Olivia Ding

Yale University · Economics · 2026–27 job market

Fields

Econometrics, Industrial Organization, Labor

Job market paper

Partial Identification and Inference for Peer Effects

Peer effects are notoriously difficult to identify. In the workhorse linear-in-means model, the central difficulty is separating the endogenous effect (how peers' average outcome affects one's own outcome) from the contextual effects of peers' average characteristics and the correlated effects arising from a shared unobserved environment —an identification failure known as the reflection problem (Manski, 1993). Existing methods restore point identification with social network data, group size variation, second moment restrictions on errors, or exclusion of some contextual effects, which are often unavailable or hard to justify. I propose a new partial identification approach that requires none of these. It provides sharp bounds on the endogenous effect by mapping the identified reduced form coefficients back to the structural parameters under economically motivated sign restrictions on the direct and contextual effects of selected characteristics. For example, when studying peer effects in classrooms, it is reasonable to assume that classmates' average parental education levels do not lower a student's own academic achievement. Inference proceeds in two steps. The first step estimates the reduced form coefficients by a within estimator that removes fixed effects at a level coarser than the peer group. The group size may be fixed or grow with the number of groups; in the latter case, averaging peers' characteristics over the group makes the regressors asymptotically collinear, leading to nonstandard asymptotics. I establish uniform asymptotic normality for the least squares estimators whose components converge at different rates and propose a consistent cluster-robust covariance estimator. The second step plugs the reduced form estimates into the sign restrictions and adaptively selects the potentially binding sign restrictions with thresholds calibrated to the different rates. The resulting confidence sets are shown to be uniformly asymptotically valid with good power properties. Applied to peer effects on academic achievement in Swedish classrooms (Diemer, 2022), the method yields informative bounds on the endogenous effect under reasonable sign restrictions.