Collinearity was not problematic (VIF: eigenvector = 3.27; betweenness = 1.13; degree = 3.45; clust = 1.29). The deviation from normality of the residuals using Shapiro–Wilk was W = 0.950, p = 0.0004 with substantive heteroscedasticity (Breusch–Pagan BP = 71.04, df = 4, p < 10^–9). In this sense, classical inference may underestimate uncertainty, so we consider robust standard errors for heteroscedasticity (HC3) and robust t contrasts for all coefficients. Given that proximity ∈ [0, 1], the specification was reviewed using a fractional regression, and the signs and relative magnitude of the effects are consistent. In addition, we inspected residuals, leverage, Cook’s distances (4/n) and partial linearity to rule out the disproportionate influence of individual cases. It is also worth noting that proximity and intermediation metrics were calculated with distance weights (transforming similarities w_ij to d_ij = 1/w_ij) to facilitate the interpretation of minimum paths when using unweighted metrics. Standard errors and t tests are HC3-robust (heteroscedasticity) (Jochmans, 2022). Predictors with superscript ^† are standardized; degree is in number of links. The 95% CIs were calculated with t-quantiles (df = 105). Point estimators and R² are OLS; Significance codes: ***p < 0.001, **p < 0.01, *p < 0.05