From Correlation to Causal Evidence
The question in this chapter is narrower than whether strong founding teams tend to perform well. It is whether losing a member of a young venture’s early team changes the firm’s subsequent trajectory, after accounting for the qualities that were already present. That distinction matters because human capital, business quality, and retention are entangled. The U.S. Census study addresses the problem with an abrupt forced-separation shock, matched comparison firms, and a difference-in-differences design. Readers should be able to say what that design makes more credible—and where its evidence still stops.
Why the descriptive relationship is not enough
The descriptive pattern is plausible. In the Census data, startups with higher-human-capital founding teams are generally more likely to survive and grow, and they tend to show stronger productivity patterns. But a relationship between team quality and firm performance has several possible explanations. High-ability people may select better ideas. They may have better access to customers, capital, or information before the firm exists. A promising venture may attract a stronger team and also be more likely to survive. A firm that begins performing well may be better able to keep its people.
Those explanations point in different directions. If the team causes the outcome, removing a member should reduce performance. If the underlying business idea causes both team quality and performance, a team comparison may exaggerate the importance of stability. If successful firms simply retain people, then stability may be partly an indicator of performance rather than its cause. Controls for industry, initial size, or initial productivity can reduce some differences, but they cannot observe every reason a person joined a venture or every quality of the idea that attracted them.
The relevant counterfactual is difficult to see directly. For a firm that kept its founding team, what would have happened if one member had left? For a firm that lost someone, what would its employment, revenue, productivity, and survival have looked like if the separation had not occurred? The causal question is about that missing alternate path, not about whether stable firms look better in the data.
A useful research design therefore needs a separation that is not primarily chosen because the firm is improving or deteriorating. It also needs a comparison firm that resembles the treated firm before the event. No design can make the alternate path observable, but a credible design can make competing explanations less plausible and show whether the treated and comparison firms were already diverging.
Imagine the counterfactual as two versions of the same young venture rather than as a comparison between a “good” and a “bad” startup. One version experiences the forced loss; the other continues without it. The second version is never literally observed, so the researchers construct the closest available comparison from firms operating in the same period and environment. The design becomes persuasive only if the comparison is similar before the event and if the post-event gap is not simply the continuation of an earlier difference.
This is also why the design focuses on an active team member rather than an employee who had already stopped contributing. The shock is meant to change the team that was currently carrying out the organization’s work. Matching on the active team count and average age helps account for exposure to the event: a firm with twelve active members has more opportunities for a death shock than a firm with two, while a younger or older team may have a different baseline risk. These variables do not capture every aspect of cohesion or capability, but they make the event less likely to be confused with an obvious difference in team structure.
Why use a premature death as the shock?
The Census researchers use premature deaths of founding-team members as an abrupt, involuntary separation. They classify a death as premature when the person is 60 or younger. For it to count as a shock to the firm, the member must have positive earnings in the quarter of death and zero earnings afterward. The researchers focus on firms in the first five years of operation, when the team is still close to the formation period. Treated firms have one premature death in that window; control firms have none.
This event has an important advantage over ordinary turnover. A member who resigns may leave because they see trouble, receive a better offer, disagree with the strategy, or know that the firm is about to change. Those motives would connect the separation to the firm’s future even before the person leaves. A premature death is less likely to be a strategic response to the startup’s prospects. It creates a sharp change in team composition that the firm did not choose as part of a growth plan.
But “less likely” is not “random.” A death can create emotional distress, managerial disruption, and financial consequences for the surviving team. The firm may also differ from an otherwise similar firm in ways that are difficult to observe. The study’s logic is therefore not that death perfectly randomizes the world. It is that an abrupt death, combined with careful comparison and evidence of similar pre-shock trends, is a more credible source of variation than ordinary decisions about hiring and departure.
The event definition also makes the estimand specific. It measures the consequence of an active founding-team member disappearing suddenly from a young employer firm. It does not directly measure the effect of a planned succession in which knowledge is transferred, a retirement after a long handoff, a voluntary departure by a poor fit, or a replacement that adds a missing capability. Those may have different consequences even if the organization ends up with the same number of people.
How the treated and control firms are constructed
For each young firm that experiences a qualifying death, the researchers seek a firm that did not experience a death and looked similar around the shock. The matching procedure requires the treated and control firms to share the same birth year, detailed four-digit industry, legal form, and state. It also matches the number of active founding-team members before the shock and the average age of those active members. These variables address several obvious sources of difference: firms at different stages or in different industries may face different survival patterns; legal form changes how owners appear in the records; and a larger or older active team has a different chance of experiencing a death.
