Social Networks, Ethnicity, and Entrepreneurship

A. Model Set-Up

We construct a simple model to illustrate how social isolation and small group size can generate ethnic entrepreneurial clustering when social interactions and production are complementary. To keep the model tractable and intuitive, we make several strong assumptions. Everyone has equal ability and is divided into two ethnic groups. Group A is the minority, with a continuum of individuals of mass N_A, and group B has mass N_B > N_A. Both groups have equal access to industries, and there is no product market discrimination, but the groups are socially segregated and spend their leisure time separately. Social interactions are random within ethnic groups, such that each person interacts with a representative sample of individuals in their own group.

We analyze how these two ethnic groups sort across two industries. Industry 1 has a production structure where self-employed entrepreneurs obtain advantages through social interactions with other self-employed entrepreneurs in the same industry. When socializing during family gatherings and religious/cultural functions, entrepreneurs in this industry can mentor each other and share industry knowledge and professional advice. The more an entrepreneur socializes with other entrepreneurs, the more knowledge is exchanged. Industry 0, by contrast, exhibits constant returns to scale with worker productivity normalized to one. This industry can equally comprise individuals working in self-employment or in larger firms; the core assumption is that private social interactions do not have the same benefit in industry 0 as they do in industry 1.

More formally, define X_l for l ∈ {A, B} as the fraction of the population in group l who are self-employed entrepreneurs in industry 1. Because social interaction is random within groups, a fraction X_l of the friends and family members of every individual in group l are also self-employed entrepreneurs in industry 1. For industry 1, denote individual entrepreneurial productivity in group l as θ(X_l). Our assumption that productivity increases when socializing with other entrepreneurs in industry 1 is formally stated as:

Denote aggregate output of industry 1 as Q₁, which is a function of the distribution (X_A, X_B): (1)

Because social interaction plays no role for industry 0, its aggregate output is simply: (2)

Demands for the two industries need to be complementary enough to avoid the complications of multiple optima possibly generated by nonconvexities. We simply assume them to be perfect complements via a Leontief utility function for consumers: (3) where v > 0 is a preference parameter, and q₀ and q₁ are individual consumption of each industry’s output, respectively.

B. The Pareto Problem

We now describe the efficient outcome. Because the outputs of both industries have unitary income elasticities, distributional aspects can be ignored when characterizing the efficient outcome. The problem simplifies to choosing an industry distribution (X_A, X_B) that maximizes a representative utility function U[Q₀(X_A, X_B), Q₁(X_A, X_B)]. A marginal analysis is inappropriate because this is a nonconvex optimization problem. We consider instead the most specialized industry distributions, where as many individuals as possible from a single group A or B are self-employed entrepreneurs in industry 1.

Figure 1 depicts the production possibilities for the two specialized distributions. Define V(X_A, X_B) ≡ Q₁/Q₀ as the ratio of industry outputs under the distribution (X_A, X_B). Along the curve with the kink V(1, 0) in the figure, group A specializes as self-employed entrepreneurs in industry 1. Starting from a position on the far right where everyone works in industry 0, members of group A are added to the set of self-employed entrepreneurs in industry 1 as we move leftward along the x-axis. When the kink at V(1, 0) is reached, all members of group A are self-employed entrepreneurs in industry 1. Thereafter, continuing leftward, members of group B are also added to industry 1 until Q₀ = 0. Similarly, along the curve with the kink V(0, 1), group B first specializes as self-employed entrepreneurs in industry 1. Members of group B are added moving leftward along the x-axis until the kink at V(0, 1), where all Bs are working in industry 1. Thereafter members of group A are also added until Q₀ = 0.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Model Depiction of Entrepreneurial Specialization

Notes: Left panel shows production possibilities with specialized occupational distributions. The ray v is the preference parameter over goods in a Leontief utility function. Along the curve with kink V(1, 0), all entrepreneurs belong to group A (belowkink), or all members of group A are entrepreneurs (above). Along the curve with kink V(0, 1), all entrepreneurs belong to group B (below), or all members of group B are entrepreneurs (above). The right panel shows the efficient occupational distribution for different values of v assuming convex productivity in interactions. The minority group A specializes as entrepreneurs if that sector is small enough.

The curve with minority specialization is above the curve with majority specialization, so long as the need for self-employed entrepreneurs in industry 1 is sufficiently small. A large fraction of those in group A are self-employed entrepreneurs in industry 1 when the minority specializes, allowing minority entrepreneurs to socialize mostly with other entrepreneurs in industry 1, improving productivity. The same is not true for the majority, because even if a large fraction of self-employed entrepreneurs in industry 1 are in group B, most Bs are nevertheless employed in industry 0.

