The Founder Is Part of the Treatment
An accelerator does not deliver the same treatment to every founder, even when the program schedule is identical. The applicant's education, prior experience, team structure, and starting resources affect what they can recognize, absorb, and implement. This chapter uses application-based evidence to examine that interaction and to show why founder-program fit is different from a simple ranking of “strong” and “weak” founders.
The distinction is useful for the value-versus-selection question. If founders who enter a program already differ in the capabilities that determine performance, graduate outcomes can reflect those differences. But if the same program produces different results depending on what a founder already knows, that is also evidence about how value is created. The founder is not merely a background variable to control away; they are part of the pathway from program input to venture outcome.
An applicant pool is a better comparison, not a random experiment
The study examined data from the Entrepreneurship Database Program, which collects information from accelerator applicants rather than only from accepted ventures. The dataset includes successful and unsuccessful applicants, ventures from 102 countries, and a follow-up survey. Most of the programs represented in the data last three to six months. This matters because an unsuccessful applicant is often a more plausible comparison for an accepted applicant than an unrelated startup: both chose to apply and were visible to the same kind of selection process.
The comparison is still imperfect. The authors do not know why an applicant was rejected. A program may have selected one venture because of founder quality, market timing, geography, product readiness, or a fit that the dataset does not record. Successful applicants may also differ in motivation or persistence before the program begins. Comparing applicants narrows the gap in observable circumstances, but it cannot guarantee that the accepted and rejected groups are equivalent in every feature that affects later performance.
The study treats program completion or participation as a quasi-experiment and uses inverse-probability-weighted regression adjustment. The method can be understood as two linked models. One estimates the probability that a venture receives the observed participation status from its measured characteristics. The other estimates the outcome associated with those characteristics and participation. The weighting step gives more influence to observations that are less expected under the treatment pattern, helping construct a comparison on the variables the researchers can see.
That is more informative than simply comparing accelerator graduates with all startups. It is not random assignment. The method cannot balance an unmeasured founder trait, a hidden admission rule, or a local opportunity that was never recorded. Nor can it identify the contribution of a particular workshop or mentor because the programs' service intensity, mission, and structure are not fully observed.
The applicant comparison changes the question from “are graduates different from startups in general?” to “do applicants with similar measured characteristics show different outcomes depending on participation?” That is a meaningful improvement in the counterfactual. It also clarifies what remains unknown. If a program rejects a venture because its market is too early, its founders lack a particular credential, or its location falls outside the program's network, the reason may affect later performance and may not appear in the adjustment model. The comparison can reduce visible selection without removing the selection process itself.