Data Source

This study utilized data from the 2015 Trends in International Mathematics and Science Study (TIMSS). The U.S. TIMSS sample comprised 10,221 students, representing over 3.8 million eighth graders nationwide. The dataset included comprehensive background questionnaires capturing students’ school experiences, socio-demographic characteristics, and contextual factors.

Data Preparation

The data was cleaned by removing 121 cases with missing peer victimization responses. Key variables were selected, including victimization items, plausible value scores for math and science, and covariates (e.g., age, gender, SES). Categorical variables were recoded as factors. Following Solberg & Olweus (2003), peer victimization items were recoded into two categories based on frequency: 1 = “Never” or “A few times a year”, and 2 = “Once or twice a month” or “At least once a week”. The analytic sample included students with complete responses on all peer victimization items (10,100). A listwise deletion would reduce the sample size to 8384.Overall, less than 1% of the individual data points across the dataset are missing (see plot below).

Measures

Latent Class Analysis

To identify subgroups of students with distinct experiences of peer victimization, we estimated latent class models ranging from 2 to 7 classes using poLCA package. Model selection was guided by theoretical interpretability and statistical criteria such as AIC, BIC, ABIC, and CAIC. Fit indices improved markedly from 2 to 3 classes, with BIC reaching its minimum at the 3-class model (see Figure below). While AIC continued to decline slightly with additional classes, gains were marginal and came at the cost of parsimony. Consequently, a 3-class model was selected as the optimal solution.

The three latent classes are:

• Non-victimized (74%): Students in this group had low probabilities of endorsing any form of victimization.
• Sometimes Victimized (21%): Students in this class showed moderate endorsement of items related to verbal and relational victimization (e.g., “Left Out”, “Spread Lies”, “Made Fun”).
• Victimized (5%): Students in this group exhibited high probabilities of experiencing multiple types of victimization, particularly more severe forms such as “Hit or Hurt,” “Stole From Me,” and “Forced Actions.”

The figure below shows item-response probabilities across the three latent classes. In the Non-victimized group, the likelihood of endorsing any victimization item was minimal-typically below 0.25. In contrast, the Victimized class demonstrated consistently high probabilities (often exceeding 0.75) across nearly all indicators. The Sometimes Victimized class displayed intermediate probabilities, particularly for relational and reputational harm such as being left out or having lies spread about them.

Propensity Score Analysis

To estimate the causal effect of peer victimization on academic performance in math and science, we implemented a multinomial propensity score weighting approach using the twang package. The exposure was students’ latent class membership: Victimized, Sometimes Victimized, and Nonvictimized. Propensity scores were estimated using a generalized boosted model with 5,000 trees. The model was optimized using both the effect size mean (es.mean) and the Kolmogorov-Smirnov mean (ks.mean) as stopping criteria.

Analysis of Treatment Effects

To estimate the impact of peer victimization on academic outcomes, weighted linear regression models were fitted separately for mathematics and science performance. The models used latent class membership as the key predictor, with the non-victimized group serving as the reference category.

The plot below presents the regression estimates and confidence intervals for each victimization group relative to the non-victimized reference, across both academic domains:

• Math Performance (blue, circles):

• Science Performance (orange, squares)

• Students classified as victimized had significantly lower scores in both math and science compared to their non-victimized peers.

• Their math scores were, on average, approximately 25–30 points lower.

• Their science scores showed a similarly negative pattern, with an estimated decrease of nearly 30 points.

• The sometimes victimized group also demonstrated lower academic performance, but the magnitude of these differences was more modest i.e, Math and science scores for this group were estimated to be 10–15 points lower on average than the non-victimized group.

All estimates were statistically significant based on robust standard errors derived from the survey-weighted regression models.

The results confirm a graded relationship between peer victimization and academic performance, with increasing exposure to victimization associated with progressively worse outcomes in both math and science. The visual representation in Figure 5 highlights this trend, showing the most substantial performance decline among the victimized subgroup. These findings provide strong empirical support for the hypothesis that peer victimization undermines cognitive and academic functioning, even after adjusting for a wide range of potential confounders.

Policy Implications

• Comprehensive Screening: Establish systematic, school-wide screening protocols to detect all levels of peer victimization, from occasional incidents to chronic patterns.
• Tiered Intervention Framework: Adopt a multi-tiered system of support (MTSS) that aligns intervention intensity with students’ victimization risk profiles, ensuring resources are allocated efficiently and equitably.
• Integrated Academic Recovery: Combine academic support services—such as tutoring and enrichment—with anti-bullying efforts to address the full impact of victimization on learning outcomes.
• Data-Driven Decision-Making: Incorporate peer victimization indicators into early warning systems and MTSS data dashboards to proactively identify at-risk students and guide timely intervention

Conclusion

Peer victimization has a measurable, negative effect on math and science outcomes. Schools and districts must act to identify, intervene, and support impacted students—both socially and academically.