The argument can be generalized to show that minority specialization is Pareto efficient so long as industry 1 is small enough. Perfect complementarity simplifies the problem of solving for the optimal allocation, because any bundle where industrial outputs are in the exact ratio v of the Leontief preferences (3) is strictly preferable to all other bundles that do not include at least as much of each industry. The Pareto optimal distribution (X_A, X_B) must therefore satisfy v = V(X_A, X_B). Define the total number of entrepreneurs in the population as M ≡ X_AN_A + X_BN_B. It follows that:

This contradicts Q′₁ ≥ Q₁.

The efficient outcome requires that a single group specializes as self-employed entrepreneurs in industry 1, and importantly, which group specializes is not arbitrary. Minority specialization is more efficient because the minority’s social isolation enables entrepreneurs in A to socialize mostly with other entrepreneurs in their small isolated group. For v ≤ V(1, 0), the transformation curve and the curve with minority specialization in Figure 1 coincide.⁸ Group A has absolute and comparative advantages as self-employed entrepreneurs in industry 1. If the demand for industry 1 is sufficiently great, however, then the minority is too small to satisfy demand by themselves. In the special case when v = V(0, 1), the demand for industry 1 is great enough for group B to specialize completely. In this case minority involvement would dilute the majority’s productivity advantage, and the Pareto efficient solution is for Bs to specialize in being self-employed entrepreneurs in industry 1.

Thus, the relationship between group size and productivity is not monotonic, and the group with the absolute advantage is the group with a population size that most closely adheres to the size of industry 1. Other production possibilities generated by more unspecialized distributions, such as X_A = X_B, are not displayed in Figure 1. Our theoretical Online Appendix proves that a convex production function in social interactions (θ″ > 0) is sufficient to ensure that at least one group specializes, in which case the efficient frontier is the outer envelope of the curves shown in Figure 1. Consequently, above a certain value of v, there is a discrete jump from minority specialization to majority specialization.

C. Model Discussion

This simple model provides a stark economic environment for considering how isolated social interactions affect the sorting of ethnic groups over industries. Although our model considers only two industries, this simplification is not as limiting as it may first appear. The model captures a setting where a small industry of self-employed entrepreneurs can benefit through nonwork interactions. Allowing the baseline industry 0 to be an aggregate of many constant-returns-to-scale industries would still lead to the efficient solution being for the small ethnic group to specialize in being the self-employed entrepreneurs if their group size matches the demand preferences for industry 1. In fact, framed this way, the baseline industry 0 would be expected to be quite large to any one industry, making it more likely that the minority group should specialize.

Another obvious simplification is that we only have two ethnic groups. Yet a complex model allowing for several small industries and also several minority ethnic groups would lead to the same conclusions. For example, consider an economy with industries 1a and 1b that have equal demand and display the same productivity benefit for social interaction. Also allow there to be two minority groups of equal size. If the demands for industries 1a and 1b are sufficiently small, then the efficient outcome is for one minority group to specialize in being self-employed entrepreneurs in 1a and for the other minority group to specialize in 1b. Which minority group specializes in which sector is arbitrary. In this multi-sector economy with sector-specific skills, otherwise-similar groups consequently specialize in different business sectors. Pushing further, if the economy has several small industries of varying sizes that benefit from these social interactions and multiple minority ethnic groups, the efficient outcome will be characterized by minority groups specializing in specific self-employment industries as much as possible.

Our Online Appendix also provides several formal extensions to the model. We analyze competitive market outcomes and dynamics and show that initial conditions matter. Social interaction will reinforce early concentrations by attracting members of some groups and pushing out others. We also demonstrate that a small group size is inherently more likely to result in high initial concentrations in one or more industries that can then become reinforced and propagate. This reinforcing mechanism and the growing stratification over time are important features, as many referral models instead show decay over time due to imperfect transfer and a lack of a sustained earnings advantage.

An additional extension considers individual heterogeneity in ability and earnings and predicts that an ethnic group can achieve greater earnings at the group level when specializing. The prediction becomes more complicated for entrepreneurs vs. wage workers within groups as it depends upon how high- vs. low-skilled members of the ethnic group are attracted by the gains from social interaction. The empirical work of Patel and Vella (2013) shows a positive earning relationship for immigrant groups and common group occupational choices, and we note below some complementary evidence from our own data. This earnings premium provides evidence that the choice to engage in self-employment and specialize is due to more than just discrimination against minority groups, which could still nonetheless play a role, and it helps distinguish the theory from being just about referral networks for opportunities.⁹

A final extension looks at endogenous interactions. Although our simple model takes social ties as given, in the extension we look at endogenous social interaction and show how a social network is formed through matching in a marriage market where social traits are diverse. We explore the potential for splinter groups to break out of the majority group in order to benefit from the increasing returns to social interaction in our model. Drawing on results from graph theory, we show that there are no such splinter groups in a first-best matching on social traits only. This demonstrates that there would be costs in terms of deteriorated matching quality if the majority were to duplicate the social structure of an (exogenously) isolated ethnic minority. Ethnicity consequently matters and can confer a productive advantage for self-employment even when interaction is endogenous.

