∗∗ (0.011) 0.105∗∗∗ (0.001) -0.161∗∗∗ (0.011) -0.035∗∗∗ (0.010) -0.112∗∗∗ (0.022) 0.028∗∗∗ (0.010) -0.135∗∗∗ (0.016)

-0.017∗∗∗ (0.002) 0.058∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.321∗∗∗ (0.007) 0.142∗∗∗ (0.007) 0.042∗∗∗ (0.008) -0.194∗∗∗ (0.010) -0.100∗∗∗ (0.020) -0.122∗∗∗ (0.011) 0.093∗∗∗ (0.001) -0.200∗∗∗ (0.011) -0.072∗∗∗ (0.010) -0.261∗∗∗ (0.018) -0.014 (0.010) -0.238∗∗∗ (0.015)

-0.014∗∗∗ (0.002) 0.060∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.322∗∗∗ (0.007) 0.139∗∗∗ (0.007) 0.041∗∗∗ (0.007) -0.190∗∗∗ (0.010) -0.086∗∗∗ (0.023) -0.120∗∗∗ (0.011) 0.095∗∗∗ (0.001) -0.193∗∗∗ (0.011) -0.067∗∗∗ (0.010) -0.260∗∗∗ (0.023) -0.016 (0.010) -0.234∗∗∗ (0.017)

-0.019∗∗∗ (0.002) 0.057∗∗∗ (0.001) -0.001∗∗∗ (0.000) -0.337∗∗∗ (0.009) 0.141∗∗∗ (0.010) 0.039∗∗∗ (0.010) -0.176∗∗∗ (0.013) -0.081∗∗∗ (0.028) -0.123∗∗∗ (0.015) 0.098∗∗∗ (0.002) -0.207∗∗∗ (0.014) -0.066∗∗∗ (0.013) -0.218∗∗∗ (0.026) -0.011 (0.013) -0.223∗∗∗ (0.021)

Yes -75166.148

0.148∗∗∗ (0.009) 0.034∗∗∗ (0.005) -0.001∗∗∗ (0.000) 0.061∗∗∗ (0.021) -0.185∗∗∗ (0.043) -0.069 (0.042) 0.303∗∗∗ (0.026) 0.752∗∗∗ (0.130) 0.182∗∗∗ (0.045) 0.046∗∗∗ (0.004) 0.383∗∗∗ (0.055) 0.286∗∗∗ (0.056) 0.249∗∗ (0.115) -0.072 (0.048) 0.287∗∗∗ (0.088) No -80975.994

-0.109 (0.070) -0.041 (0.070) 0.001 (0.001) -0.781∗∗ (0.333) -0.015 (0.323) -0.116 (0.305) 0.776∗ (0.407) 0.929 (0.824) -0.046 (0.446) 0.240∗∗∗ (0.077) -0.305 (0.453) 0.256 (0.398) 1.888∗∗ (0.758) 0.156 (0.387) 0.692 (0.579) Yes -75145.568

No -82028.122

Standard errors are in parentheses.

23

Table 9: Results for Miami 2010 Explanatory Variable Endogenous effect Own vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Contextual effects vehicles experience experience2 female married child black Asian Latino education arr.< 10 10 ≤arr.< 20 Eng.none Eng.well Eng.poor Correlated effects log-likelihood Note:

(1)

(2)

-0.012∗∗∗ (0.003) 0.022∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.315∗∗∗ (0.010) 0.225∗∗∗ (0.010) 0.130∗∗∗ (0.011) -0.170∗∗∗ (0.015) -0.112∗∗∗ (0.029) -0.129∗∗∗ (0.015) 0.089∗∗∗ (0.002) -0.228∗∗∗ (0.015) -0.096∗∗∗ (0.014) -0.115∗∗∗ (0.032) -0.001 (0.014) -0.129∗∗∗ (0.023)

-0.018∗∗∗ (0.003) 0.024∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.272∗∗∗ (0.011) 0.220∗∗∗ (0.012) 0.119∗∗∗ (0.012) -0.162∗∗∗ (0.017) -0.070∗∗ (0.033) -0.111∗∗∗ (0.018) 0.089∗∗∗ (0.002) -0.237∗∗∗ (0.018) -0.107∗∗∗ (0.016) -0.146∗∗∗ (0.037) 0.001 (0.016) -0.165∗∗∗ (0.026)

0.519∗∗∗ (0.013) -0.126∗∗∗ (0.005) 0.003∗∗∗ (0.000) -0.856∗∗∗ (0.027) 0.368∗∗∗ (0.052) 0.250∗∗∗ (0.052) 1.141∗∗∗ (0.035) -0.650∗∗∗ (0.126) -0.252∗∗∗ (0.054) 0.565∗∗∗ (0.004) 0.076 (0.068) 0.548∗∗∗ (0.069) 2.204∗∗∗ (0.136) 0.524∗∗∗ (0.058) 1.844∗∗∗ (0.102) No -62554.447

