Academic Peer Effects with Different Group Assignment Policies: Residential Tracking versus Random Assignment∗

A growing literature in the economics of education emphasizes the importance of peer effects in determining students’ academic outcomes. Most of the existing empirical literature uses natural experiments in which students are randomly assigned to peer groups to identify the causal effect of peer characteristics. However, their results provide no direct evidence regarding the policy-relevant problem of which group assignment policies produce better or worse aggregate outcomes. I address this gap in the literature by contrasting the distribution of test scores under two common policies for assigning students to residential peer groups: tracking and random assignment. I find that tracking reduces mean GPA by approximately 0.12 standard deviations (one quarter of the black-white GPA gap) relative to random assignment. This is driven by very large negative effects on the lower tail of the GPA distribution (up to 0.5 standard deviations) and near-zero, insignificant effects on the upper tail. These effects are robust to a variety of strategies to control for selection on observed and unobserved characteristics. I explore whether these effects could have been predicted using random variation in dormitory composition generated under the random assignment policy. I estimate a flexible education production function and show that the parameters do not predict the magnitude of the treatment effect of tracking. This problem may reflect out of sample prediction problems or students’ behavioral responses to changes in peer group composition, both of which are present in this setting. JEL codes: I23, I24, I25, O15 ∗I thank Manuela Angelucci, Raj Arunachalam, Emily Beam, John Bound, Tanya Byker, John DiNardo, Susan Godlonton, Andrew Goodman-Bacon, Italo Gutierrez, Brad Hershbein, Brian Jacob, David Lam, Rebecca Thornton, Jeffrey Smith, Dean Yang; seminar participants at the University of Michigan, University of Cape Town, and UM-MSUUWO Labor Day; and conference participants at CSAE, EconCon, ESSA, MEA, MIEDC, PACDEV, and SOLE for helpful suggestions. Jane Hendry, Charmaine January, and Josiah Mavundla provided invaluable assistance in obtaining the data used in this project. All errors are my own. †University of Michigan; [email protected]