What the Evidence Says—and What It Cannot
The central evidence question about startup accelerators is not simply whether graduates look successful. It is what a pooled comparison of accelerator participation and venture performance can legitimately tell us about the program's contribution. This chapter reads the available meta-analysis as an estimate with a method, a denominator, and boundaries: readers should be able to use its positive average without treating it as the guaranteed causal return for any particular founder.
That discipline matters because the previous mechanism can produce several kinds of outcome, and the research record measures them unevenly. A study of fundraising is not automatically a study of durable capability. A comparison of graduates with non-participants is not automatically a comparison of otherwise similar ventures. The meta-analysis helps by putting many quantitative findings on a common scale, but the act of pooling does not make the underlying programs, founders, or measures identical.
What the meta-analysis actually combines
The authors searched research published or produced between 2000 and 2022 and retained 21 quantitative primary studies. Three of those studies examined both accelerators and incubators. Across the included material, the analysis represents about 32,000 ventures that participated in accelerator programs; the mixed accelerator–incubator studies add a joint sample of roughly 1,100 startups. These are large-looking totals, but the number of primary studies is still limited, and multiple outcomes can come from the same study.
That last point is why the headline denominator is 68 effect sizes rather than 68 independent experiments. A primary study may report more than one relevant outcome—such as funding, employment, or innovation—and each relationship can enter the synthesis as a separate effect size. The 68 observations capture more dimensions of performance than a single number would, but they do not turn the evidence into 68 unrelated tests of randomly assigned participation.
The independent variable is simplified to a binary distinction: a venture participated in an accelerator or it did not. The included programs can nevertheless be corporate, university, ecosystem, social, public, private, investor-led, generalist, or specialized. A binary participation variable is useful for combining studies that use different labels, but it compresses those differences. It cannot tell us whether mentoring, funding, a peer cohort, office space, or a demo day produced the observed association. Nor does it record how intensively a venture used a service.
The dependent variable, “new venture performance,” is also a family of measures. The meta-analysis groups outcomes into financial, operational, and strategic categories. Financial measures include fundraising, revenue, profit, value, acquisition, or expenditure. Operational measures include employment, human resources, and firm assets. Strategic measures include digitalization, research and development, patents, copyrights, or trademarks. This taxonomy makes unlike measures easier to compare, but it should not be mistaken for one unified definition of success.
To place different outcomes on a common scale, the authors use Pearson correlation as the effect-size metric, convert reported statistics when necessary, and apply a Fisher transformation before combining them. They weight effect sizes by sample size and use a random-effects model. The choice is important. A fixed-effects model would assume that all studies are estimating one common underlying effect. A random-effects model allows the true relationship to vary across programs, ventures, and contexts, then estimates an average across that distribution. Given the diversity of the evidence, the latter is the more appropriate question: what is the average relationship across a collection of different accelerator settings?
A common scale makes comparison possible but also hides some original detail. A binary participation variable paired with a continuous performance measure can tell us whether participation and performance move together, but not whether the venture used a service, how much it used it, or what happened between admission and outcome. A correlation is symmetric about association: it does not by itself identify which direction the influence runs, whether a third factor affects both participation and performance, or whether an accelerator's signal mediates the relationship. The random-effects model acknowledges variation; it does not solve selection into participation.
The sample-size weighting adds another important nuance. Larger primary samples contribute more precision to the pooled estimate, which is generally desirable when those samples are well designed. But a large sample from one kind of accelerator does not automatically make it representative of a smaller, differently sponsored program. Precision and applicability are different properties. The meta-analysis can be statistically careful and still leave the reader to ask whether the studies most relevant to a decision resemble the studies with the most influence on the average.
The flow from studies to a pooled estimate is therefore a translation, not a purification. It makes the findings comparable enough to summarize while leaving the differences that matter for interpretation. A venture in a well-supported ecosystem and one in an institutionally weak setting may receive programs with similar labels but not similar marginal value. A financial outcome measured shortly after graduation may also capture a different process from strategic or operational change observed years later.
A positive average inside substantial variation
Across the 68 effect sizes, accelerator participation has a positive and statistically significant pooled correlation with venture performance: r = 0.102, with a reported standard error of 0.0231. The result supports a cautious answer to one narrow question: across the quantitative literature assembled by the authors, accelerator participation is associated with better venture outcomes on average.
The magnitude matters. A correlation of 0.102 is not a promise that a participating venture will succeed, and it is not a claim that the program caused an otherwise identical venture to perform 10.2 percent better. The number is on a correlation scale, not a percentage return. Its practical meaning depends on the baseline outcome, the program, the venture, and the alternative path. The authors themselves describe the average effect as modest and compatible with negative or null findings in individual studies.
The most important qualification is the amount of variation beneath the average. The meta-analysis reports a homogeneity statistic of Q = 732.31, between-study variance of tau-squared = 0.0124, and I-squared = 93.89 percent. In plain language, the observed differences among effect sizes are far larger than would be expected from sampling error alone. The estimate is not surrounded by a cloud of nearly identical program effects. It is an average across relationships that can be meaningfully stronger, weaker, or negative.
High heterogeneity changes how the result should be used. It does not make the average meaningless; an average can still summarize the direction of a diverse body of evidence. It does mean that the average is a starting point for asking what varies. If a new accelerator resembles the studies that produced stronger effects—in its venture stage, service mix, cohort design, sponsorship, or ecosystem—the average may understate or overstate its likely relationship. If it resembles a weaker or negative case, citing the overall positive result is poor decision support.
