The number of immigrants in the U.S. has increased substantially since the 1960s. Researchers have observed that immigrants from the same ethnic groups tend to cluster together in the same geographic areas (Cutler et al., 2008). It is however a puzzle as to whether this residential segregation could positively or negatively affect immigrants’ labor market outcomes. While some studies find that ethnic segregation is beneficial for immigrants (Cutler et al., 2008; Damm, 2009; Edin et al., 2003; Munshi, 2003), others show that ethnic segregation either leads to deteriorated labor market outcomes (Beaman, 2012; Borjas, 2000) or has no impact at all (Grönqvist, 2006, Song and Xu, 2020). What is more important is that ethnic segregation could impact immigrant members differently depending on their group characteristics. Edin et al. (2003) find that a higher average income of the ethnic group could positively affect enclave members’ wages, and Cutler et al. (2008) find a similar wage effect, but only for those from the highly educated groups.
In this paper, we explore whether immigrants’ own education levels and the ethnic groups’ education levels might individually and interactively explain the puzzles on the effects of ethnic segregation. We first conjecture that the impacts of ethnic segregation might vary according to immigrants’ individual education levels. For example, language barriers (Carliner, 2000), labor market discrimination (Edin et al., 2003), and overreliance on informal social networks in job searches (Borjas, 1998; Ioannides and Loury, 2004; Wahba and Zenou, 2005) could impede low education immigrants from seeking job opportunities outside of ethnic enclaves (Glitz, 2014). As a result, these immigrants could be more susceptible to segregation effects, relative to high education immigrants. On the other hand, high education immigrants, who for various reasons become ethnically segregated, could also be more negatively affected by segregation if their better matched jobs are outside of the enclaves.
Second, the differential segregation effects mentioned above may depend on the average education level of the immigrants’ ethnic group. The direction of the effect, however, is also ambiguous. On the one hand, if immigrants’ individual education levels are similar to their ethnic enclaves’ average education levels, they may benefit from better job networks, which would either alleviate the negative impacts of segregation or reinforce the positive effects of segregation. For example, it is possible that highly educated immigrants’ high group average education level may be correlated with their occupational specific skills such as computer programming or data analytical expertise. If both education and these talents are highly demanded in areas where certain industries concentrate, then ethnic segregation might create positive network effects or synergies that could help the highly educated immigrants (Cutler et al., 2008). On the other hand, matches between immigrants’ individual education levels and group average education levels may negatively impact immigrants due to increased competition for jobs (Beaman, 2012; Cortes, 2008).
Since theories cannot reconcile which of the aforementioned effects dominate, this study uses empirical strategies to examine the puzzle of ethnic segregation. We argue that it remains difficult to empirically estimate impacts of segregation because immigrants are likely to endogenously sort geographically based on observed and unobserved factors. Research using neighborhood-level segregation intensity measures (such as percentage of immigrants from the same ethnic group in the same tract/block) often find that segregation harms immigrants’ labor market outcomes. This finding, however, might be driven by lower educated immigrants sorting into ethnically concentrated neighborhoods due to housing market discrimination (Ihlanfeldt and Sjoquist, 1998) or a lack of fluency in the host country's language (Carliner, 2000), among other factors. Therefore, a number of papers adopt an alternative method by exploiting policies that exogenously place refugees in host countries (Beaman, 2012; Damm, 2009; Edin et al., 2003). They nevertheless still find conflicting effects of ethnic enclaves on immigrants’ labor market outcomes. While these results are less prone to endogeneity issues, they may not be generalizable enough to all immigrants because of potential unobserved factors specific to the refugee groups, the placement policies, and the host countries.
In this paper, instead of exploiting such exogenous variations in segregation intensity, we utilize two strategies to address the issues posed by immigrants’ sorting behaviors. First, we use a Metropolitan Statistical Area (MSA) level isolation index, as opposed to neighborhood level indices, to measure segregation intensity. Within the same MSA, lower educated immigrants are more likely to stay in ethnic enclave neighborhoods, while highly educated immigrants may move to more diversified neighborhoods with better economic conditions.
For example, Danzer and Yaman (2016) find that immigrants who have better host country language skills are less likely to ethnically segregate. Additionally, Furtado and Song (2015) find that immigrants with better observed and unobserved human capital tend to assimilate into host countries faster.
Our second strategy is to adopt a triple differences model to minimize the selection bias associated with MSA-level segregation intensity. Specifically, we control for MSA-specific time trends, group-specific time trends, and ethnicity-MSA-specific factors. Thus, our model takes into account any MSA or ethnic group factors that influence immigrants’ outcomes over time, such as the overall wage increase in a MSA and the expansion of industries that hire a large number of workers from a specific ethnic group. Our model also minimizes any time invariant MSA factors that impact outcomes of immigrants from a certain ethnic group, for example, local job training programs that target specific immigrant groups.
Following the convention in the literature (e.g., Cutler et al., 2008), we use a sample of immigrants who are between ages 20–30 and who migrated to the U.S. at age 17 or later. The reason we focus on young immigrants is to minimize the bias caused by older immigrants endogenously sorting into ethnic enclaves of different education levels and different economic conditions.
A balancing test using immigrants between ages 30–40 confirms our concern. We found that 30–40 years old immigrants with higher levels of education are less likely to segregate. More details about the balancing test will be discussed in Section 5.
