Remedying Education: Evidence from Two Randomized Experiments in India

Vadodara CAL arm: 111 municipal primary schools (55 treatment, 56 comparison) in year 2, 56 vs 55 in year 3, all grade four; 5,945 grade-four children in the year-2 sample and 5,945 in year 3; regression samples of 5,732 (year 2) and 4,688 (one-year follow-up). Over 15,000 students across both experiments and three years.

Grade-four children in Vadodara Municipal Corporation primary schools, urban Gujarat, India — poor urban-slum government schools with very low baseline achievement. (The paper's other arm, the Balsakhi remedial tutor, covered grades three and four in Vadodara and Mumbai and is not an ed-tech intervention.)

INSTRUCTIONAL intervention, not an access one — this is the paper that separates the two. The computers were largely ALREADY IN THE SCHOOLS; what the programme added was structured, level-targeted content plus a paid instructor to keep children on task. Comparison schools were free to use their own computers and were never observed doing so for instruction, which is exactly the Barrera-Osorio/Cueto mechanism showing up as the counterfactual here. DOSE AND THE INSTRUCTIONAL-TIME QUESTION, which the archive must get right: two hours per week of shared computer time, two children per machine — ONE HOUR DURING CLASS TIME AND ONE HOUR IMMEDIATELY BEFORE OR AFTER SCHOOL. So half the dose substituted for regular instruction and half was added instructional time. The +0.35/+0.47 SD is therefore not a like-for-like productivity gain against the same hour; roughly half of it is bought with extra time. `baseline` is recorded as business-as-usual because the comparison group got the normal school day and nothing else, but a fair reading discounts for the added hour. Design: school-level randomization stratified on Balsakhi treatment status, gender, language of instruction and prior-year math scores; the two experiments were crossed in year 2 so their effects and interaction could be separately estimated (the interaction is small, negative and insignificant). Treatment and comparison groups were SWITCHED between year 2 (2002-03) and year 3 (2003-04), which is why the "second year" estimate is not a two-year-exposure estimate for the same children. Attrition 3.4-3.8% in the treatment years and about 20% at the one-year follow-up, balanced across arms. MEASURE TYPE IS THE MAIN DISCOUNT. The tests were designed and administered by Pratham, the implementing NGO, and built around the competencies the Vadodara Municipal Corporation prescribes for grades one to four — the same competency list the software's games were selected to emphasise. That is a researcher-designed, treatment-aligned measure in the sense METHODOLOGY.md warns inflates effects roughly twofold against independent standardized tests. Pratham took real care with test administration (three staff per room against cheating, home make-up tests to hold down attrition), but the alignment issue is structural, not procedural. Numbers below are from the full text of NBER working paper w11904; the published QJE 122(3):1235-64 version carries the same headline estimates. Funded by ICICI, the World Bank, the Sloan Foundation and the MacArthur Network; Pratham developed and ran the programme, the academic authors designed and ran the evaluation. GRADE: B rather than A. The paper as a whole is a large multi-site experiment, but the CAL ARM — the part this record is about — is a single-city, single-grade RCT in 111 Vadodara schools whose outcome measure was built by the implementing NGO around the same competency list the software targeted. Strong single RCT, discounted for treatment-aligned measurement; the Balsakhi arm's two-city replication does not transfer to the CAL estimate.

A shared-computer math game programme run inside Indian municipal primary schools raised math scores by 0.35 SD in its first year and 0.47 SD in its second, with no effect on language (point estimates near zero, as expected since the software was math-only) — and, unlike the remedial-tutor arm, it helped children at every point in the ability distribution roughly equally. Then it went away. One year after the programme ended, the math effect was +0.097 SD (SE 0.053) and the combined score effect was +0.008; roughly 80% of the gain had evaporated in twelve months. The authors read the residual 0.10 SD as encouraging; the honest reading for a database organised around fadeout is that a large end-of-treatment effect on an implementer-designed, curriculum-aligned test decayed to something indistinguishable from noise within a year of withdrawal. Two further facts belong on the record: half the dose was extra instructional time outside the school day, and the comparison schools already had computers and never used them for teaching.

Minimal (school-level randomization, stratified, with the crossed Balsakhi design and balance checks).

The CAL result is the founding positive of the developing-country computer-assisted-learning literature and has been broadly reproduced in direction: Muralidharan, Singh & Ganimian (2019) found +0.37 SD in math from an adaptive after-school programme in Delhi, and Muralidharan & Singh (2025) found +0.22 SD after 18 months from the same product delivered inside rural government schools. It contrasts sharply with the null CAL/computer results from richer settings and from access-only interventions (Angrist & Lavy 2002; Barrera-Osorio & Linden 2009; Cristia et al. 2017). The FADEOUT recorded here has itself been replicated in the Chinese CAL literature (Mo et al. 2015). What has never been replicated is a durable CAL effect.