Inside PIPE: The Program, the Firms, and the Measurement
For PIPE, the São Paulo program at the center of this series, the grant-size question begins before any patent is counted. It begins with who is allowed to apply, how proposals are judged, what separates a funded firm from an eligible non-recipient, and which records survive long enough to reveal an outcome. Reconstructing that lifecycle shows why the study can compare grant intensity at all—and why its conclusions remain tied to this particular institutional setting.
A common application frame creates the treatment states
PIPE is the Innovative Research in Small Businesses Program run by FAPESP, the São Paulo Research Foundation. It was established in 1997 and began operations in 1999 to support scientific and technological research in innovative firms in São Paulo. The program provides non-repayable R&D grants to firms with up to 499 employees. That combination of a defined population, direct support, and project-based selection gives the grant a clear starting state: an eligible small or medium-sized firm enters a common application process with a proposed research project.
The eligibility boundary is part of the finding, not administrative background to be discarded. PIPE is designed for firms whose research projects are judged within a particular innovation system, and the study follows applicants who entered that system. It does not observe a neutral cross-section of every small business that might have used money for research. A firm that never applied, a proposal that was ineligible before technical assessment, and an eligible proposal that was not funded occupy different states in the evidence. Keeping those states separate is necessary before any later comparison can be interpreted.
The important transition is not simply from “company” to “recipient.” It is from an applicant pool to two possible treatment states. Applications in the same call are assessed under the same broad eligibility rules, information conditions, evaluation criteria, and budget constraints. Technical merit and the feasibility of the proposed research project determine selection, while prior patenting is not part of the evaluation score. A funded applicant therefore enters a state in which the project receives a non-repayable grant of a particular size. An eligible applicant that is not selected remains unfunded but has passed through the same application environment.
That shared starting point persists after the decision. Both kinds of applicants can be located in the same historical call and compared with reference to the same program context. What changes is the grant exposure. The funded firm receives resources for its project; the non-funded applicant supplies information about what happened to a similar applicant without that PIPE grant. The comparison is not perfect—technical merit may not be fully captured by the available records—but it is more informative than comparing recipients with firms that never applied or were never eligible.
Same-call comparison matters because public programs are shaped by their surroundings. Eligibility rules, the wording of a call, the available budget, the information applicants receive, and the standards used by evaluators can all change across time or across programs. Holding applicants within the same call does not make them identical, but it narrows the institutional context in which their paths diverge. The comparison asks a more disciplined question: among firms that sought PIPE support under the same broad conditions, how did outcomes differ when one entered the funded state and another did not?
What is lost at the funding transition is also important. An unfunded applicant does not receive the PIPE resources assigned to the project, so the later comparison is not a comparison of two different grant sizes among all possible firms. It is first a funded-versus-non-funded contrast, built from an applicant population. Only after that frame is understood can variation among funded awards be used to study dose. The study’s conclusions about grant size therefore inherit the selection frame created by the program.
PIPE also gives the funded state a distinctive property: beneficiary firms retain ownership of the intellectual property produced by their projects. The grant is therefore direct support for research rather than a purchase of the resulting patents by the funder. The average duration of benefits in the study’s data is reported as 3.7 years, which reinforces that the relevant object is a project unfolding over time, not a one-day cash transfer whose consequences should appear immediately.
The transition from selection to grant exposure also creates variation in dose. Awards are highly right-skewed. Most lie toward the lower end, while a long tail reaches approximately US$600,000. About 22% fall between US$100,000 and US$200,000, 10% between US$200,000 and US$500,000, and fewer than 1% exceed US$500,000. The paper reports an average expenditure per firm of US$88,143.76 between 1999 and 2019. These are properties of the observed PIPE distribution, not fixed bands that the program promises to every firm.
