Adapting for scale: Experimental Evidence on Technology-aided Instruction in India

80 schools (40 treatment, 40 control), ~6,500 students treated annually; test-score regressions on 7,994 students at the year-1 endline and 8,733 at the 18-month endline; grades 1-8

Students in grades 1-8 in integrated public "Adarsh" schools (grades 1-12) in four districts of rural and urban Rajasthan, India — larger and better-resourced than stand-alone primary schools, and the type of government school most likely to run hardware-intensive ed-tech

ADDED BY THE AGENT, not on the assigned list, because it is the missing half of the Mindspark record and the archive's central claim — efficacy-to-effectiveness decay — is directly measurable across this pair. INSTRUCTIONAL intervention delivered by SUBSTITUTION rather than addition, which is the whole point. Computer labs were installed in treated schools and timetables rewritten so that a "computer lab" period replaced regular Math and Hindi instruction: about 12.5% of all weekly classroom time (6 of 48 periods), roughly 25% of weekly Math and Hindi time in grades 1-5, and 40-50% in grades 6-8. Head teachers chose the details; in primary grades the replaced time was usually the same subject, in middle school about half came from the same subjects and the rest from non-targeted subjects and remedial instruction. The regular teacher accompanied students to the lab. Where computers were short, two students shared one. So unlike the Delhi trial this adds NO instructional time — it is a clean test of whether the software makes an existing hour more productive, which is the quantity a school actually has to decide about. Cluster-randomized at the school level with pair-strata; ITT estimates; standard errors clustered at the stratum level; independent in-class tests at baseline (July 2017) and each February endline, IRT-linked across rounds and grades and standardized to grade 5 in the control group at baseline. Students absent on test day were tracked to households. No differential attrition across arms. Pre-registered (AEARCTR-0002546). Vendor relationship, recorded rather than hidden in the enum: the project was executed in partnership with the Government of Rajasthan and Educational Initiatives, EI staff provided operational support and are thanked but are not authors, and EI received Global Innovation Fund money to develop the in-school model being evaluated. The evaluation itself was authored by UCSD and Stockholm School of Economics and funded by the RISE Programme (FCDO, DFAT, Gates Foundation). Independent by this archive's definition; the vendor's stake in the adapted model is the thing to watch. One genuine step beyond the 2019 trial, recorded here rather than in the enum: the academic authors CO-DESIGNED the scalable in-school delivery model with the vendor before evaluating it ("we adapt the PAL implementation for scalability"), so they are not purely external assessors of a product someone else built. They are still the sole authors of the evaluation and the vendor did not run it, which is what the enum turns on. Numbers below are from the full text of NBER working paper w34205 (September 2025, revised November 2025). Not yet peer-reviewed at the time of extraction, which is the main reason not to grade it above A on the strength of design alone.

The strongest positive in the ed-tech literature, re-run the way a government would actually buy it. Mindspark moved from an after-school centre with 619 self-selected Delhi volunteers to 40 Rajasthan government schools and ~6,500 students a year, with lab periods carved OUT of existing Math and Hindi instruction rather than added to the day. It still worked: +0.22 SD math and +0.20 SD Hindi after 18 months, which the authors frame as a 50-66% increase in the productivity of schooling time. But per unit of exposure the effect is roughly a quarter of the Delhi estimate (+0.37 SD math in 4.5 months there; +0.15 SD after six months here). Learning gains were proportional to time actually logged on the platform. This is what a real efficacy-to-effectiveness curve looks like when it does not collapse to zero — deflation, not disappearance — and it is the honest number to quote for adaptive software deployed at scale in a school day.

Minimal (cluster-randomized at school level with pair-strata; no differential attrition).

This is the effectiveness trial for the efficacy result in Muralidharan, Singh & Ganimian (2019) — same product, same lead authors, same vendor, a sample over twenty times larger, not self-selected, delivered inside government schools during the school day. The effect survives but shrinks sharply per unit of exposure (0.22 SD in 18 months versus 0.37 SD in 4.5 months). It also reproduces the Banerjee et al. (2007) finding that in-school computer-assisted learning in India moves math, and reproduces the pattern that gains are targeted below grade level.