What the U.S. Data Actually Show
The U.S. evidence asks a precise question about newly founded firms: after comparing ventures in similar industries, places, and founding cohorts, is the composition of their employees still related to the background of their founders? This chapter focuses on how that relationship is measured and what its size means. The goal is not to memorize a list of coefficients, but to learn how to read a founder–worker association without confusing a broad pattern with a causal rule about any individual hire.
Two data designs, two kinds of visibility
The strongest U.S. evidence comes from the Longitudinal Employer–Household Dynamics, or LEHD, data used by Josh Feng and Xavier Jaravel. The data combine administrative wage records with demographic information linked from the Decennial Census, as well as industry and location. That combination gives the researchers something unusual: a large population of workers whose earnings and demographic characteristics can be followed inside firms over time.
The design begins by identifying new firms in business-register data. To reduce the risk of treating an old company’s new establishment as a new venture, the researchers remove firms that had earlier entries in other Census datasets. They then identify the top three earners in a company’s first year and use their demographic profiles to construct a founder measure. Everyone else observed at the firm is treated as a non-founder employee for the composition analysis.
That last step is useful, but it needs to remain visible. Administrative data do not directly announce which people founded a company. A high first-year earner may be a founder, but they may also be an early senior employee. The authors therefore use the top-three classification as an operational proxy. They test one- and two-earner alternatives later; those checks matter because a result that disappeared when the proxy changed would be less persuasive.
The main LEHD sample is a balanced panel of approximately 825,000 firms founded between 1992 and 2014. Each firm has at least one non-founder employee in every year from founding year through year five, along with the industry and metropolitan-area information needed for detailed comparisons. At the firm-year level, the data contain roughly 4.947 million observations. The balanced requirement is not a minor technical detail. It prevents the measured change in workforce composition from being driven simply by the fact that some kinds of firms exit before later years. But it also selects firms that survive and continue to employ people, so it should not be read as a census of every business that was ever started.
The LEHD data also cover only 30 states and the District of Columbia. That is a major advantage over a small convenience sample, but it still leaves questions about representativeness across the United States. The paper reports rounded counts because of Census disclosure rules. These facts do not invalidate the estimates; they define the population to which the estimates most safely apply.
The second U.S. design links Crunchbase startups to LinkedIn work histories compiled by Revelio Labs. It sacrifices some administrative coverage in exchange for more direct information about roles. Founders can be identified from titles such as “Founder,” “Co-Founder,” or “Founding CEO,” and employee histories include occupation and seniority. This lets the researchers ask whether a founder–worker relationship appears among engineers, scientists, senior employees, or other roles that the LEHD data do not label as cleanly.
That visibility comes with selection. LinkedIn profiles are not a complete record of the workforce, the demographic variables are inferred from names rather than directly reported Census information, and the population is heavily tilted toward college-educated people. The balanced Crunchbase-linked sample contains approximately 44,545 firms founded in or after 2007 with non-founder profiles in each of the first six years. It is also tilted toward venture-backed and technology firms. The two datasets therefore do not duplicate one another. LEHD supplies scale, wages, and broad demographic coverage; LinkedIn–Crunchbase supplies role and seniority detail. Agreement across them is useful, but neither removes the other’s limitations.
The solid path shows how each data lane reaches a comparable outcome. The dashed links are not competing results; they mark the measurement costs attached to each lane. A useful empirical finding is strongest when its interpretation keeps both the estimate and the cost of observing it in view.
A large association remains after closer comparison
The central comparison is a share-on-share relationship. For a given firm and year, the researchers calculate the fraction of employees with a characteristic—female, Black, Hispanic, Asian, a particular age profile, or a college education—and relate it to the corresponding fraction among the firm’s founders. They then repeat the comparison within increasingly detailed cells defined by industry, geography, and founding cohort.
This is what the paper calls homophily: founder and worker composition move together. In ordinary language, firms with more founders from a particular group tend to have more workers from that group. The term describes a pattern; it does not identify the reason for it.
The coefficient is easiest to understand with a hypothetical comparison. Suppose two otherwise comparable firms differ by ten percentage points in the share of founders who are women. A female-founder coefficient near 0.37 implies an estimated difference of about 3.7 percentage points in their female employee shares, holding the comparison framework fixed. It does not say that a female founder causes a particular person to be hired, that 37 percent of all hires are determined by founder identity, or that the relationship must hold for every pair of firms.
In the founding year, the most detailed LEHD specification reports coefficients around 0.37 for female founders, 0.556 for Black founders, 0.406 for Hispanic founders, and 0.574 for Asian founders. The relationship also appears for age and education, with coefficients around 0.364 for founder age and 0.118 for the college-educated founder share. The broad lesson is more important than the ranking: the association spans gender, race and ethnicity, age, and education rather than appearing as a single isolated gender result.
