What the Evidence Actually Measures
This chapter asks a narrow but decisive question about founder departure and startup employee retention: what, exactly, is being measured when turnover rises after a founder leaves? The answer determines how much confidence we should place in the headline result. By separating the study’s descriptive association from its stronger sudden-loss comparison, we can read the evidence as a causal argument with assumptions—not as a simple story that founders leave and employees follow.
The distinction matters because the same event can have several explanations. A founder may leave after investors lose confidence, after a product fails, or after the organization has begun to deteriorate. Those conditions can also make employees more likely to leave. If we simply compare turnover before and after the founder’s departure, we may attribute to the founder an effect that actually belongs to the startup’s decline. The study’s design is valuable because it tries to make that alternative harder to sustain.
How the study turns a startup into a measurable case
The researchers combine administrative records about employers with records about employees and wages. A startup enters the underlying construction when it is five years old or younger, using its first year of positive payroll as the starting point. In plain terms, the firm has to appear as an operating employer rather than merely an idea or an entity on paper. Employee records then show who worked for the firm and how their employment changed over time.
The study cannot read a company’s incorporation documents or ask who thought of the original idea. It uses an operational definition designed for the available records: a founder is an individual employed from the firm’s first year who ranks among its three highest earners in that first year. Everyone else in the wage records is classified as a non-founding employee. This is a practical proxy for founder status, not a complete biography of entrepreneurship. It is likely to capture people who were central early in the firm’s paid operation, while potentially missing a founder who was not among the early top earners or including an early senior participant whose role was more complicated than the label suggests.
That limitation does not make the measure useless. It makes the unit of analysis explicit. The paper is not estimating the effect of an abstract “visionary founder” or every person who helped start a company. It is examining the employees and early high-earning workers that can be linked consistently in administrative data. Any conclusion about founder influence is therefore conditional on this operational definition.
The definition also shapes which startups enter the story. A firm must have enough paid employment for a wage ranking to be meaningful, and a founder who contributes capital, unpaid labor, or a later strategic role may not appear as one of the first year’s top earners. Conversely, an early high earner may have strong influence without being the person employees informally regard as the founder. These are not reasons to discard the result; they are reasons to interpret it as evidence about a measurable class of early central actors. If the effect appeared only for a narrower or different founder definition, that would change how broadly the conclusion should travel.
The broad departure sample contains roughly 135,000 startups that experienced a founder departure before the firm reached age ten between 2000 and 2015. In this part of the study, departure means that the founder no longer earns wages from the firm, for any reason ranging from retirement to dismissal. The researchers then construct a panel around the departure year, using up to five years before and after it when those observations are available. This produces a large view of what usually happens around founder separation.
The turnover outcome also requires care. For a firm-year, the main rate is the share of people employed at the focal firm who no longer earn wages there in the following year and are employed by another employer at higher wages. The wage increase is intended to make the measure closer to voluntary movement, since employees who leave voluntarily often move to better-paid jobs. But the records do not directly say whether a departure was voluntary or involuntary. A higher-wage move is informative; it is not the same thing as a survey-confirmed resignation.
This denominator makes the measure useful for comparing firms, but it also narrows the question. It does not count an employee who leaves the labor force, moves without a recorded higher wage, remains unemployed, or moves to another employer at similar pay in the same way. Nor does it capture the people who stay but become disengaged, lose trust, or reduce their contribution. A rise in the measured rate therefore signals a particular form of employee movement, not every form of organizational damage. Conversely, a stable rate does not prove that the remaining team is healthy.
The study’s administrative advantage is consistency: the same employment and wage rules can be applied across many firms and years. Its cost is interpretive distance. The records show that a worker’s employer and wage relationship changed; they do not show whether the worker missed the founder’s judgment, disliked the successor, saw a better opportunity, or simply responded to a broader labor-market change. Those reasons become hypotheses for later analysis rather than facts contained in the turnover variable.
What the broad comparison tells us—and what it cannot
The first result is intuitive: employee turnover is higher after a founder separates from the startup. In the preferred specification for this broad comparison, the increase is about 27% relative to turnover rates before separation. Because the sample is large and drawn from administrative records, this is more than an anecdote. It establishes that founder separation and employee movement are systematically related in the observed startups.
But a large sample does not automatically solve a causal problem. Consider a startup whose product has stalled. Investors may become dissatisfied, the founder may lose influence or decide to leave, and employees may begin searching for better jobs. In that sequence, the founder’s departure and employee turnover share a cause. The before-and-after comparison would still show a rise after departure, even if employees were responding mainly to the deteriorating business.
The same problem appears if the founder withdraws gradually. A formal separation date may be the last visible step in a process that has already changed the organization. Employees may have learned that the founder is no longer available, or that the mission and strategy are shifting, weeks or months before the wage records show an exit. A measured post-departure effect would then combine the final departure with earlier changes that are difficult to observe.