To estimate the differential segregation effects, we introduce a set of double and triple interaction terms between the segregation intensity, immigrants’ own education, and ethnic group average education. We find that, given the education level of the co-ethnic peers, the average segregation effect is positive for lower educated immigrants and negative for highly educated immigrants. However, when estimating the differential effects of segregation by both individual education and group average education, we find that low education immigrants only receive the benefits from ethnic segregation when being isolated with other lower educated ethnic peers. The negative segregation effects on highly educated immigrants can be reversed if they are living with many other highly educated co-ethnics. These findings support the theory of positive network effects from ethnic segregation. Finally, we do not find any significant segregation effects on young immigrant's employment prospects.
In the rest of the manuscript, we use the phrase “immigrants” in replace for “young immigrants” for brevity purposes.
This paper proceeds as follows. In Section 2, we discuss in greater detail how immigrants’ individual education levels and ethnic group average education levels may influence segregation effects. In Section 3, we introduce the data used for our empirical estimations and the measure of segregation intensity. In Section 4, we discuss the triple interaction and triple differences model. In Section 5, we test the validity of our identification strategy. In Section 6, we present the results of our main model. In Section 7, we test the robustness of the earnings results. Section 8 presents the employment results. We conclude in the last section.
In the current literature, there is no definite conclusion as to how ethnic segregation influences immigrants’ labor market outcomes. Some papers showed evidence that immigrants gain labor market advantages from ethnic segregation. For example, using the natural experiment of the Swedish immigration policy, Edin et al. (2003) showed that, when sorting is taken into account, living in enclaves improves labor market outcomes for lower educated immigrants. Munshi (2003) used an instrumental variable approach and found that Mexican immigrants with larger network sizes in the host country have better labor market outcomes. Damm (2009) showed that immigrants’ hourly wages increase with enclave size when refugees were randomly dispersed throughout Denmark.
On the other hand, other research found negative or no segregation effects on immigrants’ outcomes. For instance, Borjas (2000) showed that ethnic segregation slows down wage growth for immigrants in the U.S. Beaman (2012) found that an increase in network size can negatively impact new arrivals yet benefit tenured members. Grönqvist (2006) used Swedish data and found that the size of enclave has no effect on earnings.
Moreover, segregation effects might change with the average human capital of ethnic groups. For example, Edin et al. (2003) found that immigrants from high income groups benefit more from ethnic segregation. Cutler et al. (2008) showed that ethnic group average education levels negatively impact segregation effects. However, this relationship was reversed when they instrument the endogenous segregation intensity.
In this paper, we propose that ethnic segregation has different impacts on immigrants with different education levels and from ethnic groups with different average education levels. If our hypothesis is true, it may help explain why findings regarding segregation effects are inconsistent in the previous literature. We will analyze different cases in the following three sub-sections.
Given the same ethnic group average education level, the effects of ethnic segregation might be different for immigrants with different education levels. The direction of these effects, however, is uncertain. Living in ethnic enclaves may lead to worse outcomes for lower educated immigrants than for highly educated immigrants. Lower educated immigrants who are living in ethnic enclaves have fewer chances to speak the host country language, which slows down their assimilation process. In addition, jobs in ethnic enclaves might be of a lower quality (Cutler et al., 2008). Since highly educated immigrants may have broader networks extending beyond their local neighborhoods, if ethnic segregation has negative effects on immigrants’ labor market outcomes in general, education may in some part protect highly educated immigrants from these effects.
On the contrary, it is also likely that segregation may be more harmful to highly educated immigrants because their opportunity costs of staying in ethnic enclaves might be larger than those of lower educated immigrants’. More interactions with natives and job searching in native-dominated labor markets may bring highly educated immigrants more job opportunities and higher returns to education. Lower educated immigrants might instead benefit from ethnic segregation because ethnic enclaves could provide job networks and language environments where lower educated immigrants can obtain job market information more easily (Bayer et al., 2008).
Similarly, ethnic segregation may have different impacts on immigrants from ethnic groups with different average education levels. The direction of the segregation effects might still be unclear.
Cutler et al. (2008) concluded that group average education positively influences segregation effects. However, this result is based on an instrumental variable which we argue might violate the exclusion restriction. Our analysis suggesting this is in Appendix A.
Alternatively, high group average education levels may also strengthen the positive segregation effects for immigrants. With more highly educated immigrants in the same ethnic enclave, immigrants of all education levels may benefit from better quality in professional networks. Also, highly educated immigrants may create more job opportunities near ethnic enclaves by building targeted businesses, such as ethnicity-specific companies (Borjas, 2000).
Finally, the differential segregation effects by individual education levels and ethnic group average education levels are again indeterminant for the following reasons. Ethnic segregation may be beneficial to immigrants who are living in ethnic enclaves with many similarly educated co-ethnics, because the positive network effects of ethnic segregation could amplify and improve job matching quality. For example, it might be easier for a lower educated Chinese immigrant to find a table waiting job in Chinatown, and a highly educated Chinese computer engineer to find better job information through the Chinese community near Silicon Valley.
Conversely, ethnic segregation may lead to adverse labor market outcomes as competition between immigrants with similar education levels drives wage rates down. This negative effect may be especially harmful for lower educated immigrants who are constricted to the labor markets near ethnic enclaves.
