When Growth Begins: Timing, Selection, and the Causal Challenge
The subject in this chapter is the timing of advanced-technology adoption and venture-capital funding relative to firm growth. The question is whether employment and revenue begin to rise after those events, rather than merely being higher among firms that eventually adopt technology or receive VC. The answer will not settle every causal dispute, but it lets readers judge how much stronger a timed, matched comparison is than a static ranking—and why a larger VC effect among technology adopters is the key clue about reinforcement.
A cross-sectional comparison has an obvious weakness. Suppose a technology-using, VC-backed firm employs more people than a conventional, bank-financed firm. That difference could reflect the technology, the financing, the founder, the business idea, or a growth trajectory that began before either observed choice. Looking at the calendar does not solve the problem automatically, but it changes the question from “Which firms are bigger?” to “What changes around the point at which a firm adopts or receives funding?”
Why timing changes the interpretation
The Census paper's adoption-timing evidence comes from the 2023 Annual Business Survey, which asks firms when they first adopted AI or robotics in broad five-year intervals. Nearly half of the firms in the process of adopting one of those technologies are between zero and four years old, and nearly 70% are less than ten years old. The concentration among young firms is important because it puts the technology decision close to the formation of the business itself. A startup may be designed around a technology even if the recorded adoption occurs later, after the firm has found customers or obtained the equipment needed to use it.
That observation cuts in two directions. It supports the idea that adoption can be part of the original business model rather than a response made by an established firm. But it also strengthens the selection concern. If an idea is technology-compatible from the beginning, the same underlying quality may make it more likely to grow, attract VC, and report adoption early. The observed event then arrives after some of the relevant sorting has already taken place.
The paper's timing regression follows firms across their own histories while absorbing firm effects, year-by-industry effects, and firm age. It distinguishes the interval during which AI adoption occurs from the years afterward. In the employment specification, the adoption coefficient is 0.136 during the adoption interval and 0.168 afterward. In the revenue specification, the corresponding coefficients are 0.174 and 0.260. The interpretation is that firm size is higher during and after the recorded adoption interval than in the pre-adoption periods under this specification, with a further increase after adoption.
The same table reports a larger pattern around VC. Employment rises by 0.266 in the funding year and by 0.844 after the funding year; for revenue, the estimate during funding is 0.118 and the post-funding estimate is 1.103. The paper describes the VC magnitudes as larger than the AI-adoption magnitudes in this specification. That comparison is suggestive, not a direct rate of return: the coefficients are changes in log outcomes relative to the omitted periods and are estimated under the particular fixed-effects design.
The sequence matters more than the visual drama of any one coefficient. A firm that is already expanding before adoption cannot be treated as though its later size was created entirely by adoption. Conversely, if employment and revenue remain similar to a comparison path before the event and separate afterward, the timing is more compatible with a treatment effect. The adoption intervals make that test imperfect, but they provide information that a single technology-status indicator cannot.
There is an important measurement boundary here. A five-year window does not reveal the exact date on which a firm changed its production process, hired a specialist, bought equipment, or began selling a technology-enabled product. The paper treats the first year of the interval as the adoption year in its difference-in-differences analysis, which the authors describe as conservative because some firms counted as treated may adopt later within the interval. The estimate is therefore about a constructed event window, not a precise managerial decision.
What a matched comparison changes
Difference-in-differences compares the change in an outcome for firms that experience an event with the change for firms that do not, over the same relative period. Its value comes from subtracting a before-to-after change rather than comparing final levels. If a treated firm and a suitable control both grow at the same rate before funding, but the treated firm grows faster afterward, the difference in changes is evidence consistent with an effect of funding. The method still relies on an assumption: without treatment, the two paths would have continued in parallel.
The paper constructs matched controls to make that assumption more credible. For each treated firm, potential controls share firm age, industry, and an employment-size bin at the time of treatment. VC-treated firms are additionally matched with non-VC firms that share technology status; technology-treated firms are matched with non-tech firms that share VC status. Among those candidates, the analysis uses treatment propensity and pre-treatment log employment to select a close comparison. This is not cosmetic matching. It prevents a VC-backed high-tech firm from being compared indiscriminately with a much older, smaller, or technologically different business.
The analysis also uses more than one estimator. Standard two-way fixed effects are familiar, but treatment timing varies across firms, and effects may change as time passes. The paper therefore reports the De Chaisemartin–d'Haultfoeuille method and a pairwise calculation that averages the difference between each treated firm and its matched control. The methods differ in how they handle staggered timing and heterogeneous effects. Agreement across them does not prove the identifying assumptions, but it reduces the chance that the qualitative result is an artifact of one aggregation rule.
