Heterogeneous peer effects of college roommates on academic performance

The study investigates how college roommates influence each other’s academic performance, addressing the long-standing challenge of disentangling peer influence from peer selection in dynamic educational environments. While peers can exert both positive and negative influences on behavior and learning, establishing causal peer effects is difficult due to reverse causality, confounding factors, and homophily. Chinese university dormitories provide a quasi-experimental context: 4-person rooms, stable roommate relationships, and plausibly random assignment at entry, enabling observation of persistent peer interactions outside the classroom. The research aims to quantify the magnitude and evolution of roommate peer effects on GPA, and to assess how peer heterogeneity and a student’s ordinal rank within the room moderate these effects. The work combines a null-model framework with longitudinal regression analyses to provide evidence of peer influence and to inform education management and policy.

Prior work across social contagion and education demonstrates that peers shape behaviors and outcomes, including academic achievement, health behaviors (e.g., smoking, alcohol), and longer-term life outcomes (e.g., earnings, career choices). Classroom-based studies have documented ability peer effects and the challenges of causal identification. Randomized or quasi-random roommate assignments in higher education (e.g., Dartmouth; Chinese universities) offer evidence of peer effects in dorm contexts. Methodologically, social science studies often rely on regression with longitudinal controls to mitigate reverse causality, while network science introduces null models and permutation tests to detect non-trivial structures. Heterogeneous peer effects and the role of ordinal rank have been emphasized in recent literature, suggesting that both the composition of peer groups and a student’s relative standing may independently affect performance. This study integrates these strands by applying null models to quantify assimilation beyond chance and by estimating moderated peer effects with fixed-effects OLS models.

Data: The authors analyze longitudinal accommodation and academic records from a public research university in China, covering 5,272 undergraduates in identical 4-person dorm rooms across two cohorts (admitted 2011 and 2012). Roommate assignments were administratively determined without input on academic performance, socioeconomic background, or preferences, and reassignment was rare; roommates typically stayed together until graduation. GPA data for the first five consecutive semesters (through 2014) were collected and normalized within cohort-major-semester to percentiles R in [0,1]. All data were anonymized and IRB-approved (IRB No. 1061420210802005). Tier combination analysis: Students’ semester GPA percentiles were discretized into 2-tier, 3-tier, and 4-tier systems. For each dorm, the unordered combination of roommate tiers was recorded. To account for differing theoretical frequencies of combinations under independence, the relative ratio E = (Pa − Pt)/Pt was computed, where Pa is the observed frequency and Pt the theoretical frequency. Pairwise absolute tier differences were summarized by D = (1/6)∑|lu − lv| over all roommate pairs to capture within-room heterogeneity. Assimilation metric and null model: At the granular level, assimilation A was defined for each dorm as A = 1 − (1/6)∑|Ru − Rv|, where Ru are GPA percentiles. Under i.i.d. GPAs (no peer effects), E[A] = 0.5. A roommate null model was constructed by randomly shuffling students among dorm rooms while preserving cohort, gender, and major compositions. The shuffle was implemented 1,000 times to generate null distributions. Statistical comparisons between actual and null-model assimilation used permutation tests and Student’s t-tests. Temporal trends were evaluated by computing the percentage difference between mean actual and mean null-model A each semester; standard errors were obtained by clustering across 100 independent implementations for visualization. Regression analyses: To assess dynamic associations (Granger causality style), the authors estimated OLS models relating a student’s next-semester GPA (GPA_Post) to their own prior GPA (GPA_Prior), roommates’ average prior GPA (RM_Avg), and within-room GPA dispersion (RM_Diff), including interaction terms and controls: Model (4): G_{i}^{s+1} = b0 + b1 G_i^s + b2 RA_i^s + b3 RD_i^s + b4 (RA_i^s×RD_i^s) + gender, major FE, cohort FE, semester dummies + ε_i. Model (5): Adds the student’s in-dorm ordinal rank OR_InDorm (1=highest, 4=lowest, based on prior GPA) and its interactions with GPA_Prior, RM_Avg, and RM_Diff. Independent variables (except dummies) were mean-centered; robust standard errors were reported. Falsification tests ran identical regressions on the null-model roommate assignments.