There is also a mechanical lifecycle concern. Startup turnover may change as firms age, regardless of what happens to the founder. If the post-departure years are simply later years in the startup’s life, a rising turnover rate could be mistakenly attributed to the founder’s exit. The researchers therefore do not treat the broad 27% association as the final answer. It is the reason to ask whether a more plausibly unexpected founder loss produces a similar pattern against an appropriate comparison.
Why premature death changes the comparison
The stronger design focuses on premature founder deaths, defined as deaths before age 60. The logic is not that death creates a clean laboratory. It is that an unexpected loss is less likely to be chosen because the startup has recently deteriorated or because the founder has been slowly disengaging. The researchers further restrict the events to cases in which the individual was still receiving regular wages in the quarter of death. That condition is intended to exclude cases with signs of an earlier employment break or a sustained decline before the event.
For each treated startup, the study constructs a comparison startup that did not lose a founder. The matching uses firm size, firm birth year, four-digit industry, and state, all measured in the year before the death shock. The resulting sample contains approximately 16,000 treated startups and approximately 16,000 controls, with about 221,000 firm-year observations in the unbalanced panel. The comparison is not “a startup with no risks” versus “a startup with a death.” It is an attempt to compare firms that look similar on important observable dimensions before one experiences a sudden founder loss.
Matching is a way to make the comparison more credible, not a way to make the firms identical. Two startups with the same industry code and headcount can differ in product quality, funding pressure, founder health, team cohesion, or relationships with customers. Some of those differences may be stable and are partly absorbed by firm fixed effects; others can change at exactly the wrong time. The design gains strength from combining matching with the unexpected timing of death, the pre-event trajectory, and the difference-in-differences comparison. It would be weaker if treated firms were already losing employees faster than controls before the shock, or if another unmeasured change occurred only in the treated group at the same time.
The main estimate comes from difference-in-differences. The phrase sounds technical, but the question is straightforward: did turnover change more after the founder loss in treated startups than it changed over the corresponding period in similar control startups? The control startup receives a placebo year corresponding to the year in which its matched counterpart experienced the death. This helps separate a general passage-of-time effect from a change specifically associated with the founder-loss event.
The models account for industry, state, calendar year, and the age at which the startup experiences the death or corresponding placebo event. One version also includes firm fixed effects, which absorb stable differences between firms that can be measured poorly or not at all. The reported treatment effect remains positive after these adjustments. Relative to the pre-treatment mean of the control group, the increase in turnover is about 19%.
The event-study version adds another important check. Instead of collapsing all post-event years into one indicator, it estimates the relationship separately before and after the shock. The pre-treatment estimates are statistically indistinguishable from zero, which is consistent with the parallel-trends assumption: before the founder loss, the treated and control groups were not visibly moving along different turnover trajectories. The post-event increase is strongest in the first two years and then declines.
The pre-trend result should be read as supporting evidence, not as a guarantee that the two groups were identical in every unmeasured respect. It says that the observed turnover paths did not diverge before treatment, which makes a sudden post-event divergence more consistent with the founder loss. The later decline is also open to more than one interpretation. Some employees may adapt, new leaders may build relationships, or the employees most attached to the founder may leave early. The pattern establishes timing; it does not by itself identify the process that produces the timing.
This is why the sudden-loss result is more informative than the broad before-and-after pattern. It uses an event that is less plausibly chosen by the startup’s prior condition, compares the affected firms with similar firms, checks the trajectory before the event, and observes when the effect appears. The pieces reinforce one another. None of them, however, turns the design into a randomized experiment.
The assumptions still doing work
The first remaining assumption concerns the nature of premature death. The data do not contain the cause of death, so the researchers cannot directly distinguish an entirely unexpected event from a death preceded by a long illness. The continued-wage restriction is a sensible safeguard, but it cannot reveal everything about the founder’s health or the organization’s awareness of it.
The second concerns the experience of employees. Losing a colleague suddenly can be emotionally distressing. Employees may leave because the workplace is associated with grief, not because the founder’s specific organizational role has disappeared. A later comparison with non-founder deaths helps address this possibility, but the sudden-loss design itself cannot observe emotions directly.
The third concerns the outcome. The study observes wage and employer transitions, not employees’ stated reasons. It cannot say whether each move was a resignation, a dismissal, an attempt to escape a changed culture, or an economically attractive opportunity that happened to arise after the shock. The higher-wage restriction makes the outcome more suggestive of voluntary movement, but it does not settle the classification.
Finally, the design is most directly about a sudden founder loss. A planned succession gives employees time to learn about a replacement, transfer relationships, clarify decision rights, and decide whether the company’s direction still fits them. A founder who is removed by investors, a founder who gradually withdraws, and a founder who dies unexpectedly may all produce different patterns. The evidence should travel to those settings as a question and a warning, not as an exact forecast.
The practical reading is therefore disciplined. The broad sample shows a substantial association. The matched premature-death design provides stronger evidence that sudden founder loss increases employee turnover relative to similar startups, with an estimated increase of about 19% and no visible pre-treatment divergence. The design does not tell us that every founder is indispensable, that every departure causes harm, or that age alone creates dependence. The next question is what the timing and specificity of the turnover reveal about the founder’s position inside the organization.