A. U.S. Census of Populations Data

We analyze the 2000 Census of Populations using the Integrated Public Use Microdata Series (IPUMS). We focus on the 5 percent sample, and we use person weights to create population-level estimates. In a panel exercise, we also use the 5 percent samples from 1980 and 1990 and the five-year American Community Survey (ACS) samples for 2006–2010 and 2014–2018. We will refer to the latter two data sets as the 2010 ACS and 2018 ACS, respectively. In addition, we build instruments from 1991 information on the United Kingdom and 2011 information on Spain from IPUMS-International.

We define ethnic groups using birthplace locations and, in a few cases, language spoken. We merge some related birthplace locations (for example, combining England, Scotland, Wales, and nonspecific U.K. designations into a single group). We also utilize the detailed language variable to separate Gujarati and Punjabi Indians and to identify Armenians and Chaldeans, given their prominence. Our preparation yields 131 ethnic groups from 198 initial birthplace locations. Online Appendix Tables 1a and 1b list all ethnic groups and provide descriptive statistics on them.¹⁰

We assign industry classification and self-employment status through the industry and class-of-work variables. IPUMS uses a three-digit industry classification to categorize work setting and economic sector of employment. Industry is distinct from an individual’s technical function or occupation, and those operating in multiple industries are assigned to the industry of greatest income or amount of time spent. The class-of-work variable identifies self-employed and wage workers,¹¹ and we examine {industry, class of work} pairings. For example, a self-employed hotelier is classified differently than a wage earner in the hotel industry. The sample excludes those whose self-employment status is unknown and industries without self-employment.¹²

Our core sample focuses on males aged 22–70 and not living in group quarters. For immigrants, we require that they have migrated to the United States at age 16 or older. Our final sample for 2000 contains 2.9 million observations, representing 59 million people. Of these, 0.26 million observations, representing 5.7 million people, are immigrants.

B. Clustering in Entrepreneurial Activities

We design “overage” ratios to quantify for an ethnic group the heightened rate of self-employment displayed for a particular industry and also across the full range of industries. Our primary metrics focus on the specialization evident among self-employed individuals only, although robustness checks build samples combining wage earners and self-employed.¹³

We first define OVER_lk as the ratio of an ethnic group l’s concentration in an industry k to the industry k’s national employment share. Thus, if an ethnic group l has N_l total workers and workers in industry k, then and . This baseline metric measures the over- or underrepresentation of the ethnic group for a specific industry, and by definition both cases exist for an ethnic group across the full range of industries.

To aggregate these industry-level values into an overall measure of industry concentration for an ethnic group, our primary metric takes a weighted average using the share of the group’s self-employment by industry as the weight: (4)

Intuitively, the metric is similar to a Herfindahl–Hirschman index with an underlying adjustment for different industry sizes. Our estimations ultimately transform OVER1 to have unit standard deviation for interpretation. We also test the following variants:

1. Weighted average over the three largest industries for ethnic group , where k’ = k, such that is maximized.

2. Weighted average over the three largest industry-level overages for ethnic group , where k’ = k, such that is maximized.

3. Maximum overage: OVER4_l = max_l[OVER_lk].

We investigate our entrepreneurial concentration hypotheses over the 131 ethnic groups using the metrics. OVER1_l takes the weighted sum across industries, while OVER2_l considers the three largest industries for an ethnic group. In most cases, OVER2_l is bigger than OVER1_l as concentration is often linked to substantial numerical representation; some exceptions happen when an ethnic group is focused on bigger industries. These calculations measure extreme values, and we need to be careful about small sample size, especially for OVER3_l and OVER4_l given their emphasis on outliers. We will thus focus mostly on OVER1_l and also conduct Monte Carlo simulations of expected overage described later. We will also show the results are robust dropping ethnic–industry pairs with very limited observations.¹⁴

Figure 2 displays ethnic groups with the highest and lowest OVER1_l metrics. There is substantial entrepreneurial clustering, with immigrants from Nepal (40.7), Senegal (37.0), Zimbabwe (36.5), and Yemen (36.3) displaying the overall highest industrial concentration for entrepreneurship. The national average for ethnic groups is 8.4, and lowest concentration rates are for immigrants from Poland (1.6), Germany (1.6), Canada (1.6), and Cuba (1.4). Online Appendix Tables 1a and 1b give a detailed list of overage ratios for each ethnic group and the industries with the largest overage ratio. In most cases, the industry where the ethnic group displays the highest concentration for self-employment is the same as the industry where the ethnic group shows the highest concentration for total employment. Online Appendix Tables 2a and 2b document the strong correlations among the overage metrics.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Entrepreneurial Specialization by Ethnic Group

Notes: Figure shows the weighted average overage for the entrepreneurial concentration by ethnic immigrant group. The top 20 and bottom 20 values are shown, along with the national average.