For example, Patel and Vella (2013) find that newly arrived immigrants are more likely to choose the same occupations as their co-ethnics. Immigrants from non-English-speaking countries tend to have jobs which require less communication skills (Chiswick and Taengnoi, 2007).
In conclusion, segregation effects may vary with immigrants’ individual education levels, group average education levels, and the two factors combined. However, the direction of these effects remains unclear. This paper therefore empirically tests these differential segregation effects.
Our study uses the Integrated Public Use Micro Series (Ruggles et al., 2015) from the 2000 U.S. Census and the American Community Survey (ACS) five-year sample for 2006–2010. The census data are 5-percent samples of the U.S. population. The 2006–2010 ACS sample is a five-year estimate of all geographical areas using the 1-percent samples from 2006, 2007, 2008, 2009, and 2010.
The advantage of using this single combined 5-year estimate dataset is that it increases statistical reliability of the data for less populace areas and small subgroups. Considering this paper includes all identifiable ethnic groups, where some of them have small population sizes, we believe this is the best sample to use for more accurate identification (U.S. Census, 2022).
The outcome variables are inflation adjusted log of annual earnings and a dummy variable indicating whether an immigrant is currently employed. Immigrants who are currently students, who are not living in metropolitan areas, or who are not in the labor force were excluded from our sample.
We also restrict our sample to immigrants between the ages of 20 and 30 and migrated to the U.S. when they were 17 or older. There are several reasons for this sample restriction. First, for immigrant youths, it is more likely that their parents selected their residential locations for them (Cutler et al., 2008), unlike older immigrants who have more time to decide where to live in the host country based on their own preferences.
We run balancing tests by using samples from different age groups. We find that samples of older immigrants are more likely to fail the balancing tests. Also, when we do not restrict the age of migration to the U.S., results of the balancing tests also show systematic selections into ethnic enclaves.
Descriptive Statistics
Log earnings | 104,577 | 9.656 | 0.738 | - | - | - |
Employment | - | - | - | 129,966 | 0.919 | 0.273 |
Isolation index | 8,149 | 0.014 | 0.021 | 8,874 | 0.014 | 0.020 |
Group average education | 8,149 | 13.136 | 2.241 | 8,874 | 13.153 | 2.241 |
Immigrant share | 8,149 | 0.004 | 0.015 | 8,874 | 0.004 | 0.014 |
Education | 104,577 | 11.023 | 4.170 | 129,966 | 10.922 | 4.218 |
Age | 104,577 | 26.395 | 2.829 | 129,966 | 26.330 | 2.874 |
Male | 104,577 | 0.680 | 0.467 | 129,966 | 0.655 | 0.475 |
Married | 104,577 | 0.383 | 0.486 | 129,966 | 0.385 | 0.486 |
Have children | 104,577 | 0.301 | 0.459 | 129,966 | 0.305 | 0.461 |
White | 104,577 | 0.121 | 0.326 | 129,966 | 0.121 | 0.326 |
Black | 104,577 | 0.025 | 0.157 | 129,966 | 0.026 | 0.159 |
Asian | 104,577 | 0.182 | 0.386 | 129,966 | 0.176 | 0.381 |
Hispanic | 104,577 | 0.672 | 0.469 | 129,966 | 0.677 | 0.468 |
In this paper, we use the isolation index as a measure of ethnic segregation intensity. It thereby shows the extent of an immigrant's exposure to other immigrants from the same ethnic group in the same area. We use Equation (1) to calculate the isolation index,
This method considers the effects of the total population and the ethnic group population on isolation intensity. For example, if we use the traditional equation to calculate isolation intensity, even if the population of an ethnic group is very small, we may still get a very large isolation index.
The traditional equation for the isolation index is
Another method that is often used to measure ethnic segregation is the dissimilarity index. The equation is
We utilize the tract-level information in the 2006–2010 ACS five-year sample (U.S. Census Bureau, 2016) and the 2000 Summary Tape File 3 tables to construct the isolation index. Ethnic group is defined by an immigrant's country of origin. We have 60 different ethnic groups for the 2000 sample and 74 groups for the 2010 sample. Since we are interested in using multiple cross-sections of data across time, we will, after merging, use the 59 ethnic groups that are consistently defined across all sample years.
Following previous literature (Cutler et al., 2008; Cutler and Glaeser, 1997), we use the MSA-level isolation index instead of neighborhood level measures, such as percentage of immigrants from an ethnic group in a tract/block. We focus on comparing the outcomes of immigrants living in highly segregated MSAs with those of immigrants living in less segregated MSAs. This is because, within the same MSA, immigrants with higher socioeconomic status are more likely to move to better neighborhoods. Hence, the selection bias associated with neighborhood-level segregation intensity might be more substantial than the selection bias associated with MSA-level segregation intensity. The comparison between neighborhoods within MSAs will lead to biased estimations of the segregation effects.
In this section, we begin by discussing our use of a triple differences model to deal with the selection bias caused by immigrants’ sorting behaviors, then we discuss how we estimate the differential isolation effects.
