Abstract
Adversity and poor mental health are concentrated in the populations whose human capital matters most for escaping disadvantage, yet crisis education policy typically funds academic inputs alone. This study estimates the causal effects of a scalable, lay delivered adolescent mental health program and of structured academic tutoring, separately and together, on learning, psychological wellbeing, and attitudes toward other communities, among secondary school students living through active ethnic conflict.
We implement a two by two factorial experiment among adolescents in grades 8 to 10 in Kangpokpi district, Manipur, India, crossing a low intensity group counseling protocol (Problem Management Plus) with structured small group mathematics tutoring. The four arms are a waitlist control, counseling only, tutoring only, and the combined treatment. Randomization is at the section level, where a section is a classroom, that is, a grade cohort within a school in a given wave; each section is assigned as a whole to one arm, so that arm varies across classrooms rather than within them. Assigning whole classrooms removes contamination between treated and control students who share a room and matches how such a program is actually deployed. Because the number of independently randomized sections, not the number of students, sets statistical power, the design accumulates sections by enrolling every accessible school, by treating separate classrooms within a grade as separate sections, and by pooling successive cohorts. Assignment is blocked by grade and school type, spanning government, mission, private, and relief camp learning sites.
The study is powered for the two main effects, the effect of counseling and the effect of tutoring, on three prespecified outcome families. The primary learning outcome is performance on a standardized mathematics assessment administered at endline, with items spanning grade 8, 9, and 10 content so that it captures a wide and disrupted range of learning levels without floor or ceiling effects. The psychological family comprises distress (PSS-10) and wellbeing (WHO-5), and the social family comprises attitudes toward others, including general empathy, empathy toward other communities, and intergroup contact and comfort, with school engagement, aspirations, hope, and perceived support as additional prespecified outcomes. The psychological and social outcomes are measured at baseline and again at endline. The interaction between the two treatments, the test of complementarity between a psychological and an academic input, is prespecified as a secondary and exploratory parameter and reported with transparent confidence intervals, recognizing that a section level factorial detects an interaction with far less precision than a main effect of the same size. Prespecified heterogeneity analyses examine how effects vary with baseline psychological distress and with a baseline index of conflict exposure built from relocation, schooling disruption, and perceived safety.
Each outcome is regressed on the two treatment indicators and their interaction, with block fixed effects for grade and school type and baseline covariates, including the baseline value of each outcome that is measured before randomization, to raise precision. Treatment is clustered at the section level; given the modest number of clusters, inference uses cluster robust standard errors together with randomization inference and the wild cluster bootstrap, and balance is assessed across sections rather than across students. Corrections for multiple testing are applied within each outcome family. Because the site is an active conflict, the final number of sections, schools, and students, the assessment and survey instruments, and the timeline are fixed in the preanalysis plan following a pilot and may be revised as security conditions require; the target is as many sections as access permits across multiple relatively stabilized locations, pooling cohorts to raise the number of randomized units. School closure and displacement, which can remove an entire section and therefore an entire cluster, are anticipated and handled by oversampling sections, intensive in person tracking, inverse probability weighting on baseline characteristics, and prespecified bounds, with differential attrition reported as a first order diagnostic.