This matters because a single average can conceal the path by which an effect appears. In a dynamic comparison, the years before the event are used to ask whether treated and control firms were already moving differently; the event period asks what changes as adoption or funding is recorded; and later relative years show whether the gap persists, grows, or fades. Firms do not all receive funding in the same calendar year, so aligning observations by time relative to each firm's event is necessary for a meaningful comparison. It is also why the analysis cannot simply compare employment in one national year before and one year after a market-wide trend.
Matching and fixed effects remove particular sources of difference; they do not erase the economic decision that put one firm in the treated group. A firm can have the same observed age, industry, size bin, and pre-treatment path as its control while possessing a private signal about demand or technology fit. If that signal affects both the decision to seek VC and future hiring, the post-event divergence will still combine treatment with selection. The design is most persuasive when several pieces line up: the groups have similar pre-trends, the divergence begins near the event, different estimators agree, and the result is concentrated where the proposed mechanism predicts it should be.
Across the three approaches, VC treatment is associated with an average employment increase of 0.355 to 0.403 log points over post-treatment periods one through ten. The paper translates that range into approximately 42% to 49% relative to the matched control path. Advanced-technology treatment is associated with a smaller average increase of 0.038 to 0.065 log points, or roughly 4% to 7%. The difference in the VC effect between high-tech and non-high-tech firms is 0.123 to 0.183 log points, approximately 13% to 20% in the paper's translation.
These estimates answer different questions. The average VC effect asks how the employment path of VC-treated firms changes relative to their matched controls. The high-tech treatment effect asks what happens around technology adoption in the larger sample for which timing is available. The heterogeneity estimate asks whether VC's association with employment is larger when the firm uses advanced technology. It is the last comparison that bears most directly on the interaction in the title. A larger VC effect for high-tech firms is difficult to summarize as two independent advantages simply added together.
To see why, imagine a purely additive story. Under that story, VC would raise employment by roughly the same amount for a technology user and a non-user, while technology would raise employment by the same amount whether or not the firm had VC. The two advantages could still be large, but their effects would not systematically depend on the other choice. The reported heterogeneity is different: the VC treatment effect is larger for high-tech firms, and the difference is visible across the methods rather than in only one estimate. That pattern is consistent with a complementarity in which the financing relationship is more productive when the firm's technology creates an opportunity to scale.
But “consistent with” is doing essential work. High-tech status can still be a proxy for a kind of project that is unusually well suited to outside expertise, rapid expansion, or measurement in the available data. Matching on technology status helps compare VC-backed and non-VC firms within that broad category, yet it does not measure how suitable each firm's particular technology is or how much investor involvement it received. The heterogeneous estimate is therefore a sharper empirical clue than the four-cell ranking, not a direct observation of a VC–technology production function.
The dynamic estimates reinforce that interpretation without closing the case. At the tenth post-treatment year, the reported VC effect ranges from 0.548 to 0.712 log points, and the difference between high-tech and non-high-tech VC effects ranges from 0.285 to 0.495 log points across methods. The high-tech treatment effect at that horizon ranges from 0.067 to 0.135. The paper's figures show mostly statistically insignificant pre-treatment differences after matching, a VC effect that grows over time, and a high-tech effect that peaks around year seven. A pattern that accumulates over years is consistent with financing helping a firm scale, but it can also reflect differences that matching did not capture and that become visible only as the firm matures.
That is why the paper calls the estimates suggestive of causal effects. The design depends on parallel trends: absent VC or technology adoption, treated and control firms would have followed comparable paths. It also depends on no anticipation: firms should not materially change employment in response to a treatment they already expect before the recorded event. The pre-treatment coefficients and matched histories provide checks, but they cannot observe the counterfactual path for the treated firm. A firm may have changed hiring, product development, or financing negotiations before the database records the event.
The data also impose asymmetric limits. VC treatment is estimated on a relatively small treated population because few firms receive VC, while the technology-treatment sample is larger. Technology timing is available only in the 2023 survey and is coarse. The comparison is about employment, not a complete measure of productivity, profits, wages, innovation quality, or investor value added. And even a successful treatment estimate would not identify whether the effect came from cash, advice, hiring, network access, a signal to later investors, or better selection by the VC itself.
A stronger design would need more than additional controls. Evidence on why a particular firm entered VC, what the investor actually did, and which projects were considered but rejected could separate investor value added from investor selection. More precise technology adoption dates and measures of technological suitability could clarify whether the business was designed around advanced production before the event. An external financing shock or another credible source of quasi-random variation could further test the counterfactual. The absence of those measurements does not negate the current evidence; it marks the boundary between a more credible comparison and a fully identified mechanism.
The timed results leave a focused empirical puzzle. Employment and revenue rise during or after technology adoption, firm size rises especially after VC funding, and the VC relationship is stronger among high-tech firms. Yet the firms were not randomly assigned either choice, and the data do not reveal the channel. The next step is to make both possibilities explicit: a model in which high-potential projects sort into VC and advanced technology, while VC can still add value when a project's technology makes that effort especially productive.