What Counts as a Robot Adoption?
This chapter examines how the Census study turns a difficult-to-observe investment decision into a measurable event: a U.S. manufacturing firm importing an industrial robot in a given year. That definition supports the wage-policy question because it creates a firm-level outcome that can be lined up with a firm's state and year. It also limits the conclusion. To interpret the study correctly, we need to understand what the measure captures, why the sample is narrowed, and which forms of automation remain outside the observation.
From an invisible investment to a firm-year event
A robot installed on a factory floor is not automatically visible in a standard business dataset. A firm may buy equipment through a supplier, record it as part of a broader capital project, or operate a machine for years after the purchase. The authors therefore use an observable transaction as a proxy: they identify firms that import industrial robots in a particular year and classify those firm-years as adoption events.
The first layer is product classification. Industrial robots and their parts appear in international trade records under product codes, but the relevant codes and descriptions can change over time. The authors review ten-digit Harmonized System codes in the tariff schedules from 1992 through 2021 and track the definitions needed to identify robot products across those changes. The point is not to memorize the codes. It is to maintain a consistent bridge between a physical technology and the trade data that records its movement across the border.
The second layer is the firm transaction record. The study uses the Longitudinal Firm Trade Transactions Database to connect an imported product to a U.S. firm. That database supplies the transaction context: which firm received the shipment, when it occurred, and what kind of product was imported. The event is more informative than a national robot total because it can be associated with a particular firm and year.
The third layer is the business record. The Longitudinal Business Database supplies characteristics such as establishment age, employment, industry, location, and ownership structure. Linking the two sources lets the authors compare firms that import robots with firms that do not, and relate the event to the minimum wage in the state where the firm is located. It also provides controls for characteristics that could otherwise make adopting and non-adopting firms look different for reasons unrelated to wage policy.
The result is a deliberately simple outcome: did this firm import an industrial robot in this year? That is a useful definition because it is concrete and repeatable. It is also narrower than “the firm automated.” A firm could automate through software, redesign a process, purchase a domestically produced robot, or obtain equipment through an intermediary without creating the particular import event the study observes.
Each arrow answers a different question. Product classification identifies what kind of item moved. The trade record identifies an importing firm and timing. The business record locates and describes that firm. The sample restriction makes local policy exposure easier to assign. The final event is the outcome used in the analysis. The dashed paths are not defects that make the measure useless; they are reminders that a proxy should not be stretched beyond what it records.
Why the sample is deliberately narrow
The main plant-level sample focuses on single-unit manufacturing firms—firms with one establishment. That choice reduces a problem created by multi-plant ownership. If a firm owns factories in several states, the state minimum wage at any one location may not describe the firm's overall labor cost or the place where a robot is installed. A single-unit firm gives the researchers a clearer geographic link between the firm, the establishment, and the local policy environment.
The same choice narrows the population. The results do not automatically describe large multi-plant manufacturers, firms in services, or businesses that use robots without importing them directly. A sample restriction can improve the clarity of one comparison while reducing generalizability. Both facts must travel with the estimate.
The authors also exclude wholesale, finance, insurance, real estate, and specified professional-service firms. The concern is that such firms may appear in trade data as robot integrators, resellers, or intermediaries rather than as end users deploying the equipment in their own production. A transaction associated with an intermediary could be a real robot shipment but a poor indicator of that firm's own adoption decision. Restricting attention to manufacturing and removing likely integrators makes the outcome more closely match the economic question about a production firm changing its process.
These exclusions are not merely data-cleaning preferences. They define the unit whose incentives the study is trying to interpret. The mechanism in the previous chapter concerned a firm comparing labor-intensive and robot-enabled production. A firm that imports equipment for another customer does not face that same comparison. At the same time, exclusion cannot guarantee perfect classification. Some genuine users may be omitted, and some firms may still have complex supply or ownership arrangements.
Location matters for the policy link as well. The study merges minimum-wage data from the Department of Labor with the business sample using the firm's location. That makes it possible to ask whether a firm's adoption probability changes when the wage floor in its state changes. The geographic assignment is more credible for a single-unit firm than for a diversified enterprise, but it still does not reveal the firm's full exposure to labor costs. A manufacturer may sell nationally, draw workers from across a local area, or face other state policies at the same time.