A right-skewed distribution also makes the word “average” easy to misuse. A small number of relatively large awards can pull the mean upward even when most projects receive less. The mean describes total spending divided by firms over the stated period; it does not describe the typical project’s experience or the amount at which the response is strongest. The long tail is valuable for estimating how outcomes change across support levels, but it also means that comparisons must keep their denominator and time period visible.
The later intensity analysis works with the range that the program actually produced. That is an empirical advantage: there is variation to examine. It is also a boundary: the study cannot use its observed curve to make confident claims about grant levels that PIPE did not meaningfully assign, or about a funding distribution with a different mix of small and large awards. The amount is therefore both an input to the project and a feature of the institutional process that generated the data.
That distinction is crucial. If every funded project received the same amount, the research question would largely be whether the funded state produced a different outcome from the non-funded state. Because grant amounts vary, the funded state contains multiple exposures. A low award may not cover a necessary researcher or piece of equipment; a larger one may support a broader project but also demand more coordination. The grant record therefore preserves more than a yes or no. It preserves the amount that later makes a dose–response analysis possible.
The lifecycle up to this point is institutional. It tells us how an applicant becomes funded or remains a same-call non-recipient, what persists across that transition, and where grant size enters. It does not yet tell us whether a higher dose generated a better result. To answer that, the program lifecycle has to be joined to a measurement lifecycle.
The diagram follows records and treatment states, not a claim that every project passes through identical operational steps. Its purpose is to show the comparison that the data retain: a common application frame, a funding decision, a dose for recipients, linked pre-treatment characteristics, and a delayed innovation observation.
From a grant decision to a patent observation
The study assembles the second lifecycle by linking several administrative sources. The FAPESP PIPE data supply beneficiary firms and grant-level information. These records are joined to employment, industry, location, and tax-classification information from Brazil’s Ministry of Labor, and to patent records from the Brazilian Intellectual Property Office. The linkage turns a program event into a firm-year panel: a firm can be observed before treatment, at the time of the funding decision, and after the project has had time to produce an outcome.
This linkage changes the unit of attention. A grant is recorded at the program level, but the analysis asks what happened to a firm over time. Employment, industry, and location do not become outcomes simply because they are available; they are context variables that help describe the firm and support a fairer comparison. The patent record, in contrast, is used as the principal innovation-output measure. Separating these roles prevents every column in the administrative data from being treated as evidence of the same thing.
What persists through the linkage is not an abstract measure of “success,” but a set of defined variables. Employment helps characterize firm size. Industry and location preserve differences in technological and regional context. The patent record supplies an observable output that can be tracked against earlier patent activity. The final assembled dataset contains 23,736 firm-year observations for 2,293 unique firms, including 1,313 firms that received at least one PIPE grant. After the relevant timing and comparison restrictions, the time-range construction uses 1,313 treated firms and 980 controls. The treated count here refers to firms that received at least one PIPE grant, while the control count refers to eligible non-funded applicants retained for the comparison after the stated restrictions. Firm-year observations and unique firms are consequently different denominators: one firm can contribute more than one year of information, while the treatment and control counts describe firms in the restricted panel.
Those numbers describe successive states of the evidence rather than interchangeable denominators. The raw program population is not the same as the final estimation sample. Proposals deemed ex ante ineligible are excluded. Observations without the necessary pre-treatment covariates are also excluded, because the comparison requires information about firms before funding. Timing restrictions matter too: an observation must have the relevant pre-treatment history and a sufficient post-grant window to be useful. Each restriction sacrifices some raw records in exchange for a more interpretable comparison.
The primary outcome is firm-level fractional patent applications. Fractional counting is intended to reflect the firm’s share when a patent involves multiple applicants, rather than treating every record as a whole patent belonging exclusively to one firm. Patent activity is tracked from 1996 onward. This gives the study a measurable output that can be compared over time, but it also defines what the evidence cannot see. A new process kept as a trade secret, an improved product that is never patented, or a capability that has not yet produced an application may be economically important without appearing in the main outcome. This is why a change in the study’s patent measure should be read as an output signal, not as a complete balance sheet for the firm. It is also why fractional results require care: an estimate such as an average change in patents can summarize partial contributions across firms and applications rather than a promise that each firm produced a whole, commercially valuable patent.
