Does Leveling the Classroom Prevent Drop-Out? Impact Evaluation of the Complex Instruction Program in Hungary

Treatment schools receive: (i) a 30-hour CIP training workshop delivered by experienced CIP trainers in June-August 2026; and (ii) ongoing mentoring via lesson plan feedback, class visits and individual meetings on demand for the following academic year. Trained teachers implement one CIP lesson approximately every two weeks per subject for grades 5–8 students over a nine-month period beginning September 2026. Control schools receive equivalent training and begin CIP implementation in June 2028, after the evaluation period.

2026-06-29

2027-06-10

1. Mathematics test score
2. Reading test score
3. Dropout by and after 10th grade in secondary school

Primary outcomes are specified on the student-level.
1. Mathematics test score (NABC) standardized to mean 0, SD 1 using the control group distribution.
2. Reading literacy test score (NABC) standardized to mean 0, SD 1 using the control group distribution.
3. Dropout, binary outcome equal to 1 if student dropout out by and after 10th grade in secondary school, note: this data becomes available only in 2030, the earliest.

Student-level outcomes:
1. Predicted drop-out probability
2. Upper-secondary school quality (after grade 8)
3. Student attendance
4. Student grades
5. Non-cognitive skills (from student surveys): Pre-specified scale indices for non-cognitive skill variables (under construction).

Teacher-level outcomes:
1. Behavior management difficulty
2. Low student achievement attributions
3. Growth mindset
4. Generalized self-efficacy
5. Teaching enjoyment
6. Job-related stress and burnout
7. Job satisfaction

Student-level outcomes:
1. Predicted drop-out probability: Using the observed correlation between short-term outcomes and drop-out in the ADMIN4 dataset, we will construct a predicted drop-out index. This provides the primary mechanism for long-run inference within the project timeline.
2. Upper-secondary school quality (after grade 8): The quality of the secondary school to which a student is admitted after completing grade 8, measured by average NABC scores of the secondary school. Admission is merit-based; higher-quality placement indicates stronger performance.
3. Student attendance: Number of days absent, collected from school administrative records.
4. Student grades: End-of-year grades in mathematics and Hungarian language, collected from school administrative records.
5. Non-cognitive skills (from student surveys): Pre-specified scale indices for non-cognitive skill variables (under construction). Each scale will be standardized using baseline mean and SD. A combined non-cognitive skills summary index (Anderson; 2008) will also be constructed.

Each teacher-level index will be standardized using the baseline mean and SD. A combined summary index (Anderson; 2008) will also be constructed. The precise set of questions asked from teachers is already available and attached to this registration.

We use a multi-site, school-level cluster-randomized trial in Hungary with a delayed-rollout control design. Schools willing to implement CIP are recruited and offered coverage of initial implementation costs (teacher training and mentoring).

Not available

The unit of treatment assignment is the school site. Assignment follows a 1:1 ratio between treatment (early rollout) and control (delayed rollout). Randomization is conducted within matched groups (strata) formed using administrative records from 2024. Matching cells are defined by two categorical variables: settlement type (city vs. village) and school type (state vs. church). Within each cell, school sites are ranked by the first principal component of a PCA using the following school-level variables for grades 5–8: share of disadvantaged students, share of cumulatively disadvantaged students, share of grade repeaters, share of students missing 10–250 hours of school, and share of students missing more than 250 hours. School sites are then grouped into quadruplets based on this ranking.

School sites that belong to the same school are handled as follows. The treatment status assigned to the main school is applied to all of its sites. To preserve balance, we identify a number of similar matched schools equal to the number of school sites belonging to the main school. For example, five matched schools for a school with five sites, two for a school with two sites. Treatment is then assigned between the multi-site school (together with its sites) and the matched schools as a unit.

Within each matched group, treatment is assigned by a computer-generated random number sequence and is conducted by our research team in Stata. As we proceeded with school agreements, we randomized school sites on a rolling basis in three waves. Each randomization wave is logged.

School site level

Yes

84 school sites

84 school sites and 8400 students

Staggered design:
Treatment group includes 42 school sites with early rollout teacher training
Control group includes 42 school sites with late rollout teacher training

We compute minimum detectable effect sizes (MDEs) under a two-sided 5% test and 80% power, with 42 schools per arm. For test scores we assume 100 students per site, since short-run score changes are observed across all four grade cohorts (grades 5–8); for drop-out we assume 50 students per site, as it is observed with a two-year delay. Baseline estimates of means, standard deviations, and intracluster correlations come from 2019 administrative data matched to the study schools, using the 8th-grade cohort. The intracluster correlation is 0.14 for mathematics, 0.11 for reading, and 0.14 for drop-out. Including the pre-specified baseline covariates substantially improves precision across all outcomes. For mathematics, the MDE falls from 0.24 SD without covariates to 0.14 SD with the nine covariates; for reading, from 0.21 SD to 0.13 SD. For drop-out, with a baseline rate of about 16% and eight covariates, the MDE is 0.08 — that is, we can detect a reduction in the probability of dropping out of roughly 8 percentage points.