The Paradox of More Connections
For the Indian high-growth startups under study, the striking result is not that women founders lack a network. It is that women founders show more potential connections in the broad measures while showing fewer measured relationships involving boards and angels. The useful judgment is therefore not whether one estimate is “good” or “bad,” but whether the two estimates describe different stages of network conversion.
More potential centrality, not fewer contacts
The broad network measures look for educational and professional matches among the people and firms visible in the data. They ask whether a startup appears connected to a wider set of people through shared background, and they count the number of such matches. After controlling for firm age, tier-1-city location, state, and sector, the woman-founder coefficient is positive for every one of these outcomes.
For the indicator of having any connection, the coefficient is 0.0408. In a linear probability model, that is a conditional association of about 4.08 percentage points. The coefficient is 0.0439 for having an educational connection and 0.0214 for having a professional connection—roughly 4.39 and 2.14 percentage points on their respective indicators. The count outcomes are also positive: 18.09 for the number of any connections and 18.03 for the number of educational connections. The professional-connection count coefficient is 0.0692. All of these estimates are reported on the study’s constructed measures, not on a survey question asking founders how many people they know.
The sample for these regressions contains 54,585 firms. The outcome means help put the estimates in context. The average indicator is 0.635 for any connection and 0.622 for an educational connection, while the professional-connection indicator averages 0.0724. The corresponding count means are 104.5, 104.3, and 0.181. Those different scales are one reason not to place the coefficients in a single ranking. Eighteen additional units in a count measure and four percentage points in an indicator are not interchangeable descriptions of the same thing.
Still, the direction is consistent. Once the observable controls and state and sector differences are accounted for, the broad potential network measure is not smaller for women founders. That result challenges the simplest explanation for a capital gap: that women begin with fewer contacts. It also changes the burden of explanation. If the potential network is at least as large in the observed data, then a difference later in the funding process must involve what those contacts represent, how they are recognized, or whether they become consequential relationships.
The controls matter because startups do not enter the comparison as identical units. A firm’s age can give it more time to accumulate recorded links. A tier-1-city location can put it closer to investors and institutions. State and sector effects account for persistent differences in place and industry. After those adjustments, the positive woman-founder coefficients describe a remaining association within the study’s model. They do not mean that gender itself creates eighteen useful contacts, nor that every woman-founded firm has more network resources than every other firm. Regression coefficients compare patterns across groups; they do not turn a constructed variable into a direct account of lived relationships.
There is also an important difference between “more potential” and “better positioned.” A broad network can be larger because it contains many weak or indirect matches. It may be smaller but more valuable when it includes a person who has decision authority, relevant expertise, or the willingness to sponsor a founder. The available estimates cannot value the links one by one. They can, however, show that a later shortage of measured role relationships cannot be explained simply by assuming that potential connections were unavailable at the outset.
The word “centrality” can make this result sound more powerful than it is. In network analysis, centrality is a way of describing a position in a pattern of links. It does not mean prestige, influence, or personal closeness by itself. A founder can have many recorded matches because the firm’s people are associated with a large institution or because biographies contain common professional language. Those matches may widen the set of plausible contacts without increasing the chance that any one person will offer advice, join a board, or invest.
This is the difference between an opportunity structure and a usable resource. An opportunity structure tells us that a route may exist. A usable resource depends on relevance, availability, reciprocity, and the decision authority of the person on the other end. The broad estimates support the first claim. They do not establish the second.
Fewer measured relationships around the venture
The relationship-specific results move in the opposite direction. Here the study examines whether a founder is connected to a board member, whether a founder is connected to an angel investor, and whether a board member is connected to an angel investor. With state and sector fixed effects, the woman-founder coefficients are −0.0764, −0.0519, and −0.0733 respectively. Read as percentage-point associations for these indicator outcomes, they correspond to differences of about 7.64 points for founder–board, 5.19 points for founder–angel, and 7.33 points for board–angel relationships.
The outcome means show that these are not relationships observed in every firm. In the controlled specifications, the means are approximately 0.454 for founder–board, 0.333 for founder–angel, and 0.360 for board–angel. The unadjusted woman-founder coefficients are also negative—−0.0892, −0.0627, and −0.0752—so the contrast is not created only by adding fixed effects. But the adjusted results are the more useful comparison when firms differ systematically by state and sector.
The two sets of findings are not mathematically inconsistent. A broad match can be counted without becoming a role-specific link. A founder may share an educational or professional background with many people visible in the ecosystem, while none of those people appears as a board member or angel connected to the venture. Conversely, a role-specific relationship might arise through a route the broad education and experience measures do not capture. The measures overlap conceptually, but they are not nested stages in which one mechanically guarantees the next.
The difference in sample size reinforces that point. The broad centrality regressions cover 54,585 firms, while the controlled relationship models cover 3,707 firms for founder–board links, 2,387 for founder–angel links, and 2,093 for board–angel links. The smaller samples require more complete information about the people and roles attached to a venture. A role link can be missing because the relevant person is not recorded or the biography is incomplete, not only because the relationship is absent. The negative estimates are still informative about the observed subset, but they should not be read as a census of every social tie around every Indian startup.
