Activist homophily, activist signaling, and the acquisition of social capital by Black entrepreneurs: a field experiment

2026-04-15

2026-06-29

Connection is coded 0 if the mentor declines the connection request and 1 if the mentor accepts the request.

Potential variables that would be constructed are based on any data we can collect on founders on LinkedIn and other archival data

In follow-up studies we are also sending a message with the connection request. Response to the message from the mentor is a potential secondary outcome to explore. A potential additional outcome would be to use AI and measure the tone or helpfulness of the reply across the two groups.

A randomized experiment on requesting LinkedIn connections from entrepreneurs.

We randomized the Phase II sample using a stratified design implemented in Stata (randtreat command, seed 20260318). The Phase II design extends the binary treatment structure of Phase I to a 2×2 factorial experiment crossing Entrepreneur Race (White vs. Black profile) with Request Type (simple connection vs. advice request), yielding four arms: White + Simple connection (WS, control), Black + Simple connection (BS), White + Advice request (WA), and Black + Advice request (BA).
Randomization was stratified on three variables: gender (female vs. non-female), LinkedIn activity level (High or Moderate vs. Low or Inactive), and top university attendance (binary). These replace the Phase I YC stratification variables (gender × followers × red state × pro-bono signal) for two reasons. First, the pro-bono signal is unavailable in the YC directory scrape as it was a linkedin feature that we did not collect in second phase. Second, state-level location is missing for 31.3% of founders, so including the red-state indicator as a stratification variable would exclude those founders from randomization; it is retained instead as a pre-specified heterogeneity variable. The three-way crossing produces eight strata cells (minimum cell size 109), sufficient for four-arm within-cell assignment. All 5,488 founders were entered into randomization; 12 observations with missing arm assignments from randtreat remainder allocation were excluded, yielding a final randomized sample of 5,476 founders (1,369 per arm).

Phase I comprised multiple studies across three mentor networks (SCORE, Founder Institute, and Y Combinator pro-bono alumni) and one general YC alumni sample, with a two-arm design (White profile vs. Black profile). Phase II constitutes a new study within the same pre-registration and extends the design in two key respects.
5.1 Design Change: 2×2 Factorial Experiment
Phase I used a two-arm design (Black vs. White profile, simple connection request). Phase II extends this to a 2×2 factorial design crossing Entrepreneur Race with Request Type. The motivation is to test whether racial discrimination increases with commitment level, addressing the concern that Phase I’s low-stakes context may have attenuated discrimination, while controlling for the temporal confound introduced by the post-election DEI backlash environment by manipulating both factors within the same 2025 wave. The four arms and sizes are: WS (control, n = 1,369), BS (n = 1,369), WA (n = 1,369), BA (n = 1,369); total N = 5,476.
5.2 Updated Hypotheses
We pre-register four hypotheses structured around the research question: do Black entrepreneurs face discrimination when building social capital in elite entrepreneurial networks, and does this vary with the stakes of the request?
H1 (Main effect of race): YC alumni are less likely to accept a connection request from a Black entrepreneur than from an otherwise identical White entrepreneur, averaged across request types. This tests the baseline presence of discrimination across the full 2×2 design.
H2 (Discrimination in the high-stakes condition): YC alumni are less likely to accept an advice request from a Black entrepreneur than from an otherwise identical White entrepreneur (WA vs. BA). This is the primary test of whether discrimination manifests when the ask imposes real time and commitment costs on the recipient, directly addressing the theoretical claim that discrimination concentrates in higher-stakes interactions.
H3 (Stakes intensify discrimination): The racial gap in acceptance is larger for advice requests than for simple connection requests. Formally, β₃ in
Accept = β₀ + β₁Black + β₂AdviceRequest + β₃(Black × AdviceRequest) + ε
is negative and significant. H3 is designated secondary given its limited statistical power at the expected effect size (see Section 5.5) and is tested one-sided given the directional theoretical prediction.
H4 (Exploratory. Testing cost effects by race): We explore whether the cost of the ask operates symmetrically across race. H4a tests whether White founders are less likely to receive acceptance on an advice request than a simple connection (WS vs. WA);
H4b tests the equivalent for Black founders (BS vs. BA). If H4b shows a larger penalty than H4a, this suggests that higher-commitment asks compound racial barriers rather than merely imposing uniform costs. These comparisons are explicitly exploratory and will be reported as such.
H5 (Exploratory. Temporal comparison, simple connection condition): The racial gap in acceptance in Phase II’s simple connection condition (WS vs. BS) is larger than the equivalent gap observed in Phase I’s YC sample, reflecting an increase in discrimination following the post-2024 collapse in the legitimacy of diversity and equity considerations in organizational decision-making (Bitektine & Haack, 2015).
Our hypothesis here is designated exploratory given that Phase I and Phase II draw from different sampling frames (LinkedIn search results versus the official YC directory) which introduces a potential confound. Accordingly, this comparison will be estimated on the matched subsample of founders confirmed in both phases (n ≈ 1,503, see Appendix B), with controls for observable compositional differences. The test is one-sided given the directional prediction. Formally, the test is whether β₁ (Black profile coefficient) in the Phase II WS/BS subsample exceeds its Phase I equivalent in magnitude, estimated via a pooled model interacting treatment with a Phase II indicator (Treatment × Phase II) and restricting to the simple connection condition and matched sample.
Finally, pre-specified heterogeneity analyses will test whether H1 and H2 vary by: (a) gender of the YC founder; (b) URM status; (c) red-state vs. blue-state location; (d) LinkedIn activity level; (e) top university attendance, mirroring the underdog thesis and regional conservatism hypotheses from Phase I.

