Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

Arianna Gambardella i6209324

Summary: Peer Effects with Random Assignment: Results for Dartmouth


Roommates (Sacerdote, 2001)
The standard approach to evaluating peer effects regresses the outcomes of student i to the
outcomes of its peers (all j ≠ i). Three problems arise from this approach. Firstly, the selection
effect (people self-select themselves inside a group) is hard to separate from the peer effect.
Second, there is no clear effect of student j on student i and vice versa since they influence
each other simultaneously. Lastly, it is hard to separate contextual effects (people’s
background) from endogenous effects (people’s behavior). Sacerdote addresses the problem
by using data of peers that are randomly assigned in university dorms. This way, he aims to
eliminate the selection problem. Table II shows the results of the regression of student i’s
characteristics on peer j’s characteristics (roommate of i). The results indicate that,
conditional on student i’s answers to the questions provided, there is no relationship between
i’s background and roommate j’s background. This holds only for randomly assigned pairs.
The inclusion of blocking variables does not alter this finding, not even when the author
accounts for different functional forms in order to control for them. That might be because
students give noisy answers or because “true” answers are not correlated with their
background. The author then estimates the following regression: GPA_i=a_0+
a_1*ACA_i+a_2*ACA_j+e, where ACA is the academic index, representing whether the
student is in the bottom 25%, the middle 50%, or the top 25%. Since students are randomly
assigned to each other, a_2 is not driven by selection. Therefore, Sacerdote does not need to
include any control variables for the students background. Table III shows the results to the
aforementioned regression which measures peer effects in GPA. In addition, Sacerdote
includes measures for peer effects in late graduation. Column 1 shows that for an increase of
1 point in student j’s GPA, student i’s GPA increases by 12%. However, we cannot infer any
causal interpretation from this because it might be due to a reflection problem (regressing
outcomes on outcomes). Moreover, another problem arises: common shocks to the whole
dorm. Therefore, Sacerdote includes (Column 2) dorm level fixed effects for 29 different
dorms and shows that, even then, the coefficient of student j’s GPA remains significant.
Moreover, he picks pairs of students from the same floor and shows there is no correlation,
eliminating the assumption of a common shock. Column 3 indicates that this GPA effect
vanished in the last year. Academic index is still important. Columns 4 to 6 show the effects
of j’s background on i’s GPA. Sacerdote includes four dummies of whether or not the
academic scores of i and j are in the top or bottom 25%. The middle 50% is the omitted part.
Column 4 suggests that i’s GPA increases by 6% (significant at 5%) if j is in the top 25%.
Adding dummies for i’s academic score makes this result drop in significance to 10%. If i is
in the top 25%, his GPA will increase by 17.4%, whereas if his roommate is in the top 25%,
i’s GPA will increase by 4.7%. This finding is, however, not very robust. Column 6 shows
that, if j intends to graduate with honors, i’s GPA will be positively affected. Column 7
indicates that there is no correlation for i’s late graduation and j’s late graduation. Moreover,
the choice of major is not affected by peer’s choices. Table 6 shows that roommate j in the
top percentiles help student i if he is also a top student or if he is in the bottom percentiles.
The latter estimate is significant at a 10% level.

You might also like