Equation (2) shows the triple differences model. Here,
Since we utilize multiple cross-sections of data across years, we can adopt the triple differences technique in order to control for ethnic group-specific time effects ( More examples of the triple differences model can be found in Yelowitz (1995) and Ravallion et al. (2005). Angrist and Pischke (2008) also briefly discuss the triple differences model.
We are also interested in measuring how segregation effects vary with immigrants’ individual education levels and ethnic group average education levels. Cutler et al. (2008) investigated the differential segregation effects on incomes of immigrants from ethnic groups with different group average education levels. The estimation equation can be written as the following:
Cutler et al. (2008) used MSA and group fixed effects model to identify segregation effects. In Equation (3), we modify their estimation equation by using the triple differences model. Cutler et al. (2008) found that
In this paper, we complement the existing literature by introducing additional two-way interaction terms among the isolation index, individual education, and ethnic group average education, as well as the triple interaction term of the three variables. The following represents the main regression:
Before addressing our results, we need to test whether our primary identification strategy—the triple differences model—eliminates, or at least attenuates, the potential biases that are caused by immigrants’ sorting behaviors. Immigrants are likely to choose their residential areas based on observed and unobserved characteristics, such as language, social connections, dietary habits, etc. It is thus likely that immigrants in segregated areas have different attributes from immigrants in less segregated areas. These attributes may influence their labor market outcomes and may exaggerate or undermine the effects of segregation. For example, immigrants who cannot speak native languages may prefer to reside within their ethnic enclaves (Bleakley and Chin, 2010). This lack of communication skills may cause them to have fewer high-paying employment opportunities regardless of where they live. Thus, we are more likely to find that ethnic segregation negatively correlates with immigrants’ wages.
To address these issues, we conduct several balancing tests to determine if immigrants’ characteristics can predict their own ethnic segregation intensity in the triple differences model. If the isolation index is unrelated to immigrants’ observed characteristics, then our estimated results should not be influenced by immigrants’ sorting behaviors based on observed characteristics. Also, as Altonji et al. (2005) and Bifulco et al. (2011) have pointed out, the extent of individuals’ sorting behaviors based on observables can be used as a guideline for the extent of their sorting behaviors based on unobserved variables. Thus, a lack of correlation between immigrants’ observed characteristics and the isolation index suggests that the triple differences model minimizes the likelihood that the estimated results are driven by immigrants’ selections into ethnic enclaves based on unobserved characteristics.
We also examine whether the interaction terms between immigrants’ individual education levels and other characteristics are correlated with the isolation index. If the interaction terms between individual education levels and other individual characteristics fail the balancing tests, then the estimated differential segregation effects associated with immigrants’ individual education levels might be biased. For example, if we observe that the coefficient of the interaction between age and education is negative in the balancing test, then the implication is that immigrants who are older and have higher education levels are less likely to sort into their ethnic enclaves. If they are more likely to be harmed by ethnic segregation, then the estimated results of how immigrant education influences segregation effects might be underestimated.
In Table 2, we present balancing test results using the triple differences model.
In order to show the magnitudes of the coefficients, we multiply the isolation index by 1000 in all balancing tests. We also use the earnings regressions sample for all balancing tests. When we use the sample for employment regressions, individual characteristics and their interactions with individual education are still insignificant as a group. A commonly used method to test the validity of double differences models is to examine the parallel trends, i.e., without any treatment, the treatment group and the control group should have similar changes in outcomes over time/across different locations. Olden and Møen (2022) show that the triple differences model is essentially the difference between two sets of double differences, and it does not require both double differences to have parallel trends at the same time. Instead, the two double differences just need to have biases in similar direction and magnitude. By following their method, we present evidence of parallel trend for our model in Appendix B.
Balancing Tests, Triple Differences Model
Education | −0.007 (0.007) | 0.042 (0.060) |
Age | −0.006 (0.006) | 0.014 (0.021) |
Male | 0.053 (0.042) | 0.144 (0.127) |
Black | −0.285 (0.313) | −0.000 (0.870) |
Other Nonwhite | 0.088 (0.215) | −0.026 (0.395) |
Age × Education | - | −0.002 (0.002) |
Male × Education | - | −0.008 (0.011) |
Black × Education | - | −0.022 (0.060) |
Other Nonwhite × Education | - | 0.009 (0.027) |
|
1.05 | 0.45 |
|
0.385 | 0.771 |
|
102,475 |
Significance is defined as follows: ***p<0.01; **p<0.05; *p<0.10.
In this section, we begin by discussing the estimated relationship between the isolation index and log earnings and how ethnic group average education affects the segregation effects. In Table 3, the first three columns show the empirical results of Model (2) with an increasing number of control variables in individual characteristics and education levels. The estimated coefficients of the isolation index in the first three columns are consistently statistically insignificant and negative. This suggests that there is no significant difference between earnings of immigrants living in various degrees of ethnic segregation.