C. Social Isolation and In-Marriage Rates

We measure social isolation and concentrated group interactions through within-group marriage rates for child arrivals to the United States evident among ethnicities. This metric is a strong proxy if sorting in the marriage market is similar to sorting in other social relationships.¹⁵ High marriage rates within an ethnic group, also termed in-marriage, suggest greater social isolation and stratification. Significant levels of in-marriage are often present in minority groups and along religious lines, with members of the ethnic group devoting more energy towards interacting with coethnics and ultimately transmitting the group’s traits to future generations (for example, Bisin and Verdier 2000; Bisin, Topa, and Verdier 2004). Such choices can come at the expense of better access to the formal labor market that can come through intermarriage with natives (Furtado 2010).¹⁶

We calculate in-marriage rates for ethnicities using a second data set developed from IPUMS. We focus on women and men who immigrated to the United States when 0–15 years old and who are aged 22–70 at the time of the census. Importantly, this sample is mutually exclusive from the earlier sample used to calculate our overage metrics, where we consider men who migrated at age 16 or older. By focusing on children at the time of migration, we also circumvent the joint migration of married couples to the United States.

Most immigrant groups are socially segregated with respect to marriage, some very strongly so. With random matching for marriage and equal male and female migration, in-marriage rates would roughly equal a group’s fraction of the overall population. Group in-marriage rates (also shown in Online Appendix Table 1a) average 48 percent and often exceed 80 percent. Pairwise correlations of 0.31 and 0.45 exist for in-marriage rates and the OVER1l and OVER2l metrics, respectively. We later introduce some alternative metrics for social isolation.

D. Ordinary Least Squares Empirical Results

To quantify whether smaller and more socially isolated ethnic groups have greater industrial concentration for entrepreneurship, we use the following regression approach: (5) where SIZE_l is the negative of the log value of group size and ISOL_l is the in-marriage rate of the group. We take the negative of group size so that our theoretical prediction is that β₁ and β₂ are positive. We report all coefficients in unit standard deviation terms for ease of interpretation. Our baseline regressions winsorize variables at their 1 percent and 99 percent levels to guard against outliers, weight estimations by log ethnic employment for each group, and report robust standard errors .***,**, and * indicate statistical significance at the 1 percent, 5 percent, and 10 percent levels, respectively.

Column 1 of Panel A in Table 1 measures that a one standard deviation decrease in group size is correlated with a 0.58 standard deviation increase in average entrepreneurial concentration across all industries. Similarly, a one standard deviation increase in the in-marriage rate is correlated with a 0.33 standard deviation increase in overage. Panel B introduces controls for the traits of the ethnic group in 2000: share who are aged 36–55, share who are aged 55–70 (reference group is aged 22–35), share who are married, share who speak English well, share who have some college education, and share who have a college degree or higher (reference group is high school or less). The coefficients are more equal at 0.47 and 0.45, respectively, in the presence of these controls.

The next columns consider robustness checks on our metric design. Column 2 considers the metric that uses all employed workers for the ethnic group. Column 3 compares industry-level overages only to rates of other immigrant groups by excluding natives from the calculations of industry sizes. Column 4 drops ethnic–industry settings where fewer than three observation counts exist. Column 5 excludes new arrivals to America during the prior five years, as some forms of employer-based migration are tied to specific jobs. Column 6 excludes the taxicab industry, which is a frequent industry of maximum overage. The coefficients are stable across these variations.

Table 2 continues with additional robustness checks on the OVER1_l outcomes. Columns 2 and 3 drop sample weights and winsorization steps, respectively. Column 4 introduces fixed effects for each origin continent, Column 5 uses a median regression format, and Column 6 bootstraps standard errors. Columns 5 and 6 should be compared to Column 2 given their unweighted nature. Column 7 adds an additional control to capture any mechanical relationship between ethnic group size and entrepreneurial overage. For each ethnic group we conduct 100 Monte Carlo simulations using the same count of self-employed as observed for the group but randomly picking the industry in accordance with the aggregate U.S. distribution for self-employment. From these simulations, for each ethnic group we calculate the average expected overage. Introducing these controls does not significantly impact our estimations, except that the size relationship diminishes modestly.¹⁷

Table 3 shows our other forms of the overage metric. Column 2 shows that a focus on the three largest industries for an ethnic group (that is, OVER2 discussed above) increases the relative importance of social isolation for predicting overages. Columns 3 and 4 examine extreme values using the OVER3_l and OVER4_l metrics defined above. The estimates remain statistically significant and now show a smaller connection to group isolation relative to group size.¹⁸

Online Appendix Tables 3a and 3b further test the relationships of relative size and isolation on entrepreneurial clustering by using nonparametric regressions. We partition our size and isolation variables into terciles and create indicator variables for each combination of {smallest size, medium, largest size} and {most isolated, medium, least isolated}, and we assign ethnic groups that fall into [largest size, least isolated] as the reference category.