0.319∗∗∗ (0.070) -0.064∗∗ (0.029) 0.002∗∗ (0.001) 0.469∗∗ (0.228) 0.196 (0.235) -0.087 (0.246) 1.390∗∗∗ (0.320) 0.723 (0.752) 0.306 (0.352) 0.561∗∗∗ (0.023) -0.294 (0.363) 0.165 (0.327) 1.277∗ (0.739) 0.612∗ (0.328) 0.841∗ (0.510) Yes -56239.917

(3) 0.843∗∗∗ (0.002)

(4) -0.017 (0.247)

(5) 0.684∗∗∗ (0.006)

(6) -0.043 (0.714)

-0.004 (0.003) 0.017∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.260∗∗∗ (0.008) 0.201∗∗∗ (0.009) 0.133∗∗∗ (0.010) -0.028∗∗ (0.012) -0.044∗ (0.026) -0.020 (0.013) 0.105∗∗∗ (0.001) -0.202∗∗∗ (0.014) -0.085∗∗∗ (0.013) 0.035 (0.028) 0.028∗∗ (0.012) -0.035∗ (0.020)

-0.028∗∗∗ (0.003) 0.026∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.287∗∗∗ (0.009) 0.213∗∗∗ (0.009) 0.120∗∗∗ (0.009) -0.207∗∗∗ (0.012) -0.090∗∗∗ (0.023) -0.122∗∗∗ (0.014) 0.072∗∗∗ (0.002) -0.230∗∗∗ (0.014) -0.114∗∗∗ (0.013) -0.180∗∗∗ (0.023) -0.015 (0.012) -0.187∗∗∗ (0.018)

-0.023∗∗∗ (0.003) 0.025∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.296∗∗∗ (0.009) 0.217∗∗∗ (0.009) 0.123∗∗∗ (0.009) -0.197∗∗∗ (0.013) -0.098∗∗∗ (0.026) -0.124∗∗∗ (0.014) 0.077∗∗∗ (0.002) -0.229∗∗∗ (0.014) -0.108∗∗∗ (0.013) -0.159∗∗∗ (0.029) -0.011 (0.012) -0.168∗∗∗ (0.020)

-0.025∗∗∗ (0.003) 0.026∗∗∗ (0.001) -0.000∗∗∗ (0.000) -0.286∗∗∗ (0.011) 0.216∗∗∗ (0.012) 0.118∗∗∗ (0.012) -0.184∗∗∗ (0.016) -0.071∗∗ (0.031) -0.126∗∗∗ (0.017) 0.075∗∗∗ (0.002) -0.229∗∗∗ (0.018) -0.106∗∗∗ (0.016) -0.183∗∗∗ (0.034) -0.007 (0.016) -0.193∗∗∗ (0.025)

Yes -53135.245

0.185∗∗∗ (0.012) -0.058∗∗∗ (0.005) 0.001∗∗∗ (0.000) -0.073∗∗∗ (0.025) -0.026 (0.047) -0.007 (0.047) 0.505∗∗∗ (0.032) -0.138 (0.113) 0.002 (0.049) 0.129∗∗∗ (0.005) 0.184∗∗∗ (0.061) 0.264∗∗∗ (0.063) 0.816∗∗∗ (0.123) 0.178∗∗∗ (0.052) 0.694∗∗∗ (0.092) No -58471.578

0.108 (0.079) -0.008 (0.037) 0.000 (0.001) 0.017 (0.310) 0.099 (0.278) -0.070 (0.260) 0.731∗ (0.407) 0.695 (0.683) -0.143 (0.364) 0.117∗ (0.063) 0.013 (0.419) 0.236 (0.369) -0.004 (0.662) 0.268 (0.318) -0.138 (0.483) Yes -53125.483

No -59724.547

Standard errors are in parentheses.

24

Table A1 Industry Categories Construction Manufacturing - food and kindred products Manufacturing - textile and apparel Manufacturing - paper, printing, and allied products Manufacturing - Chemicals, rubber, and allied products Manufacturing - leather products Manufacturing - lumber, furniture, fixtures Manufacturing - metal and machinery industries Manufacturing - transportation equipment Manufacturing - professional and scientific equipment Manufacturing - toys and misc. equipment Transportation - urban transit and trucking Transportation - warehousing, USPS Transportation - utility transportation Communications Utilities Wholesale Trade Retail trade - lumber, hardware Retail trade - department, retail stores Retail trade - food, grocery stores Retail trade - vehicles, gas Retail trade - apparel, shoes Retail trade - furniture, appliances Retail trade - eating/drinking Retail trade - drugs, liquor Retail trade - hobby, crafts, jewelry Retail trade - misc. retail Finance, Insurance Real Estate, Advertising Personnel, business services Automotive and misc. repair services Personal services - households, hotels Personal services - beauty, laundry, misc. Entertainment services Medical offices, hospitals, health facilities Legal services Schools, colleges, vocational, educational services Child and residential care services Museum and membership organizations Engineering, architectural, and surveying services Accounting and research services Management, public relations, and misc. services Public Administration - general government Public Administration - Justice, public order, and safety Public Administration - agencies 25