Triple differences,
−0.217 (0.469) | −0.233 (0.436) | −0.156 (0.417) | 0.438 (1.721) | −0.498 (1.599) | −0.582 (1.556) | |
|
- | - | - | −0.039 (0.187) | 0.059 (0.175) | 0.061 (0.170) |
|
- | - | - | 0.055*** (0.013) | 0.049*** (0.012) | 0.023* (0.012) |
- | - | 0.028*** (0.002) | - | - | 0.028*** (0.002) | |
+characteristic control | No | Yes | Yes | No | Yes | Yes |
|
- | - | - | 0.04 | 0.06 | 0.07 |
|
- | - | - | 0.960 | 0.943 | 0.933 |
|
102,475 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
This insignificant segregation effect, however, could be because of a lack of control variables for group education levels. For example, if the segregation effects are opposite for different group education levels for reasons suggested in Section 2, then the average segregation effect might be statistically insignificant. We therefore present the estimated results of Model (3), which is Model (2) with group average education level and its interaction term with the isolation index, in the last three columns of Table 3. The consistently statistical insignificant coefficients of the isolation index and its interaction term suggest that the effect of isolation on log earnings is not influenced by group average educations.
Our findings in Table 3 are inconsistent with the OLS model findings in Table 7 of Cutler et al. (2008). There are two potential reasons behind the difference. First, we use a triple differences identification strategy. As shown in Appendix Table A.2, estimations without the triple differences controls do not entirely take immigrants’ endogenous sorting into account, which could bias results. Second, we also conjecture that the reason could be stemming from differences in samples. Our sample contains a larger number of ethnicity-city-year cells than Cutler et al. (2008).We also use yearly data (the 2000 and 2010 samples) to replicate findings in Cutler et al. (2008).The results are shown in Appendix Table A.1.
Moreover, the group average education variable becomes less significant after individual education level is controlled for. Therefore, we further investigate whether individual education level could alter the effects of segregation on immigrants’ labor market outcomes, both individually and collectively with group average education. We show the results of regressions with all two-way and three-way interaction terms between the isolation index, individual education, and group average education level in Table 4. To properly interpret the coefficients of these interaction terms, we center all the involving variables at their mean level. Specifically, we replace
Triple Differences and Triple Interactions Model (Centered), Earnings Results
−0.136 (0.414) | 0.375 (0.846) | 0.415 (0.847) | 0.327 (0.837) | |
−0.174*** (0.021) | −0.029 (0.023) | −0.160* (0.083) | −0.216** (0.096) | |
|
- | 0.109 (0.169) | 0.117 (0.169) | 0.098 (0.168) |
|
- | - | −0.036* (0.184) | - |
|
- | - | - | 0.505*** (0.138) |
|
- | - | - | −0.048** (0.021) |
|
- | 0.010*** (0.001) | 0.011*** (0.001) | 0.011*** (0.001) |
|
- | 0.007 (0.012) | 0.006 (0.012) | 0.007 (0.012) |
0.040*** (0.002) | 0.063*** (0.002) | 0.066*** (0.002) | 0.065*** (0.002) | |
|
33.97 | 0.69 | 1.10 | 3.09 |
|
0.000 | 0.559 | 0.354 | 0.009 |
|
102,475 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
We begin with the regression containing only the isolation index and its interaction with immigrants’ individual education levels. The main effect of segregation is negative but statistically insignificant, but the coefficient of the interaction term is negative and statistically significant. This suggests that highly educated immigrants who are living in ethnic enclaves have lower earnings than they would in diverse or native communities. Specifically, one percent increase in segregation reduces highly educated immigrants’ earnings by 0.174%. This is consistent with the hypothesis in Section 2.2., in that higher opportunity costs penalizes highly educated immigrants who live in ethnic enclaves, while lower educated immigrant can benefit from local labor market networks and less reliance on the host country linguistic skills. Additionally, higher individual education level also raises earnings, as per conventional wisdom.
In Column 2 of Table 4, we show the estimated results for the regression that only includes double interaction terms between isolation index, group average education and individual education. The estimated coefficient for the interaction between isolation index and individual education is consistently negative but has become statistically insignificant. The coefficients of the isolation index and its interaction with group average education are both statistically insignificant. Lastly, the positive estimated coefficient of the interaction term between group average education level and individual education level suggests that generally, immigrants tend to earn higher wages if they are similarly educated as their ethnic groups, irrespective of where they reside. One possible explanation is that ethnic groups with a certain level of group average education may concentrate in and have labor market advantages in certain industries. When immigrants’ own education level matches their ethnic group average education, they could benefit either from the ethnic neighborhoods’ group network effect or from the overarching shared expertise in a specific industry even if they do not live with their co-ethnics.
As mentioned in Section 2.4, there could be a more intricate effect of isolation that depends on both individual education and ethnic group average education, so in Column 3 of Table 4, we include the triple interaction term in the regression. The estimated coefficient of this triple interaction term is negative and statistically significant at 10% level, suggesting that match or mismatch between an immigrant's own education level and their group average education level play an important role in the effect of isolation. Even though the model specification in Column (3) points to the importance of the triple interaction term, the interpretation of its coefficient is still unclear. As discussed in Section 2.4, the negative coefficient of the triple interaction term might have different implications for immigrants with different levels of individual education.
As a result, instead of using one triple interaction term, in Column 4 of Table 4, we parse out the effects of segregation by considering highly educated and lower educated ethnic groups separately.
Specifically, we generate two new variables:
When considering all statistically significant terms involving the isolation index, the model shows that for lower educated immigrants who live in ethnic enclaves with low average education levels, a one percent increase in the isolation index increases earnings by 0.168 percent.