The results continue to support the theory, as depicted in Figure 3. The [smallest size, most isolated] groups have entrepreneurial concentrations that are 1.8 standard deviations greater than the [largest size, least isolated] groups. Equally important, the pattern of coefficients across the other indicator variables shows the relationships are quite regular and not due to a few outliers. For example, holding the ethnic group size constant, higher levels of social isolation strongly and significantly correspond to larger overages. Flipping it around and holding social isolation constant, smaller group sizes mostly promote greater concentration within each isolation category.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Nonparametric Estimations of Specialization

Notes: See Table 1. Ethnic groups are divided into equal-sized bins based on group size and social isolation using terciles. Within each isolation triplet, groups are ordered smallest to largest as shown for the most isolated groups. Effects are measured relative to largest and least isolated ethnic groups. Coefficient estimates and 95 percent confidence bands are reported. Full results are provided in the Online Appendix.

E. Panel Data Models and Assimilation

We next consider panel estimations to remove time-invariant features of the data. Some ethnic groups may face persistent discrimination that contributes to both social isolation and entrepreneurial specialization. This could be particularly true for nonwhite immigrants, who feature prominently in Figure 2. Our cross-sectional results could also be overly dependent on a single wave of migration to the United States, possibly to fill short-term needs around the year 2000, and thus could be incomplete for the longer-term dynamics we hope to capture. Showing similar results with a different source of identifying variation provides greater confidence in our estimations, and we can use panel models also to study the process of immigrant assimilation and the persistence of entrepreneurial specialization.

Table 4 extends our work to a panel model covering 107 ethnic groups over the five time periods of 1980, 1990, 2000, 2010, and 2018. The 24 excluded groups lack information for one or more years because of changes in the birthplaces recorded in IPUMS. Preparation steps are consistent across the time periods, and the controls for ethnic groups’ traits are time varying as well in Panel B. We cluster standard errors by ethnic group.

Column 1 finds a longitudinal size relationship that is much stronger than that observed with the 2000 cross section, although the group isolation is comparable in economic magnitude. Column 2 adds the control for expected overage based on Monte Carlo simulations with ethnic observation counts in each year. With this control, the results look even more like those measured in the cross section. Column 3 adds a linear time trend interacted with the 1980 level of overage as an alternative control strategy. Overall, the panel data model is quite consistent with the results present in the 2000 census.

The process of assimilation of new arrivals receives great attention in the immigrant literature. Our model of entrepreneurial specialization does not undertake a detailed treatment of the issue and how later generations can be affected. It would be feasible, for example, for entrepreneurial specialization to weaken assimilation, being statically efficient and dynamically inefficient by creating “cul-de-sacs” of entrepreneurial specialization that limit further assimilation (for example, Andersson Joona and Wadensjö 2009). Furtado and Song (2005) also speak to the growing wage premiums connected to marrying a U.S. native since 1980. On the other hand, greater earnings with entrepreneurial specialization can be a route for new immigrants to afford better educations and future career opportunities for their children.

The results in Table 4 shed some light on this issue. First, the panel coefficient for social isolation is very similar to the cross section. This suggests that continued assimilation of an ethnic group into the United States as measured by reduced in-marriage rates would be connected to continued declines in entrepreneurial clustering. That said, the data suggest that this is not happening for many ethnic groups. From the 1980 and 1990 censuses to the 2010 and 2018 ACS, the measured in-marriage rates among child arrivals to the country increased on average by eight percentage points. Indeed, it may be difficult to find same-origin partners in small groups, leading to in-marriage rates increasing as the group grows in size.

Additionally, the unreported age controls for the group in Panel B capture the aging of the migrants in the United States. Conditional on in-marriage rate adjustments, aging as captured by these controls does not connect very strongly to lower entrepreneurial clustering. This is similarly true when considering changes over decades in the share of the ethnic group that has been in the United States for longer than 15 years. Future research with data that combine the records of parents and children can further investigate the assimilation outcomes and long-term consequence of entrepreneurial clustering by first-generation immigrants.