It is useful to think of the broad measures as a map and the role measures as a map with landmarks. A map may show that a route exists between two regions; a landmark tells us that a particular destination is documented. Neither map tells us whether the road is open today, whether the traveler has permission to use it, or whether the destination is the one needed. In the same way, potential centrality can identify a field of possible connection while a founder–angel or founder–board link identifies a more specific recorded relationship. The conversion question begins exactly where the map stops: who recognizes the route, who controls the next step, and what happens when the founder tries to use it?
The missing middle in the diagram is the important part. The data can reveal that a potential connection is present and can reveal whether a role-specific link is recorded. They do not directly reveal who made the introduction, whether the founder was invited into a meaningful conversation, whether the relationship carried status in both directions, or whether a board or angel role was offered for reasons unrelated to the measured match.
That limitation makes it tempting to fill the gap with a single explanation, usually discrimination or founder choice. Both may be relevant in some settings, but these estimates do not decide between them. At least five possibilities remain open.
First, the broad measure may capture quantity while the role link depends on relevance. A shared institution is more useful when it connects to the startup’s sector, stage, or immediate financing need. Second, recognition may be uneven: the same background can be treated as a strong credential for one founder and as ordinary information for another. Third, role selection is constrained by who has authority and willingness to join a venture. A potential contact may be busy, geographically distant, or unwilling to take on risk. Fourth, founders may strategically avoid or postpone a relationship if previous interactions suggest that the expected cost is high. Fifth, biographies and firm records may be incomplete, so an apparently absent link can be missing from the data rather than absent in reality.
These explanations also imply different remedies, which is why the distinction is more than statistical vocabulary. If relevance is the bottleneck, a larger contact list will not help as much as a relationship tied to the startup’s sector and financing stage. If recognition is the bottleneck, the same credentials may need an intermediary who can translate them into credibility. If role authority is the bottleneck, an event that creates conversations but excludes board or investment decision makers will have limited reach. If measurement is the bottleneck, a program may appear ineffective simply because the database cannot see the ties it created. The current study does not test these remedies, so they belong as hypotheses about what further evidence should examine, not as proven solutions.
A causal claim would require more than the current cross-sectional contrast. We would want to observe when a potential match existed, when a relationship was activated, who initiated it, what decision followed, and how the result changed relative to a credible comparison. Longitudinal data could show whether a role link precedes funding or is created after a firm has already become attractive. Interviews could reveal whether founders experienced a shared institution as access or as an irrelevant label. Investor-level records could help separate a lack of opportunity from a decision not to pursue it. Without those observations, the safest conclusion is narrower but still important: the study finds a gender difference in measured network layers, and the layers do not move together.
These explanations are not equally supported by the World Bank tables. The tables directly support a difference in the measured potential and role-specific outcomes. They do not measure trust, anticipated rejection, pitch quality, household constraints, or the identity of the person who made an introduction. A disciplined interpretation therefore says that the conversion gap is observed in the recorded network structure while its causal mechanism remains open.
What the paradox changes
The result changes how a founder’s network should be evaluated. Counting alumni or former colleagues can be useful for finding possible routes into an ecosystem, but it cannot tell us whether the route reaches a decision maker. A founder can be well connected in the broad sense and still be weakly connected to the roles that confer validation or supply early capital. The relevant diagnostic is not “How many people are in the network?” but “Which people can make which decision, and is the relationship active at the moment that decision is needed?”
It also changes how an investor or ecosystem institution should read a networking intervention. A program that increases the number of introductions may raise potential centrality without changing who receives board invitations, who is treated as credible, or who obtains an angel conversation. That does not prove the program failed. It means that success must be measured at the stage it intends to change. If the objective is financing, an organizer should not stop at attendance or contact counts; the meaningful outcomes would be consequential conversations, role participation, funding status, amount, or timing. The evidence here does not test such a program, but it shows why a contact-count metric can be too far upstream.
The difference between potential and realized relationships is also a warning about denominators. The broad network regressions use 54,585 firms. The relationship-specific outcomes have smaller samples: 3,707 firms for the controlled founder–board result, 2,387 for founder–angel, and 2,093 for board–angel. These are not interchangeable populations. The smaller samples likely reflect the fact that the role-specific variables require more complete information about people attached to a venture. A negative coefficient in that group should not automatically be generalized to every startup in the broader sample.
Nor does the pattern prove that women founders are less capable of building relationships. It says that the relationships visible under these definitions are less common, even while broad potential centrality is higher. That distinction matters ethically as well as statistically. Turning the result into advice that women should simply network harder would ignore the possibility that the bottleneck lies in recognition or gatekeeping. Turning it into a definitive accusation against every investor would ignore the study’s missing interaction data. The evidence supports a sharper question than either response.
That question is whether the two network layers matter differently for capital. If broad potential connections are associated with a higher probability of receiving funding but not with a larger amount, the conversion problem is about entering the funnel and accumulating capital on different margins. If direct role links predict funding amounts for firms generally but not more strongly for women founders, then the value of a role relationship may depend on an additional stage the data do not capture. The next analysis must therefore move from network structure to funding probability, amount, investor type, and timing—without discarding the distinction that made the paradox visible.