What the proxy can see—and what it cannot
The import measure has one important strength: it creates a time pattern that can be compared with another measure of robot activity. The authors aggregate the value of U.S. imports of the identified robot products and compare the resulting national trend with International Federation of Robotics installation data. The levels do not coincide, but the movements over time track one another, including downturns around the dot-com bust and the financial crisis and sustained growth afterward.
That comparison is a validation check, not a certification of complete observation. It suggests that robot imports capture meaningful variation in the timing of adoption at the national level. It does not prove that every importing firm installed the machine in the same year, that every installation was imported directly, or that the proxy measures the number of robots rather than the occurrence of a transaction. Similar trends can support the usefulness of a measure without making its blind spots disappear.
The authors identify two especially important sources of undercounting. First, robots produced domestically in the United States do not appear as imports. Second, a firm may obtain a robot through a third party that acquired or resold it, so the end user would not be recorded as the importer. The paper therefore treats the import-based measure as likely to miss some adopters. This matters for interpretation: a low observed adoption rate may reflect both genuinely rare adoption, particularly in earlier years, and incomplete capture.
The event also says little about what happens inside the plant. It does not identify the exact task performed by the robot, the number of workers affected, whether the machine complements or substitutes for labor, or whether the investment remains productive. It cannot by itself establish a productivity gain. Those questions require different outcomes and often different data. The import record is the starting point for studying the policy relationship, not a compressed measure of all consequences.
There is a further distinction between the extensive and intensive margins of adoption. The study's indicator answers whether a firm crosses the line into an observed import event. It does not tell us whether the firm purchased one robot or several, whether the equipment was installed immediately, or how much of the production process changed. A wage policy could affect the decision to begin automating, the scale of a project already underway, or both, but this particular outcome is most directly suited to the first question. Treating it as a measure of automation intensity would add information the transaction record does not contain.
The timing also requires care. An import year is a useful anchor, but the underlying decision may have been made earlier, and the economic consequences may appear later. The authors use a lagged minimum-wage measure in their firm-level specifications because investment responses need not occur at the same moment as a policy change. That timing choice fits the mechanism, yet it does not reveal the date when managers first considered the project or the date when workers experienced a task change. The event marks an observable point in a longer process.
Why rarity changes the meaning of the estimate
Robot adoption is rare in the study's long panel. Among approximately 240,000 single-unit manufacturing firms in the full sample, about 2,600 ever import a robot. The mean adoption indicator is 0.00341. In the border sample, there are roughly 87,000 firms and about 900 robot-importing firms, with a mean adoption indicator of 0.00304. The figures reflect a period that begins in 1992, when industrial robots were much less common than they became later, as well as the undercounting problem.
Those numbers change how the headline estimate should be read. When the authors say that a 10 percent minimum-wage increase is associated with an approximately 8 percent increase in adoption relative to the sample mean, they are describing a relative change in a rare event. They are not saying that 8 percent of firms adopt robots, or that 8 percent of workers leave their jobs. A modest change in the probability of a rare event can be economically meaningful for the affected margin while still leaving most firms without an adoption event.
Rarity also makes selection important. Firms that adopt may differ from firms that do not in size, industry, location, demand, financing, engineering capability, or prior technology. The study controls for several observable characteristics and uses comparison strategies to narrow those differences, but the outcome definition itself cannot tell us whether an unobserved capability helped a firm respond to the wage change. The paper's background reports that robot adopters tend to be larger and concentrated in geographic “robot hubs”; those patterns reinforce the need to avoid treating the average as a universal firm-level rule.
A careful reader should therefore carry four statements at once. The import event is a useful, operational measure of one form of industrial-robot adoption. The linked business data make it possible to relate that event to firm characteristics and local policy. The validation comparison supports the measure's broad timing signal. And the measure misses domestic, indirect, and non-robot forms of automation while leaving the task and worker response unknown.
That combination is not a contradiction. Good empirical work often depends on a measure that is narrower than the concept it helps us study. The question is whether the boundary is explicit and whether the analysis respects it. Here, the mechanism suggests why a wage floor could influence adoption, while the proxy defines which adoption events can be tested. The next step is to assemble those pieces into a testable claim: what is directly measured, what is inferred, and why the study needs both broad patterns and tighter comparisons before the estimate can be interpreted.