The study then follows firms for five years after the event and compares their outcomes with a difference-in-differences framework. “Difference” appears twice. First, the analysis compares the treated firm with its matched control. Second, it compares the change after the event with the relationship between the firms before the event. Firm fixed effects remove stable differences between the matched firms, while industry-by-year effects and firm-age controls account for common changes associated with sector and age.
The event-study version makes the timing visible. The researchers estimate outcome differences for years before and after the shock rather than reporting only one before-and-after average. The pre-shock coefficients indicate whether treated and control firms were already moving apart. If the treated firms had been shrinking before the death and the controls had been growing, a later gap could not confidently be attributed to the death. If their trends are similar before the event and diverge afterward, the forced separation becomes a more plausible explanation for the change.
This timing view also prevents a common interpretive mistake. A single post-event average could conceal a brief dip followed by recovery, a delayed effect that appears only after several years, or a gap that was already widening. An event study asks whether the break occurs near the shock and whether it persists. Persistence would fit the idea that a team member carried relationships or coordination capacity; a quick return to the earlier path would be more consistent with a readily replaceable vacancy. Neither pattern alone proves a mechanism, but the timing changes which explanations remain plausible.
The outcomes are chosen to keep the comparison from depending on one narrow definition of success. Employment captures scale, revenue captures the firm’s ability to generate business, labor productivity captures output per worker, and survival captures the extensive margin of whether the firm remains active. Looking across them allows the researchers to distinguish a temporary staffing change from a broader change in the venture’s operating trajectory. It also creates a demanding test: a story about lost organizational capital should have consequences beyond the payroll count, while a story about a single vacancy should be more limited.
The sample contains about 26,000 treated and control firms, although the paper rounds observation counts to protect confidential information. In the shock year, the groups have similar measured firm age, founding-team age, employment, revenue, and productivity. The matching is not a guarantee that every relevant characteristic is identical. It is a way to make the comparison more disciplined and to test whether the most important observable preconditions line up.
What the pre-trend check supports
The authors report no evidence of differential pre-trends for the main outcomes after the design’s controls. That result supports the idea that the treated and control firms were following comparable paths before the death. It matters because difference-in-differences relies on a parallel-trends assumption: in the absence of the treatment, the treated firms would have changed roughly like the control firms.
The pre-trend result does not prove that the assumption holds forever or that unobserved differences are absent. It shows that the researchers cannot detect a divergence in the observed pre-event window. A hidden difference could still emerge after the event, and a death may affect the firm through channels that are not the same as ordinary turnover. Evidence of similar prior trends raises confidence in the comparison; it does not turn a working-paper estimate into a randomized trial.
It is helpful to separate three levels of confidence. The data support that the groups looked similar on the measured characteristics at the shock and that their observed outcome paths did not differ detectably beforehand. The design supports interpreting a later difference as associated with the death under the parallel-trends assumption. It does not establish that every unmeasured characteristic was balanced, that the death had no emotional or financial side effects, or that the same effect would appear after a carefully managed replacement. A good causal reading preserves all three levels instead of treating “quasi-experimental” as a synonym for certainty.
The design also has population limits. It covers non-farm employer startups that can be linked across the relevant administrative records, not every new business or every partnership. Revenue is available for only about 80% of the Longitudinal Business Database frame used in the revenue analysis. The treatment is limited to deaths at or before age 60, during the first five years, among members who were active immediately before the event. A result for that group may be especially relevant to formation-stage vulnerability, but it should not automatically be described as the average effect of team turnover throughout a company’s life.
There is a second interpretive limit: the shock may combine several mechanisms. Losing the member’s skills and relationships is one possibility. The suddenness of the event may also create grief or coordination costs. The authors address this concern by pointing to the persistence of the effect and to larger effects for key and higher-human-capital members, but they do not claim to observe every mediating process separately. The design establishes a durable consequence of forced separation in the studied firms; it does not identify one pure variable called “organizational capital.”
That distinction is useful for practical reasoning. If a planned departure includes documentation, overlap, relationship transfer, and a search for missing capability, it is not the same treatment as an unexpected death. The evidence may still make a manager take continuity seriously, especially when a person carries concentrated context. But it does not supply a numerical forecast for a planned change. New evidence on voluntary departures with and without handoffs, or on older firms, could change how broadly the result should be applied.
The design thus earns a stronger conclusion than the descriptive correlations, but a carefully bounded one: among young employer firms, an abrupt loss of an active founding-team member is followed by a worse trajectory than in comparable firms without that shock, under the study’s identification assumptions. The next question is what that worse trajectory looks like—whether the gap is a short vacancy effect, an increase in exits, weaker performance among survivors, or all three.