We next consider two complements to the panel model. We have established a tight empirical relationship of the in-marriage rate to ethnic entrepreneurial specialization, but we should consider other measures of social isolation. We undertake this comparison next to better ground the use of the in-marriage rate and learn more about other types of social distance between groups. We then test for reverse causality concerns: for example, that growing entrepreneurial specialization leads to more in-marriage among the ethnic group. For this, we use instrumental variables (IV) models that exploit sources of variation outside of the United States.

F. Additional Measures of Social Isolation

Table 5 considers additional measures of social isolation. We first measure the residential segregation of the ethnic group. Ethnic enclaves can be important early homes for new arrivals, with links to social isolation like those we measured via in-marriage rates. Although residential segregation could generate self-employment activity to satisfy local consumer demand of the ethnic group, extensive specialization of entrepreneurial activity would require serving customers from other ethnic groups. Many common industries of entrepreneurial specialization, such as taxi drivers, construction and building trades, and landscape services, could be well aligned with self-employed members traveling to other local areas to serve customers.

Our data here are limited to exploiting the Public Use Micro Areas (PUMA) of residence within metropolitan areas captured by the 2000 census. We only consider metro areas with more than one PUMA, and we calculate residential segregation for an ethnic group relative to 100 randomized counterfactuals that considered if an equivalent number of census observations were drawn at random in proportion to local population from PUMAs in the same metropolitan areas where the ethnic group resides. Transformed to have unit standard deviation for comparability, residential segregation is also a strong predictor for entrepreneurial clustering in Column 2 and with an economic magnitude comparable to the in-marriage rate.

Columns 3–5 alternatively take data from Spolaore and Wacziarg (2016) on the genetic, linguistic, and religious distance of countries to each other. We applied these country-based distances to our setting by measuring a weighted average for an ethnic group from the ethnic composition of the United States as measured by country of birth for U.S. residents. Metrics are again expressed in unit standard deviations. Regressions cluster standard errors by 120 unique observations from Spolaore and Wacziarg (2016) that we map to our sample. Although we can map measures of genetic distance for our full sample, linguistic and religious distances are only available for 113 groups (112 overlapping).

Without controls for ethnic group traits, genetic and religious distance most closely connect to entrepreneurial clustering, although linguistic and religious distance are strongest in the presence of the controls. When combining all of our measures together in Column 6, in-marriage rates stand out, with genetic distances also being important in Panel A. These results, in combination with their longitudinal consistency in Table 4, suggest that our measure of social isolation via in-marriage rates captures a salient part of the group’s social dynamics that is not just due to residential segregation, linguistic isolation, or an even more fixed component like genetic distance.

F. Instrumental Variables Empirical Tests

We next consider IV specifications to test against reverse causality concerns (for example, where isolated business ownerships lead to greater social isolation or lower group sizes) or omitted variables. Some omitted factors could center on sector-specific skills gained by ethnic groups abroad that are then ported to the United States with migration (especially if booming local demand for an ethnic group’s services leads them to draw more migrants with similar skills from their home country). Others could be due to local traits, such as state-level adoption of stringent employment verification procedures (for example, Amuedo-Dorantes and Bansak 2012; Orrenius and Zavodny 2016) leading to more social and workplace isolation.

Our primary IV approach uses as instruments the predicted ethnic group size from a gravity model and in-marriage rates from the United Kingdom in 1991. To instrument for ethnic group size, we use a gravity model to quantify predicted ethnic size based on worldwide migration rates to the United States. The original application of gravity models was to trade flows, where studies showed that countries closer to each other and with larger size tended to show greater trade flows, similar to the forces of planetary pull. This concept has also been applied to the migration literature, and we similarly model (6) where DIST_l is the log distance to the United States from the origin country and POP_l is the log population of the origin country. For this purpose, we estimate log ethnic group size in the United States as the dependent variable (without a negative value being taken as in earlier estimations). Unsurprisingly, lower distance (β₁ = −1.43, SE = 0.24) and greater population (β₂ = 0.42, SE = 0.05) are strong predictors of ethnic group size in the United States. We take the predicted values from this regression for each ethnic group as our first instrument.

For our instrument of in-marriage rates in the United States, we calculate the in-marriage rates in the 1991 U.K. Census of Populations. This approach is attractive as the social isolation evident in the United Kingdom a decade before our study is most likely to be predictive of U.S. self-employment rates to the extent that the British isolation captures a persistent trait of the ethnic group. The instrument is not completely foolproof (for example, a third factor like specialized ethnic-specific skills could be present in the diaspora in both countries and lead to similar outcomes), but the instrument does provide assurance against some of the most worrisome endogeneity arising in local areas. A limitation of this instrument is that we are only able to calculate it for 34 broader ethnic divisions. We map our observations to these groups and cluster the standard errors at the U.K. group level.