The transition from grant exposure to outcome is deliberately delayed. The study measures the change in patent stock three years after the grant, because R&D spending and observable innovation typically take time to materialize. This window is not a claim that every project needs exactly three years. It is the study’s principal horizon for allowing research work, complementary inputs, and the process of preparing a patent application to unfold. Choosing a window also determines which projects are visible to the evaluator. A project that is still developing at the end of the window is not necessarily a project that failed, just as a patent observed within the window is not by itself proof of commercial success. The three-year measure is a disciplined common checkpoint, not a complete account of the project’s life.
That time horizon changes how one interprets a firm’s state after funding. Immediately after the decision, the data can tell us that a grant was assigned and in what amount. They cannot yet tell us whether the project’s knowledge has become an observable patent. At the three-year transition, the study can compare patent outcomes. The event-study evidence discussed later shows that effects emerge gradually, becoming statistically significant from the second year rather than appearing as an instant jump. A short evaluation window could therefore misclassify a project whose research process is still underway as unsuccessful.
The measurement state is also narrower than the policy objective. Patents are a useful and comparable indicator of technological output, but they are not a complete inventory of innovation. The paper does not measure forward citations or patent-family size, so its data cannot directly establish the quality or commercial importance of each patent. Nor do the records capture all possible spillovers, such as enhanced reputation, collaboration, network effects, or knowledge that moves to other firms. A higher patent count is evidence of additionality for this outcome; it is not automatically evidence of higher total social value. Likewise, no observed patent difference would be evidence about this measured outcome under this horizon, not evidence that the grant created no knowledge or capability. The distinction between what is observed and what is absent from the measure protects the analysis from both overclaiming success and overclaiming failure.
This boundary matters for the grant-size question. Suppose a large award produces no additional patent within three years. Several explanations could remain possible: the project may have been poorly suited to patenting, the result may appear after the chosen horizon, or the grant may have built an unmeasured capability or spillover. Conversely, a patent count can rise without telling us whether the project generated sales or productivity gains. The study’s design does not erase these possibilities; it makes one important outcome comparable while keeping the unobserved remainder outside the estimate.
The same discipline applies to the program’s geographic and institutional scope. PIPE is a place-based program operating in São Paulo, an important Brazilian industrial and innovation center. The firms in the database are applicants to this program, not a random sample of all Brazilian SMEs. Their projects were evaluated under PIPE’s rules, their awards came from PIPE’s funding distribution, and their outcomes were recorded in the surrounding São Paulo innovation system. Those features are what make the case concrete, but they are also what make a PIPE grant amount a poor candidate for an automatic rule elsewhere. A funding level has meaning only relative to the project requirements, the firm population, the local evaluation process, and the distribution of awards available in that setting. Changing any of those conditions can change which applicants receive support, what the amount buys, and how quickly an outcome becomes visible.
The two lifecycles meet at the firm-level outcome. The institutional lifecycle explains why the grant amount differs across funded applicants and why a same-call non-recipient can serve as a comparison. The measurement lifecycle explains how the study sees pre-treatment characteristics and delayed patent activity. Neither lifecycle alone answers the causal question: program rules do not guarantee comparable firms, and linked records do not guarantee that patents capture all value. Together they produce a more transparent object for evaluation, with both a defined treatment and defined limits.
By the end of this lifecycle, the study has preserved the ingredients needed for the next question: who received support, who remained in the same-call comparison group, how much the funded firms received, what their pre-treatment characteristics looked like, and what patent activity followed after three years. The resulting panel is stronger than a simple list of winners and losers, but it is still a structured comparison with assumptions. The next step is to ask whether the difference in patenting can reasonably be interpreted as an effect of funding rather than as a pre-existing difference between applicants.