Optimal Cohort Partitions
Research Seminars: Virtual Market Design SeminarThe paper presented in this Virtual Market Design Seminar studies optimal partitions of a heterogeneous population into distinct groups. Through the lens of majorization theory, the authors provide an optimization framework for part-additive objectives that characterizes the fractional and integral partition polytopes, and identifies new classes of optimal policies. The main result shows that, under edge-quasi-convex objectives, there exists an optimal fractional partition that is locally uniform, mixing adjacent types evenly across consecutive sections. Among integral partitions, there exists an optimal pairwise extremal partition, where every pair of sections, holding the others fixed, is maximally balanced or maximally polarized. Under Schur-convex objectives the assortative partition is optimal, while under Schur-concave objectives the uniform fractional and balanced integral partitions are. The authors next apply their framework to student assignment with peer effects. Using the workhorse linear-in-means (LIM) model, they characterize the optimal partitions for performance, inequality, welfare, and diversity, which generally conflict. The assortative partition maximizes welfare but produces the greatest inequality and the least diversity. The inequality-minimizing partition changes from balanced to dispersion-maximal as peer effects weaken. The performance-maximizing partition moves from assortative to upper-uniform to uniform as the planner's weights shift from higher-toward lower-performing sections. The paper further characterizes sorting-neutral models that, despite being non-linear, exhibit LIM's invariance of total effort. Finally, the authors complement their theory with an empirical exercise on a novel dataset with quasi-random classroom assignments. They find that greater sorting is associated with sizable gains for lower-achieving students in the most sorted cohorts. Higher achievers exhibit non-significant and even negative effects.
People