Randomization done in office by a computer using stata. We will upload the log file of the randomization code.

The updated log file for phase 2 will be uploaded as well.

We randomize at the founder level.

Yes

Phase I YC stratification: gender × followers × pro-bono signal × red state. Phase II stratification: gender × LinkedIn activity level × top university. Pro-bono signal is unavailable in the 2026 scrape. Red-state is excluded from stratification (missing for 31.3% of sample) and retained as a heterogeneity variable. LinkedIn activity replaces follower count as the activity-based variable and is available for 100% of the sample. Eight cells, minimum N = 109.

Phase I drew its YC sample from LinkedIn search results using self-reported affiliation (n = 2,921). Phase II uses a near-complete census from the official YC directory (raw N = 9,104; after Phase I exclusion and US restriction: 5,488 founders, 3,500 companies; randomized N = 5,476). This is a 1.9× increase over Phase I and eliminates the self-selection bias of the Phase I sampling frame.

Eight cells, minimum N = 109.

Phase I YC baseline. The full Phase I YC sample (n = 2,686 with controls) produced a treatment coefficient of −0.009 (SE = 0.014, p > 0.10) — a null result — against a White profile acceptance rate of 17.5% (regression constant = 0.175***). The corresponding raw gap was 1.1 pp (White 17.4%, Black 16.4%, p = 0.44). The most relevant Phase I signal for Phase II is the pro-bono mentoring interaction: Treatment × Mentoring = −0.120** to −0.122**, indicating a gap of approximately 11–12 pp among founders who signaled selective engagement. This pattern — discrimination concentrating when founders make deliberate, higher-commitment decisions — is the direct Phase I precedent for H2. Phase II power calculations use the manuscript regression baseline of 17.5% for the simple connection condition and assume a 40–50% reduction (to ~10%) for the advice request condition based on the literature on commitment-level effects. Minimum detectable effects. With n = 1,369 per arm (α = 0.05, two-tailed, 80% power), MDEs are as follows. For H1, the pooled race main effect (n = 2,738 per race), MDE = 2.52 pp; power to detect a 2 pp gap is 60% and a 3 pp gap is 92%. For H1 (simple connection only, WS vs. BS, baseline ~17.5%), MDE = 4.1 pp two-sided; power to detect a 3 pp gap is 57% and a 5 pp gap is 96%. For H2, the race gap in the advice request condition (WA vs. BA, assumed baseline ~8.5%), MDE = 3.0 pp two-sided (2.65 pp one-sided); power to detect the expected 3 pp gap is 80%. H2 is thus the best-powered individual test in the design, consistent with the theoretical and Phase I pro-bono evidence that discrimination concentrates in selective, higher-commitment contexts. Power for H3 (DiD/interaction). Under the expected scenario (a 3 pp gap in advice requests and a 1.1 pp gap in simple connections) the expected DiD is 1.9 pp. Two-sided power for this DiD is approximately 19% (28% one-sided). Power reaches 40% at a 3 pp DiD, 63% at 4 pp, and 83% at 5 pp. The study would require a much larger sample to achieve 80% power for the expected 1.9 pp DiD which is not feasible as we used the full YC sample. We address this limitation through three design choices. First, ANCOVA with pre-treatment covariates (followers, activity score, top university, elite employer) reduces residual variance by an estimated R² of 10–18%. This meaningfully improves power for H2 (lowering the advice-request race gap MDE from 3.0 pp to 2.70–2.83 pp and raising H2 power from 80% to 84–87%) but yields only modest gains for H3, raising DiD power from 19% to approximately 22% two-sided. Second, H3 is tested one-sided given the unambiguous directional prediction and Phase I precedent (Treatment × Mentoring = −0.120**, p < 0.01), raising power for the expected 1.9 pp DiD from 19% to 28%. Third, H2 is designated primary and H3 secondary, so the study’s inferential burden falls on H2 where power is adequate. Informative null. If H2 is null, that result is substantially more informative than Phase I's null. Phase I could rule out gaps larger than approximately 3.6 pp in the YC network (a 20% relative effect at the 17.5% baseline). A null on H2 would rule out gaps larger than 3.0 pp in the advice request condition — a 35% relative effect at the ~8.5% advice-request baseline. Ruling out a 35% relative discrimination effect in a high-stakes context would constitute meaningful evidence that discrimination is not operating, directly addressing the AE's concern that Phase I's low-stakes context precluded conclusions about higher-stakes interactions.