Ind1990 codes 60 100-130 132-152 160-172 180-212 220-222 230-262 270-350 351-370 371-381 390-392 400-410 411-412 420-432 440-442 450-472 500-571 580-590 591-600 601-611 612-622 623-630 631-640 641 642-650 651-662 663-691 700-711 712-721 722-741 742-760 761-770 771-791 800-810 812-840 841 842-861 862-871 872-881 882 890-891 892-893 900-901 910 921-932

References Aizer, A. and Currie, J. (2004). Networks or neighborhoods? correlations in the use of publiclyfunded maternity care in california, Journal of Public Economics 88: 2573–2585. Ajilore, O. (2014). Identifying peer effects using spatial analysis: the role of peers on risky sexual behavior, Review of Economics of the Household . Alijore, O., Amialchuk, A., Xiong, W. and Ye, X. (1974). Uncovering peer effects mechanisms with weight outcomes using spatial econometrics, The Social Science Journal 82(6): 1063–1093. Bayer, P., Ross, S. L. and Topa, G. (2008). Place of work and place of residence: Informal hiring networks and labor market outcomes, Journal of Political Economy 116(6): 1150–1196. Becker, G. (1974). A theory of social interactions, Journal of Political Economy 82(6): 1063–1093. Borjas, G. (2014). Immigration Economics, Harvard University Press, Cambridge. Boucher, V., Bramoull´e, Y., Djebbari, H. and Fortin, B. (2014). Do peers affect student achievement? evidence from canada using group size variation, Journal of Applied Econometrics 29: 91– 109. Bramoull´e, Y., Djebbari, H. and Fortin, B. (2009). Identification of peer effects through social networks, Journal of Econometrics 150(1): 41 – 55. Calvo-Armengol, A., Patacchini, E. and Zenou, Y. (2009). Peer effects and social networks in education, Review of Economic Studies 76: 1239–1267. Edin, P.-A., Fredriksson, P. and Aslund, O. (2003). Ethnic enclaves and the economic success of immigrants – evidence from a natural experiment, Quarterly Journal of Economics 118(1): 329–357. Fryer, R. G. and Torelli, P. (2010). An empirical analysis of ‘acting white’, Journal of Public Economics 94(5-6): 380–396. Hellerstein, J. K., Kutzbach, M. J. and Neumark, D. (2014). Do labor market networks have an important spatial dimension?, Journal of Urban Economics 79: 39–58. Hellerstein, J. K., Neumark, D. and McInerney, M. (2008). Spatial mismatch or racial mismatch?, Journal of Urban Economics 64(2): 464–479. Ioannides, Y. M. (2013). From Neighborhoods to Nations: The Economics of Social Interactions, Princeton University Press, Princeton, New Jersey. Lee, L.-f. (2007). Identification and estimation of econometric models with group interactions, contextual factors and fixed effects, Journal of Econometrics 140: 333–374. Lee, L.-f., Liu, X. and Lin, X. (2010). Specification and estimation of social interaction models with network structures, The Econometrics Journal 13: 145–176. Lin, X. (2010). Identifying peer effects in student academic achievement by spatial autoregressive models with group unobservables, Journal of Labor Economics 28(4): 825–860. Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem, Review of Economic Studies 60: 531–542. 26

Manski, C. F. (2000). Economic analysis of social interactions, Journal of Economic Perspectives 14(3): 115–136. Ong, P. M. and Miller, D. (2005). Spatial and transportation mismatch in los angeles, Journal of Planning Education and Research 25(1): 43–56. Patacchini, E. and Venanzoni, G. (2014). Peer effects in the demand for housing quality, Journal of Urban Economics 83(0): 6–17. Rossi-Hansberg, E., Sarte, P.-D. and Owens III, R. (2010). Housing externalities, Journal of Political Economy 118(3): 485–535. Ruggles, S., Alexander, J. T., Genadek, K., Goeken, R., Schroeder, M. B. and Sobek, M. (2010). Integrated public use microdata series: Version 5.0 [machine-readable database]. Sacerdote, B. (2001). Peer effects with random assignment: Results for dartmouth roommates, Quarterly Journal of Economics 116(2): 681–704. Singer, A. (2004). The Rise of New Immigrant Gateways, Center on Urban and Metropolitan Policy, The Brookings Institution, Washington, D.C. Weinberg, B. A. (2007). Social interactions with endogenous associations, NBER Working Paper 13038 . Weinberg, B. A., Reagan, P. B. and Yankow, J. J. (2004). Do neighborhoods affect hours worked? evidence from longitudinal data, Journal of Labor Economics 22(4): 891–924. Zenou, Y. (2009). Urban Labor Economics, Cambridge University Press, New York. Zhu, P., Liu, C. Y. and Painter, G. (2014). Does residence in an ethnic community help immigrants in a recession?, Regional Science and Urban Economics 47: 112 – 127. Zimmerman, D. J. (2003). Peer effects in academic outcomes: Evidence from a natural experiment, Review of Economics and Statistics 85(1): 9–23.

27