Imagine two programs with the same participation label. One offers a small, intensive cohort, specialized mentors, and access to a thin local investor network. Another runs a large generalist batch in a mature ecosystem where founders can find comparable advice elsewhere. If the first produces a larger relationship and the second a weak one, the pooled result is not contradictory; it is averaging different marginal contributions. The program label remains constant while the resource gap, service fit, and outside alternative change. Heterogeneity is therefore a prompt to investigate the conditions of value, not a nuisance to be averaged away.
The authors conduct a leave-one-out analysis and do not find one single outlier that explains the variation. That is useful, but it does not solve the deeper problem. No single study is solely responsible for the heterogeneity, because the heterogeneity is distributed across program designs, samples, methods, regions, and outcomes. Removing an outlier can protect against one distorted observation; it cannot turn a varied population of programs into one uniform treatment.
This is also why statistical significance should not be allowed to dominate interpretation. Significance addresses whether the pooled relationship is distinguishable from zero under the model. It does not answer whether the effect is large enough for a founder to exchange equity for it, whether it persists after graduation, or whether a particular program is better than the founder's next-best alternative. Those are questions of magnitude, fit, durability, and opportunity cost.
Bias, measurement, and the limits of generalization
A second uncertainty concerns which studies become visible in the literature. The meta-analysis finds that published studies have a larger mean effect, r = 0.123, than unpublished studies, r = 0.057. The difference between the groups is statistically significant. Its funnel-plot analysis and a positive, significant Egger test—beta = 1.18 with p = 0.032—also indicate small-study effects. Smaller studies with large standard errors appear more likely to report unusually positive findings, while small studies with null results may be less visible.
These diagnostics are evidence of possible publication bias and small-study effects, not proof that every positive study is wrong. They do, however, change the strength of the headline claim. The pooled positive estimate may be somewhat higher than the relationship that would appear if all relevant null and negative findings were equally available. The fact that the authors searched for unpublished work is a strength of the design; the difference between published and unpublished estimates shows why that effort matters.
The asymmetry also has more than one possible explanation. Small programs may genuinely differ from large programs, may serve unusually risky ventures, or may report outcomes with greater uncertainty. Selective publication can coexist with those real differences. The available diagnostics point to a reporting problem but do not identify exactly how much of the pooled relationship it explains. This is why the correct response is not to replace one certainty with another—“all positive findings are hype”—but to lower confidence in an unqualified positive narrative and look for evidence from less selective comparisons.
The composition of the outcomes creates another boundary. Forty-nine of the 68 effects are financial, compared with 10 operational and nine strategic effects. The financial subgroup has r = 0.112, the operational subgroup r = 0.092, and the strategic subgroup r = 0.044, which is not statistically significant. This does not establish that accelerators cannot improve strategy or innovation. It shows that the supplied literature measures financial outcomes more often and finds its clearest pooled relationship there.
The imbalance is consequential because financial outcomes are often the easiest to observe soon after a program. A follow-on investment can be recorded even when the venture has not yet built a durable operating capability. Revenue or valuation can rise with market conditions, investor sentiment, or the credibility conferred by affiliation. Conversely, learning, customer understanding, organizational routines, and resilience may matter greatly without appearing in a short-term funding measure. The available evidence is therefore stronger for a relationship with early or medium-term financial performance than for long-run strategic growth or survival.
The distinction between “no significant effect” and “no effect” is essential here. The strategic subgroup's estimate does not support a confident positive claim, but it also cannot prove that strategic benefits are absent. There are only nine strategic effects in the synthesis, the underlying programs differ, and the authors note that many studies focus on short-term outcomes. A non-significant result may reflect a small evidence base, noisy measures, short follow-up, or genuine absence of a broad effect. The responsible conclusion is that strategic value remains less clearly established.
The meta-analysis also inherits the limitations of the primary studies. Participation is usually not randomly assigned. Programs differ in admissions, resources, missions, and local conditions. Some studies use surveys and self-reported performance; others use archival data. Prior control groups may not be equivalent, and program features such as mentorship intensity or specialization are not reported consistently. The pooled estimate can summarize the available evidence without resolving those underlying design differences.
For a decision maker, this creates a simple evidence hierarchy. The pooled average is useful for rejecting the claim that accelerators never help anyone. Moderator patterns are useful for generating fit questions about design, venture stage, and context. A study that compares applicants is more informative about selection than a study that compares graduates with the entire startup population. None of these levels, by itself, identifies the exact value of a particular program for a particular venture. That value still requires a local counterfactual: what resource, capability, or signal would the venture obtain otherwise, at what cost, and with what delay?
The strongest supported conclusion is consequently narrow but useful: accelerators are associated with modestly better venture performance on average in the available quantitative literature, especially where financial outcomes are measured. The evidence is not strong enough to say that every accelerator creates value, that the average is a causal return, or that a program's financial signal implies durable strategic improvement. The positive relationship becomes a decision-relevant claim only after it is conditioned on the design and setting of the program, the venture's starting capabilities, and the outcome being sought.
That is the point at which the average stops being an answer and becomes a clue. The next question is not whether one program feature is universally best, but how cohort scale, duration, services, sponsorship, venture size, and institutional context may change the marginal value of participation.