The paper also expresses some of these patterns as percentage-point differences between firms with different founder compositions. That presentation can be intuitive, but it is not interchangeable with the regression slope. A slope describes how employee share changes as founder share changes across the observed sample; a percentage-point comparison depends on the particular founder profiles being compared. Keeping that distinction prevents a vivid headline from being mistaken for a universal effect size.
The association becomes smaller when the comparison is made more demanding, but it does not disappear. Broad industry differences and regional labor markets clearly matter, so the researchers compare firms within detailed industry-by-cohort and industry-by-metropolitan-area-by-cohort cells. The remaining relationship says that firms with similar broad surroundings can still have different employee compositions associated with their founders. That is stronger than a raw comparison between, say, a technology company in one labor market and a retail company in another. It is not the same as having randomized founders.
The founder definition provides a useful stress test. When the pooled six-year analysis classifies only the highest-paid first-year worker as a founder, the female coefficient is about 0.162. With the top two, it rises to about 0.277; with the default top three, it is about 0.364. The increase is plausible if three people capture more of the initial leadership group than one person does. It also shows why the coefficient should not be described as a pure personal trait effect. The operational definition changes the object being measured—from one leading earner toward a broader founding leadership.
Even the conservative one-founder specification produces positive and quantitatively meaningful associations across the demographic dimensions. Weighting the employee shares by earnings and giving larger firms more influence produces estimates close to the unweighted results: about 0.360 for female, 0.544 for Black, and 0.377 for Hispanic founder shares. This makes it less plausible that the headline pattern is created only by small firms or low-paid workers. It still does not eliminate selection into the firms that remain visible in the data.
Using more granular zip-code comparisons makes the estimates somewhat smaller, as one would expect if local labor-market conditions had been only partly controlled by metropolitan-area comparisons. They remain large, however. The same pattern of “smaller but present” is the right way to read these robustness checks. They reduce some alternative explanations; they do not transform an observational relationship into an experiment.
The LinkedIn–Crunchbase results provide a second check with a different founder measure. In the founding year, female-founded firms have an estimated 17.4 percentage-point higher fraction of female employees, while Black-founded and Hispanic-founded firms show estimated differences of about 10.4 and 24.3 percentage points, respectively, in the balanced sample. After adding the available industry, state, and cohort controls, the corresponding year-zero slopes are smaller—roughly 0.144, 0.103, and 0.240—but remain positive and statistically distinguishable from zero.
The difference in magnitude between the two U.S. samples should not be treated as a contradiction. The LEHD measure includes a much broader set of new firms and uses directly linked demographic information, while the Crunchbase sample concentrates on startups tracked by a venture-oriented database and workers with LinkedIn profiles. Name-based classification can also attenuate a relationship. The more defensible conclusion is that founder–worker similarity appears under both measurement systems, with magnitudes that depend on which firms and workers can be observed.
What the evidence establishes—and what it leaves open
At this stage, the U.S. studies establish a robust descriptive relationship. Founder composition is associated with employee composition across several demographic dimensions, and the relationship survives detailed comparisons, alternative founder counts, weighting, and a second data source that identifies founder roles more directly. The result is difficult to dismiss as a simple difference between broad industries or as a pattern limited to the smallest firms.
The evidence does not yet tell us why the relationship exists. A founder may demand workers with similar backgrounds because of organization-specific productivity, shared networks, communication, or trust. Workers may be more likely to apply to, accept, or remain at such firms. The firm may serve a customer base or operate in a submarket that attracts both founders and workers from the same group. Some of these explanations can coexist.
The U.S. results also do not establish that founder background determines the quality of a job or the worth of a worker. They concern composition, not a ranking of people. Nor do they imply that mixed-background teams are impossible or that every founder hires narrowly. A positive slope across firms describes a tendency in the data; it leaves substantial variation around the tendency.
That distinction matters for how the finding should be used. For a researcher, it identifies a relationship worth explaining with wages, worker histories, and organizational mechanisms. For someone evaluating a startup labor market, it suggests that founder composition may be relevant evidence about opportunity—but not sufficient evidence about a particular employer’s practices. For a policy reader, it indicates why access to entrepreneurship could have effects beyond the people who receive capital or permission to start a business, while leaving the size and mechanism of those effects unresolved.
The next step is therefore not to repeat the coefficient in a longer list. It is to ask what happens to the association as firms add employees, lose employees, and move away from their founding moment. If the pattern fades quickly, it may describe a short-lived first impression. If it remains, the organization’s early composition may be doing more than recording who happened to be present at the start.