Experimental and Quasi-Experimental Analysis of Peer Effects: Two Steps Forward?

Abstract

In the past 10 years, there has been an explosion of well-identified studies that measure peer effects across many settings and for many outcomes. The emphasis on natural experiments and randomization is a highly useful one; in more standard observational studies, the self-selection of people into peer groups can make the measurement of peer effects extremely difficult. In the absence of exogenous variation, knowing that people have similar outcomes as their friends, classmates, and coworkers may tell us little about peer effects. I examine the successes, failures, and findings of experimental analyses of peer effects. I draw three broad conclusions. First, even more than in other areas of social science, the size and nature of peer effects estimated are highly context specific; peer effects in student test scores and grades are prominent in some cases and absent in others. That said, there is a pattern across studies suggesting that social outcomes (e.g., crime, drinking behavior) and career choices show larger peer influences than do test scores. Second, researchers have shown that the linear-in-means model of peer effects is often not a good description of the world, although we do not yet have an agreed-upon model to replace it. Third, despite potential temptation, we have not reached the point at which we can reliably use knowledge of peer effects to implement policies that improve outcomes for students and other human subjects.