This is the net effect of isolation when the general effect of 0.216 is reduced by the triple interaction effect of −0.048. The net effects in the rest of this section are all calculated accordingly. We have also conducted the experiment using categorical education variables instead of years of schooling. The model specification is more complex due to the nature of necessary interaction terms in the model design. The results are presented in the tables in Appendix C. Using the categorical education variables makes the interpretation of the estimated coefficient of Isolation Index itself more complex, but the directions of the effects are similar to those found in the preferred model, i.e. for lower educated immigrants who isolate, the benefits are larger if they live amongst lower educated co-ethnics; and for highly educated immigrants, being isolated with co-ethnics of high average education could reduce the negative effects of isolation.
These results are notable in two important ways. First, while the average effect of isolation is positive for lower educated immigrants and negative for highly educated immigrants, this effect can be significantly influenced by whether the immigrants’ individual education matches the group average education. For lower educated immigrants, the reduction in segregation effect is significantly larger if they are isolated with highly educated co-ethnics than if they are isolated with lower educated co-ethnics. This means while the competition with similarly educated co-ethnics could hurt lower educated immigrants’ wages, removing them from their similarly educated co-ethnics, and therefore removing their access to education-specific networks, would reduce their wages so much that it reverses the general positive effect of isolation. On the other hand, while isolation causes an average negative effect for highly educated immigrants, being isolated with other highly educated co-ethnics could provide sufficiently high-quality networks to counteract that effect, reversing the isolation effect to positive. Both scenarios point to the same crucial role education-specific networks play.
Second, there is a seemingly contradiction between the average isolation effect found in Column 1 and the net effects of isolation found in Column 4. This contradiction can be reconciled by further investigation into the composition of immigrant groups. Because a significant portion of immigrant population are lower educated, lower educated immigrants are 3.5 times more likely to be isolated with other lower educated co-ethnics than with highly educated co-ethnics. Similarly, highly educated immigrants are 2.87 times more likely to live in ethnic enclaves with low average education than in ethnic enclaves with high average education. Therefore, the positive segregation effects for lower educated immigrants and the negative segregation effects for highly educated immigrants dominates.
In summary, our findings suggest a rather clear story: the net effects of ethnic segregation on earnings depends on the match or mismatch between immigrants’ own education levels and their group average education levels. While on average, ethnic segregation hurts highly educated immigrants and benefits lower educated immigrants, the opposite is true in the case of being segregated with co-ethnics of high average education levels.
As shown in Table 2, the estimated coefficients of male dummy and the interaction term between male dummy and individual education are both statistically insignificant in the balancing test regression. This suggests that any potential residential sorting according to gender should not affect our main empirical results. However, since 68% of the sample are male immigrants, we nevertheless test whether our model is robust when males and females are estimated separately. The results in the first two columns of Table 5 show a rather similar picture for male and female immigrants, except that the coefficients for female immigrants are less precisely estimated than for male immigrants.
Robustness Tests, Triple Differences and Triple Interactions Model (Centered)
−0.945 (1.146) | 1.962 (1.618) | −0.572 (1.001) | 0.294 (1.262) | |
−0.281*** (0.095) | −0.108 (0.082) | −0.179 (0.115) | −0.158 (0.129) | |
|
−0.124 (0.221) | 0.449 (0.329) | −0.187 (0.212) | 0.115 (0.281) |
|
0.554*** (0.156) | 0.439* (0.233) | 0.513*** (0.183) | 0.368* (0.222) |
|
−0.060*** (0.021) | −0.035* (0.020) | −0.038 (0.025) | −0.033 (0.029) |
|
0.010*** (0.001) | 0.012*** (0.001) | 0.010*** (0.001) | 0.009*** (0.001) |
|
−0.011 (0.015) | 0.014 (0.025) | −0.001 (0.019) | −0.010 (0.025) |
0.060*** (0.003) | 0.071*** (0.003) | 0.065*** (0.003) | 0.063*** (0.003) | |
|
3.68 | 1.88 | 1.73 | 0.73 |
|
0.003 | 0.095 | 0.123 | 0.600 |
|
69,166 | 31,718 | 51,828 | 32,188 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10
As stated in Section 3, we have restricted our sample to immigrants who migrated to the U.S. after age 17 to avoid endogenous choices in education post arrival in the host country. However, their education decisions may still be partially influenced by their experience in the U.S. Therefore, we further restrict our sample to immigrants who migrated after age 21 and age 23 to test the robustness of our findings, because these immigrants’ education levels are more likely to have been determined before arrival in the U.S.
In Columns 3 and 4 of Table 5, we re-estimate the model using only immigrants who migrated after age of 21 and immigrants who migrated after age of 23, respectively. The estimated coefficients are largely robust in terms of signs, but with lower level of statistical significance for some variables. This is not surprising because the sample size has decreased by almost 50% and 70% after the sample restrictions. Given the number of variables in the triple differences regressions, we expect the estimation to be less precise.
The literature of immigration has suggested that immigrants’ English language skills (Bleakley and Chin, 2004) and their unobserved degrees of social assimilation (Furtado and Song, 2015) could play an important role in labor market outcomes. Since it is likely that immigrants’ education is correlated with these factors, we further test whether the estimated segregation effect according to education is robust to additional controls of linguistic and social skills.
We start with conducting a second series of balancing tests with two additional variables: an indicator of whether the immigrant has insufficient English-speaking skill and the number of years the immigrant has lived in the U.S.