The first-stage results with this instrument set are quite strong. The first two columns of Table 6 show that these instruments have very strong individual predictive power with and without the ethnic group controls. The second-stage results in Column 3 are similar to the ordinary least squares (OLS) findings. The IV specifications in Panel A suggest that a one standard deviation decrease in ethnic group size increases overage by 0.46 standard deviations. A one standard deviation increase in isolation leads to a 0.32 standard deviation increase in entrepreneurial concentration. These results are well measured and economically important. The results are close enough to the OLS findings that we cannot reject the null hypothesis in Wu–Hausman tests that the instrumented regressors are exogenous. These IV results strengthen the predictions of our theory that smaller, more isolated groups are more conducive to entrepreneurial clustering.

Ideally, we would be able to build a broader instrument that used in-marriage rates from many countries for an ethnic group. This would help counteract any persistent bias due to similarities for immigrant experiences in the U.K. and U.S. economies, and it would overcome measurement error in the instruments. Unfortunately, the data requirements for our in-marriage rate calculation are steep, especially for knowing detailed countries of birth of spouses within a household, and the only additional source we could identify from IPUMS International is Spain 2011. These data have 60 ethnic origin groups that we can map to the U.S. data.

In Columns 4–6, we use average in-marriage rate for an ethnic group from the U.K. 1991 and Spain 2011 as instruments for U.S. 2000 in-marriage rates. As anticipated, the results are a bit sharper, and, due to the growth of the isolation coefficient in the second stage, we are now more likely to reject that the instrumented regressors are exogenous. We remain cautious of the Spain instrument but take comfort in the overall stability evident in this modification.¹⁹

Online Appendix Tables 5a–8b show robustness checks to the instruments. Results are very similar with simple adjustments like excluding sample weights and dropping winsorization. Some results for the social isolation metric have larger standard errors when bootstrapping and including ethnic group controls, which is not too surprising given the smaller number of underlying U.K. groups. Another weak spot is that the expected overage controls from simulations can crowd out the size instrument in a dual IV as the instrument and predicted overage are being built upon the same data, making it hard to separate them. Beyond these caveats, however, the IV is quite robust overall. We also find very similar results when expanding the gravity equation to have a squared distance term or an indicator for Canada and Mexico as bordering countries or when using underlying components of the gravity equation as direct instruments.²⁰

G. Extension: Earnings

Our model predicts that members of an ethnic group can achieve greater earnings when entering a common entrepreneurial setting. In our framework, social complementarities produce a positive relationship between earnings and entrepreneurship at the group level. Evidence for this prediction helps show discrimination is not solely responsible for our findings, and this also helps differentiate our work from job search networks. To the extent that our person-level controls on education and language fluency capture skill levels, we may also anticipate that self-employed individuals earn more.²¹ This net relationship must be empirically investigated in the data, and an earnings premium for self-employed workers would provide evidence against the entrepreneurial clustering being due to herding behavior or other forms of inefficient entry.

Patel and Vella (2013) comprehensively show a positive earning relationship for immigrant groups and common group occupational choices using the 1980–2000 Census of Populations data. Table 7 provides complementary pieces of evidence that look at variation within metropolitan statistical area (MSA)–industry cells and within ethnic groups. As in our prior estimations, the sample includes immigrant males who arrived in the United States after age 16 and are aged 22–70 in 2000. The outcome variable is log annual income.²² Estimations include fixed effects for the following person-level traits (category counts in parentheses): age (5), age at immigration (5), education (4), and English language fluency (2). Regressions use person weights and cluster standard errors by ethnic group. Explanatory variables are transformed to have unit standard deviation for easy comparison and interpretation.

Panel A considers self-employed individuals, and Panel B considers wage workers. The first column simply considers the share of an individual’s ethnic group who are self-employed in the industry of the focal worker. There is a positive relationship for both worker types, even conditional on MSA–industry fixed effects. For the self-employed, a one standard deviation increase in the concentration of the ethnic group for self-employment in the industry is associated with about a 7 percent increase in annual earnings. For wage workers, the relationship is measured to be 4 percent.

Column 2 adds into the estimation the overall share of the ethnic group who are self-employed—which is very predictive of group earnings, per the model—and Column 3 further adds the total ethnic group employment in the focal industry. Columns 4 and 5 add ethnic group fixed effects, which absorb the group’s overall rate of self-employment and focus on variation across industries within each ethnic group. Looking across these estimations, there is strong confirmation of the model’s prediction that members of an ethnic group can achieve greater earnings through entrepreneurial clustering. The whole group earns more when entrepreneurial activity is higher, and the earnings of the self-employed in an industry show a tight relationship to other members of the ethnic group being self-employed in the same industry space.

