What These Studies Prove—and What They Leave Open
The two studies support a useful account of early professional opportunity, but they do not support every conclusion that account might invite. This chapter places their designs side by side and turns their boundaries into practical rules for reading claims: ask what was assigned, what was measured, who entered the sample, and whether the outcome is a later tie, a first destination, or something else entirely. That audit keeps a credible finding about access from becoming an unsupported promise about earnings or success.
Two designs, two levels of claim
The workplace study is built to identify a causal effect on tie formation. New hires are randomly assigned to initial project teams within offices, and the network is observed again later. The researchers then ask whether features of that starting network—especially the presence of indirect ties—change the probability that a later project tie forms. The randomization is not a random assignment of a career outcome. It is a source of exogenous variation in the initial position from which later relationships develop.
The timing makes the causal interpretation possible. If people with indirect ties later collaborate more often, the result is less likely to be explained solely by their initial personal characteristics because the initial project position was assigned rather than chosen. Repeated network observations also avoid treating a finished network as if it revealed its own history. The method can compare later outcomes after different feasible starting arrangements, while conditioning on the realized employees and the potential-outcome relationship represented by the network’s evolution.
That claim still relies on conditions. The analysis needs enough variation in the starting arrangement for the relevant network feature to take different values. It also needs an exclusion restriction: for the focal pair, the selected feature must be a defensible route through which the initial network affects the later tie. If an indirect-tie indicator is merely standing in for every unmeasured aspect of the starting team, its coefficient would not have the intended interpretation. The design is strong because it makes those assumptions visible; it is not strong because randomization makes all network questions causal automatically.
The paper’s inferential machinery is designed around that same boundary. When the relevant network feature has sufficient support across possible assignments, inverse-probability weighting can estimate an average effect over the defined set of employee pairs. When richer features do not have full support, within-office comparisons identify a more local effect over the assignments that can actually occur. The distinction is easy to lose in a headline, but it matters: an estimate based on feasible reshuffling within these offices is not automatically an effect for every organization or every possible network.
The study also uses the assignment permutations to test a sharp null—the claim that the initial network feature has no effect on the later tie—and derives conservative uncertainty bounds from the office-level structure. These procedures avoid pretending that the observed employees were randomly sampled from a universal population of networks. They make the analysis more exact for the design that was actually run, while narrowing the population and treatment interpretation to that design.