The Impact of Restorative Practices on Violence, Psychological Well-being, and Social Cohesion in Bogotá’s Schools

The intervention is rooted into Restorative Practices, and it consist of 4 sessions:
1. Sensitization on school violence and conflict dynamics
2. Empathy and perspective-taking, identifying the different roles in a conflictive situation.
3. Emotional self-awareness and self-regulation, identifying my role in current conflictive situations
4. Assertive and cooperative conflict resolution exercises and tools provision.
Delivery: Facilitated by trained staff from the Secretary of Education. Each session lasts 1.5 hours. Schools and head teachers are provided advice and encouraged to do folllow-up activities with the classroom.
Timing: Four-session intervention delivered over a 2–4 weeks period.

2025-05-01

2026-04-30

Four confirmatory families plus a human-capital umbrella, each summarized by a control-referenced index (first factor score, z-scored to control; equal-weighted z-average as robustness): (1) Violence — violence_index (15 EPVS items; emotional and physical sub-indices; a discrimination sub-index; Sistema de Alertas incidents; discipline records); (2) Psychological well-being — phsyco_wellbeing_index (self/peers/teachers connectedness + school motivation); (3) Mental health — index of GAD-7, PHQ-9, and stress; (4) Social cohesion — belonging, positive/negative network ties, willingness to spend time together, and cooperative-game team performance. Human capital is a primary umbrella with three separate sub-indices: cognitive performance (in-survey cognitive games), academic performance (in-survey knowledge test and self-reported/administrative grades and promotion), and school trajectory (attendance, retention, transfers, dropout, repetition).

Indices are first-factor scores standardized to the control group. There are seven primary average-effect tests (the four family indices plus the three human-capital sub-indices). Inference uses randomization inference (within-pair permutation) with Romano–Wolf family-wise correction; Benjamini–Krieger–Yekutieli FDR q-values reported. Tests are two-sided at α=0.05.

Mechanisms: socioemotional skills (empathy, self-awareness, conflict-resolution skills); attitudes toward violence; reporting norms (willingness to report under a randomized public vs. anonymous frame); preference for reparation vs. sanction; identity shift (moral/prosocial, agency, victim, reputation, defender); and behavioral prosociality games (dictator, trust, cooperation).

Each mechanism family is aggregated into a control-referenced index (first factor score, z-scored to the control group; equal-weighted z-average of standardized items as robustness), directions harmonized so that higher values denote improvement, with constituent items also reported. These outcomes are secondary/exploratory: they trace the channels through which the intervention operates and help interpret null or attenuated primary effects. Estimation uses the same model as the primary outcomes (role interactions, pair/stratum and wave fixed effects, SEs clustered at the classroom, randomization inference); tests are two-sided at α=0.05 and adjusted for multiplicity within family using Benjamini–Krieger–Yekutieli FDR q-values. The reporting-norms module additionally identifies the norm-shift as the treatment × public-frame interaction (the imagined-audience "tax" on reporting), not the main effect.

H1 RP reduces violence; H2 improves well-being; H3 mental-health effect a priori ambiguous, registered as role-conditional; H4 improves cohesion; H5 improves human capital (cognition, grades, trajectory). H6 (by role): effects are stronger for students in conflict pre-treatment — victims gain on learning but may experience null-to-worse short-run mental health (welfare trade-off, consistent with Cilliers, Dube & Siddiqi 2016), while aggressors gain across fronts. H7: effects concentrated among sub-populations more likely to be victims (LGBTQ, female, disability, victim per Alertas, displaced/migrant). H8: mechanisms via socioemotional skills, norms (incl. reporting), and prosociality. Directional predictions, two-sided tests.

At the government's request to scale the program, in ~40 schools the intervention expands from one to several nominated classrooms per school (up to four additional). Each of these schools nominates 4 or 6 grade 6–11 classrooms in conflict, and I randomize half within the school to treatment. I measure the same outcomes as in the main trial, but with administrative data only (attendance, retention, grades, promotion, transfers, dropout, and Sistema de Alertas incidents), unless additional funding allows survey collection. This within-school design lets me (i) test spillovers through grade/school treatment saturation, and (ii) for classrooms not randomized, apply a difference-in-differences design against control schools. This component is just beginning.

Real time randomization done in a computer

Unit of randomization: School level.

Yes

Sample Size: 216 classrooms total — 108 treatments, 108 control — from 216 schools. ~33 students per classroom in average

All students in sampled classrooms will be interviewed (N=~9000, with 30 students per classrooms in average)

108 contro, 108 treatment.

Pooled ≈100 per arm (target 102/102), α=0.05, power=0.80, two-sided; realized class size ≈26.5 (power_calculations_phase2.py). ATE MDEs ≈0.09–0.11 SD (violence 0.113, well-being 0.106, stress 0.089, belonging 0.091); within-victim/aggressor MDEs ≈0.16–0.18 SD. ICCs: violence 0.045, well-being 0.035, stress 0.013, belonging 0.016; prior-wave cognitive ICC ≈0.060.To increase statistical power, the experiment will include a pure control arm that won’t be invited to the intervention nor surveyed. I will test impacts using administrative data (which includes violence and educational attainments).