The split-sample approach in Table 7 does not quantify whether self-employed earn more than immigrant wage workers in the same setting, as the fixed effects and controls can change values. Online Appendix Table 9 shows a combined analysis with self-employment interactions and groups traits, thus requiring control variables to have the same values. These estimations confirm that within the same MSA–industry cell and conditional on covariates, the self-employed do earn more. Given the challenges for measuring entrepreneurial income noted earlier, this differential is likely also an underestimate.

These results support the model’s structure and are consistent with a potential positive benefit from immigrant entrepreneurial concentration. It is important for future theoretical and empirical work to consider both owners and employees of firms. Empirical work could particularly target employer–employee data sets to observe more detailed hiring and wage patterns; such work could also evaluate job transitions during the assimilation of new members of ethnic groups, perhaps ultimately leading to starting their own business.

H. Extension: Industry Variation

We conclude our analysis with two extensions that consider industry and metropolitan variations. The Pareto version of our model, presented in Section II, makes the compelling prediction that ethnic groups should match in terms of size with the industry of self-employment; that is, smaller ethnic groups are a better fit for small self-employment industries, and larger groups should be in larger sectors.²³

Figure 4 shows descriptive evidence in this regard. We plot for five aggregated groups the cumulative distribution in self-employment as we move from the smallest industries for self-employment, starting with “petroleum and coal products” (left-hand side, #1) to the largest industry of “construction” (right-hand side, #126). The solid line captures the self-employment distribution of U.S. natives. We parse immigrant ethnicities into four equal-sized groups based on whether they are above/below the median social isolation and group size.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Industry Distributions of Self-Employment

Notes: Figure shows the cumulative distribution of self-employment for groups moving from the smallest [#1] to largest [#126] industries for self-employment. Immigrant ethnic groups are divided equally into four groups based on being above or below median group size and social isolation.

The figure visually aligns with the model’s prediction. All immigrant groups are shifted to the left of the cumulative distribution of U.S. natives, indicating a greater share of self-employment work in smaller sectors. The smallest and most isolated ethnic groups are the most concentrated in smaller industries, followed by the smallest and least isolated ethnic groups. The figure highlights some of the industries where concentration emerges (for example, taxis, grocery stores, physicians, eating and drinking places).²⁴

Table 8 confirms these patterns with regressions, including adding controls in Panel B for ethnic group traits. Columns 1 and 2 show that smaller and more isolated groups have their self-employment activity concentrated in industries with smaller sizes as measured in terms of self-employed workers only or all workers, respectively. Columns 3 and 4 find similar results when isolating the largest industry of self-employed workers for an ethnicity. Columns 5–8 show these results are not present for wage workers. The wage worker results are an interesting extension beyond our model as they suggest the coethnic hiring of immigrants, which has been frequently observed, is not so extensive as to replicate the industry concentration pattern that is experienced for self-employment.

At an aggregated level, we can also use the industries in Figure 4 to provide some calculations broadly consistent with the model’s mechanism of interactions. The 2016 Annual Survey of Entrepreneurs (ASE) asked entrepreneurs and small business owners their sources of advice for business. The publicly available ASE data are only available at the two-digit NAICS level, so we compare “accommodation and food services” (NAICS 72) to “construction” (NAICS 23) and a composite of other industries. Table 9 shows that entrepreneurs and small business owners in “accommodation and food services” report the greatest likelihood of collecting advice from customers, family, and friends. “Construction” has a higher reported reliance on colleagues, and “legal and professional advisors” feature more strongly among other two-digit NAICS sectors.²⁵

I. Extension: Metropolitan Variation

We close our study by examining variation in ethnic group size across metropolitan areas. Our theoretical framework is built around a single economy and does not include spatial variation. Although some industries of self-employment concentration are spatially distributed by nature (for example, the concentration in motels by Gujarati Indians), many industries like taxis and landscape services are oriented towards local markets. This localization of service does not necessarily prevent a group from consistently specializing in an industry, as there can be sharing over communities and regional gatherings. Also, Basso and Peri (2020) show that the most recent immigrant arrivals have the highest rates of internal migration across locations within America. Such migration can transport a local specialization to new locations, such as the spreading of Vietnamese nail care salons and Gujarati motels from their points of origin in California in 1975 and 1942, respectively.²⁶

To examine whether local group size connects with local entrepreneurial clustering, Table 10 presents regressions with group size measured at the metropolitan level.²⁷ We include metropolitan fixed effects to control for the overall scale of local activity, and we control for the in-marriage rate measured nationally.²⁸ Column 1 provides the estimation with size by itself, and Columns 2 and 3 add the expected overage based on Monte Carlo simulations for the local ethnic group observation count. Column 3 further adds ethnic group fixed effects. Across these specifications, there is again very consistent evidence that smaller ethnic group size is connected to greater entrepreneurial clustering. We hope that future research can develop frameworks to quantify jointly industry and geographic spans for entrepreneurial concentration of ethnic groups and their dynamics.