We categorize immigrants who self-report as speaking English not well or not at all as having insufficient English-speaking skill.
Balancing Tests with Assimilation Measures, Triple Differences Model
Education | −0.011 (0.007) | 0.017 (0.061) |
Age | −0.002 (0.008) | 0.008 (0.023) |
Male | 0.051 (0.042) | 0.142 (0.129) |
Black | −0.278 (0.313) | −0.051 (0.867) |
Other Nonwhite | 0.097 (0.215) | 0.023 (0.409) |
English Insufficiency | −0.132*** (0.050) | −0.255* (0.154) |
Years in the U.S. | −0.012 (0.008) | 0.007 (0.019) |
Age × Education | - | −0.001 (0.002) |
Male × Education | - | −0.008 (0.011) |
Black × Education | - | −0.017 (0.060) |
Other Nonwhite × Education | - | 0.006 (0.028) |
English Insufficiency × Education | - | 0.012 (0.014) |
Years in the U.S. × Education | - | −0.002 (0.002) |
|
1.73 | 0.064 |
|
0.097 | 0.695 |
|
102,475 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
Therefore, in Table 7, we test the robustness of our model by including the indicator of insufficient English-speaking skill, number of years in the U.S., and their interactions with the isolation index.
Even though in our data we already exclude immigrants who have not completed their secondary education, we argue there is value in identifying the heterogeneous effects for immigrants with different degrees of English language skills. First, different home countries offer varying degrees of English language education and therefore there is still sufficient variation in English language skills in the data. Second, given the variation in the English language skills for the immigrants in the sample, we could also use English language skill as a proxy for social skills. Third, the estimated coefficients we gather in Column 2 in Table 6 can be considered a lower-bound effect of English language's effect as immigrants who have not completed secondary education upon arriving in the U.S. would tend to have worse English language skills and labor market outlooks. Additionally, in Appendix A Table A.4, we also show the results by using the sample excluding immigrants who identified themselves as English speakers. There aren’t significant changes in the results either.
Mechanisms and Heterogeneity Tests, Earning Results
0.327 (0.837) | 0.320 (0.843) | 0.673 (0.809) | |
−0.216** (0.096) | −0.247*** (0.090) | −0.244*** (0.094) | |
|
0.098 (0.168) | 0.132 (0.171) | 0.115 (0.164) |
|
0.505*** (0.138) | 0.516*** (0.139) | 0.544*** (0.141) |
|
−0.048** (0.021) | −0.056*** (0.020) | −0.057*** (0.021) |
−0.183*** (0.012) | - | ||
- | 0.046 (0.131) | - | |
- | 0.037*** (0.002) | ||
- | - | −0.041** (0.017) | |
|
3.09 | 3.69 | 3.85 |
|
0.009 | 0.003 | 0.002 |
|
102,475 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
The results in Column 3 of Table 7 suggest that additional years of living in the U.S. increase immigrants’ earnings, consistent with the literature that ceteris paribus, an increase in social assimilation induces bettering labor market outcomes (Furtado and Song 2015). Moreover, immigrants who have stayed in the U.S. for longer time would suffer wage loss if they live in ethnic enclaves, possibly due to the lack of access to a wider labor market as suggested in Section 2. Similar to the results in Column 2, the estimated coefficients of
To summarize, results in this sub-section show that the baseline results are robust to linguistic skills or social assimilation, which suggests that the differential segregation effects are not likely caused by immigrants’ socio-communication skills. However, we can only test a limited number of variables here. The baseline results could be caused by other unobserved factors that are correlated with education levels, ethnic enclave residence, and earnings. For example, immigrants from families that are heavily influenced by traditional culture and values may be more likely to live in ethnic enclaves. If these immigrants are from wealthy families, they are more likely to get higher levels of education but may give up job opportunities that are far away from their ethnic enclaves. On the contrary, if they are from poor families, they would have lower educational attainment and are less capable in getting higher quality jobs.
In this section, we test whether ethnic segregation has differential segregation effects on immigrants’ employment outcomes. Similar to the earnings results, in Table 8, we did not find differential segregation effects on employment for immigrants from ethnic groups with different average education levels. In Table 9, only the coefficient of the double interaction between isolation index and individual education in Column 1 is statistically significant and negative, suggesting that on average, ethnic segregation reduces the likelihood of being employed for highly educated immigrants but improve lower educated immigrants’ employment. However, this coefficient becomes statistically insignificant in the other model specifications. Moreover, the coefficients of the other double interaction and triple interaction terms show no discernable segregation effect on employment.
