Xuandong Chen

How should regulators design the information available to employers to balance efficiency and fairness in hiring? I study a model in which applicants differ along multiple attributes, some potentially protected, such as criminal history, and others less sensitive, such as education. I measure discrimination along each attribute by the difference in its average value between hired and rejected applicants. I show that prohibiting employers from observing a protected attribute can backfire, exacerbating discrimination along both the concealed attribute and others. This finding sheds light on evidence from Agan and Starr (2018), who document a widening racial gap in callbacks following the removal of criminal-history questions. Moving beyond outright bans, I characterize the full set of optimal information policies and show that every such policy induces employers to hire exactly half of all applicants. Under optimal disclosure, having more of an attribute the firm values can lower an otherwise identical applicant’s hiring probability. This reversal can occur without randomization: holding all other attributes fixed, some applicants with lower values and others with higher values of that attribute are hired, while those in an intermediate range are rejected.

Can two agents, Anne and Bob, agree on which experiment weakly reduces their expected disagreement when different decision problems favor different experiments? I extend the information-validates-the-prior (IVP) theorem of Kartik et al. (2021) from Blackwell informativeness to Lehmann accuracy. Under familiar ordering requirements, Lehmann accuracy preserves the separation of posterior beliefs between lower and higher signals that underlies IVP. For each experiment, consider signals below a cutoff, with the cutoffs chosen so that Bob assigns the same probability to observing such signals. Greater accuracy lowers his average posterior mean conditional on these signals and makes them more likely under Anne’s prior. The corresponding comparison considers signals above cutoffs chosen to give the same probability under Anne’s prior. Greater accuracy raises her average posterior mean conditional on these signals and makes them more likely under Bob’s prior. Besides, under each agent’s own prior, the posterior mean averages to the prior mean. As a result, both agents expect a weakly smaller gap between their posterior means, even when the experiments cannot be ranked by their usefulness across all decision problems.

How do beliefs about what an opponent might insist on affect bargaining power when commitment becomes unlikely? I study continuous-time reputational bargaining in which the probability of commitment vanishes while its distribution across demands can change. A more credible demand encourages the opponent to concede, while a larger demand makes waiting more attractive. I show how this trade-off determines the shares each player can secure, accounting for the information conveyed by strategic demand choice. These guarantees bound payoffs and the surplus lost through delay in every equilibrium. For all but countably many relative discount rates, they exhaust the surplus, yielding common limiting expected payoffs and vanishing delay.