• Over-representation of homogeneous tier combinations: Under 2-, 3-, and 4-tier schemes, combinations with similar or adjacent tiers show significantly positive E, while those with distant tiers show negative E. Under 2-tier, 1111 and 2222 are over-represented (P<0.001), while 1112 and 1122 are under-represented (P<0.001). • Assimilation beyond chance: Mean actual assimilation A=0.549 versus null-model mean A=0.496, a 10.7% increase; means differ significantly (Student’s t-test, P<0.001). Results are robust by semester. • Increasing peer effects over time: The percentage difference between actual and null-model mean assimilation increases from semester 1 to 5 overall, indicating stronger peer effects as roommates live together longer (with a notable rise by semester 3; slight disruption by semester 5). No significant gender differences in assimilation. • Positive roommate average effect (OLS): Without controls, RM_Avg → GPA_Post coefficient b=0.365 (P<0.001). With controls and fixed effects, RM_Avg remains positive and significant (b≈0.050; P<0.01), while GPA_Prior has the largest effect (b≈0.801). Effect size: a 100-point rise in RM_Avg associates with ~5-point increase in GPA_Post, about 6% of the own-prior effect and ~10% of average GPA. • Moderation by peer heterogeneity: RM_Diff alone is not significant, but the RM_Avg × RM_Diff interaction is negative (Table 1: −0.089*), indicating the positive RM_Avg effect is stronger when peers are similar (slope b=0.055; 95% CI [0.040,0.070] at high RM_Diff per text’s plotted interpretation) and weaker when peers are dissimilar (slope b=0.028; 95% CI [−0.001,0.057]). Falsification on the null model shows nontriviality. • Ordinal rank effects: OR_InDorm predicts higher GPA_Post (e.g., b=0.011**, then b≈0.006** with full controls; note larger rank = lower prior GPA standing). The OR_InDorm × RM_Diff interaction is negative (−0.022**): the positive ordinal-rank effect is significant when RM_Diff is low (slope b=0.007; 95% CI [0.002,0.012]) but not when RM_Diff is high (slope b≈0.000).

The findings indicate that roommates influence academic outcomes beyond what random assignment and chance would produce. Higher assimilation compared to null-model expectations, and its increase over time, align with mechanisms of sustained interaction and spillovers in close-knit dorms. Regression results show that, after accounting for a student’s strong path dependence (own prior GPA), roommate average performance positively predicts future performance, with the effect moderated by heterogeneity: peer influence is more pronounced when roommates are similar. Additionally, the student’s ordinal rank within the dorm exerts an independent influence, particularly in more homogeneous rooms. These results suggest the educational significance of everyday peer environments outside classrooms and support policies that consider roommate composition to enhance learning, while acknowledging that effect sizes, though modest, match magnitudes reported in prior literature.

This study integrates null-model and longitudinal regression approaches to quantify and unpack college roommate peer effects. It contributes an assimilation metric and permutation-based null modeling to assess effect sizes comparably across datasets, and demonstrates that peer effects strengthen with time together. It further shows that peer heterogeneity moderates the benefits of high-performing peers and that ordinal rank within the room independently predicts future performance. Future research should generalize across institutions and countries, incorporate richer pre-college and demographic data to test randomization more directly, refine GPA normalization methods, and probe mechanisms (e.g., peer pressure, identity, social networks, classroom interactions) that mediate assimilation and influence.

• External validity: Data come from two cohorts at a single Chinese university; generalizability to other contexts is uncertain. • Random assignment verification: Assignment is plausibly random by procedure, but lack of comprehensive pre-college and demographic data limits direct testing of randomness. • Measurement/normalization: GPA percentiles normalized within cohort-major facilitate comparisons but may discard distributional information; alternative normalizations could improve precision. • Omitted mechanisms: Factors beyond dorm life (orderliness, classroom interactions, social networks, behaviors, common shocks) may mediate effects; mechanisms like peer pressure and identity were not directly measured.