Triple differences,
−0.106 (0.151) | −0.109 (0.152) | −0.092 (0.153) | 0.014 (0.514) | −0.061 (0.512) | −0.068 (0.512) | |
|
- | - | - | −0.014 (0.050) | −0.006 (0.050) | −0.005 (0.050) |
|
- | - | - | −0.001 (0.005) | −0.001 (0.005) | −0.004 (0.005) |
- | - | 0.003*** (0.000) | - | - | 0.003*** (0.000) | |
+characteristic control | No | Yes | Yes | No | Yes | Yes |
+individual education control | No | No | Yes | No | No | Yes |
|
- | - | - | 0.37 | 0.31 | 0.31 |
|
- | - | - | 0.690 | 0.736 | 0.736 |
|
127,769 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
Triple Differences and Triple Interactions Model (Centered), Employment Results
−0.089 (0.153) | −0.009 (0.249) | −0.012 (0.249) | −0.018 (0.249) | |
−0.013** (0.005) | −0.005 (0.006) | 0.005 (0.017) | 0.001 (0.018) | |
|
- | 0.027 (0.056) | 0.026 (0.056) | 0.025 (0.056) |
|
- | - | 0.003 (0.004) | - |
|
- | - | - | 0.042 (0.037) |
|
- | - | - | 0.002 (0.004) |
|
- | 0.001*** (0.000) | 0.000** (0.000) | 0.000** (0.000) |
|
- | −0.005 (0.004) | −0.005 (0.004) | −0.005 (0.004) |
0.004*** (0.000) | 0.005*** (0.001) | 0.005*** (0.001) | 0.005*** (0.001) | |
|
2.97 | 0.49 | 0.57 | 0.68 |
|
0.052 | 0.687 | 0.686 | 0.635 |
|
127,769 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
Table 10 presents a similar picture for male and female immigrants as well as immigrants who migrated after age of 23. The model becomes slightly more significant for the sample of immigrants who arrived after 21. Given the results are robust when using other three samples, this result is likely caused by random sample variations as opposed to a different labor market mechanism for immigrants who migrated when they were 21 and 22. Lastly, Table 11 suggests that controlling for English language skills and length of stay in the U.S. does not affect the statistical significance of the isolation index and its interaction terms with education variables.
Robustness Test, Triple Differences and Triple Interactions Model (Centered), Employment Results
0.134 (0.304) | −0.346 (0.464) | −0.119 (0.280) | −0.133 (0.332) | |
−0.004 (0.016) | −0.001 (0.029) | 0.039** (0.018) | 0.030 (0.019) | |
|
0.038 (0.065) | 0.006 (0.113) | 0.013 (0.069) | 0.115 (0.077) |
|
0.034 (0.046) | 0.068 (0.057) | 0.036 (0.051) | 0.029 (0.065) |
|
0.001 (0.004) | −0.001 (0.007) | 0.009** (0.004) | 0.007 (0.005) |
|
0.001*** (0.000) | −0.001 (0.000) | 0.001** (0.000) | 0.000 (0.000) |
|
−0.005 (0.005) | −0.007 (0.009) | −0.009 (0.006) | −0.011 (0.007) |
0.005*** (0.001) | 0.004*** (0.001) | 0.005*** (0.001) | 0.005*** (0.001) | |
|
0.67 | 0.62 | 1.41 | 2.09 |
|
0.647 | 0.686 | 0.219 | 0.063 |
|
83,164 | 42,838 | 65,149 | 40,642 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.10.
Mechanisms and Heterogeneity Tests, Employment Results
−0.018 (0.249) | −0.028 (0.249) | 0.024 (0.251) | |
0.001 (0.018) | −0.000 (0.018) | −0.001 (0.018) | |
|
0.025 (0.056) | 0.031 (0.056) | 0.027 (0.056) |
|
0.042 (0.037) | 0.042 (0.037) | 0.043 (0.038) |
|
0.002 (0.004) | 0.001 (0.004) | 0.001 (0.004) |
- | 0.045 (0.043) | - | |
- | −0.021*** (0.004) | - | |
- | - | −0.006 (0.007) | |
- | - | 0.003*** (0.001) | |
|
0.68 | 0.65 | 0.57 |
|
0.635 | 0.664 | 0.721 |
|
127,769 |
Significance is defined as follows:
p<0.01;
p<0.05;
p<0.01.
Residential segregation may have positive or negative impacts on immigrants’ labor market outcomes. On the one hand, segregation might provide social networks where immigrants can exchange job information in their own languages. On the other hand, segregation might also slow immigrants’ assimilation process by isolating immigrants from natives. In the current literature, there is no consensus as to whether ethnic segregation is beneficial or harmful to immigrants. In this paper, we attempt to use young immigrants’ individual education levels to gain further insights on segregation effects and to measure the impacts of ethnic group average education on segregation effects.
Our estimation shows that on average, ethnic segregation reduces the earnings of highly educated young immigrants and boosts the earnings of lower educated young immigrants, but it does not have differing impacts on employment. Moreover, the education compositions of ethnic enclaves lead to different net segregation effects for immigrants with different education levels. Specifically, the net segregation effects are only positive when immigrants’ own education levels match the average education levels of their ethnic enclaves. This is consistent with the labor market network theory: removing the lower educated co-ethnic networks could reduce lower educated immigrants’ earnings, even when they live in ethnic enclaves. On the other hand, providing a high-quality network of similarly educated co-ethnics for highly educated immigrants would reverse the average negative effect of isolation and provide a net benefit in their earnings.
From a policy standpoint, this paper suggests that when policy makers promote diversity in ethnic neighborhoods, they should take into consideration the valuable benefits of ethnic segregation for lower educated immigrants, especially for those living with many other lower educated co-ethnics. While diversified neighborhoods have their own advantages, how to compensate for the potential reduction in earnings due to a loss of access to the education-specific co-ethnic labor market networks should be considered in the process of integration promotion.