Laptops in the Long Run: Evidence from the One Laptop per Child Program in Rural Peru

531 rural primary schools from the Peruvian OLPC randomization; school census data 2007-2019; 22,861-28,516 individual student observations in the pooled trajectory regressions; a 2013 survey of 70 treatment and 70 control schools covering 167-168 teachers and ~2,130 students

Students and teachers in rural Peruvian public primary schools, cohorts entering second grade 2009-2016, followed through primary and secondary completion and university application to 2019

Long-horizon follow-up of the Peruvian OLPC cluster-randomized experiment, using the national school census (2007-2019), the national standardized examinations in mathematics and reading (grade 2 annually 2007-2016, grade 4 in 2016 and 2018, grade 8 in 2015/2016/2018/2019, individually matched across grades), and the SIAGIE enrolment register covering the full national student population 2012-2019. Randomization strata-by-cohort fixed effects; standard errors clustered at the primary school level; Benjamini-Hochberg q-values reported alongside p-values because of the number of outcomes. IMPORTANT SAMPLING FACT, easy to get wrong: this is NOT a re-survey of the schools in Cristia et al. (2017). Those 320 schools are explicitly EXCLUDED, because the government treated their controls once that study ended in 2011. The 531 schools here come from the same national randomization but are a disjoint sample, so this is an independent long-run estimate on sibling clusters rather than a longer look at the same clusters. Internal validity is preserved because the original selection was done at the randomization-strata level. Treatment intensity persisted: treated schools were near one XO laptop per student from 2009, control schools received almost none through 2010 and drifted up to about 0.4 per student by 2019, leaving a large and significant gap throughout. No differential change in internet access, desktops, non-XO laptops, computer labs, electricity, water, libraries, science labs, pupil-teacher ratio or enrolment, so the estimate is attributable to personal laptops and not to a bundle. Exam coverage falls over time (59-71% of schools 2007-2013, 39% by 2016); the paper analyses attrition and shows participation was unrelated to treatment with results robust to dropping low-coverage years. MECHANISM DATA IS THE MOST VALUABLE PART for this archive: teachers got the training (+35pp received any XO training, over a 28% control base) but gained no digital skills (XO test +0.157 ns, PC test -0.255 ns, internet test -0.317*) and barely changed classroom practice (+0.8 hours/week of computer use on a 3.7-hour base, ns overall). Students used the laptops significantly more for entertainment (+14pp) than for academic purposes (+10pp, ns). This is the same mechanism Barrera-Osorio & Linden (2009) documented in Colombia, measured a decade later on a different continent. Published as an NBER working paper (w34495, November 2025) and in the Journal of Public Economics (2025). Run by GRADE (Peru), the Inter-American Development Bank, Northwestern and UT Austin; no vendor involvement.

Ten years of administrative data on 531 randomized rural Peruvian primary schools: nothing. School-level academic performance shows no discernible pattern over eight years and rules out effects above 0.05 SD; the pooled individual index of academic performance is -0.046 (SE 0.045), ruling out effects above 0.04 SD. Grade progression is 1.0 percentage point WORSE in treated schools (control mean 86.9%), and on-time primary completion is 2.1 points lower (p<0.10, though not after multiple-testing adjustment). No effects on secondary completion, eighth-grade scores, university application or enrolment, or total years completed. Mechanism data explains why: teachers received the training, gained no digital skills, and did not change how they taught; students got better at using the XO (+0.41 SD) and used it mostly for entertainment. And the cognitive-skills story from the 15-month evaluation does not survive — Raven's +0.046 (ns), coding +0.034 (ns), pooled index +0.136 (ns) after four years of exposure.

Minimal (cluster-randomized at school level, strata-by-cohort fixed effects, documented baseline balance on 2007-2008 census and examination data).

Confirms and extends the short-run nulls of Cristia et al. (2017) and Beuermann et al. (2015) at a ten-year horizon on administrative data, and confirms Fairlie & Robinson (2013) and Malamud & Pop-Eleches (2011) on the computer-skills-yes/learning-no pattern. It is also the decisive non-replication of the cognitive-skills result in Cristia et al. (2017) — four years into exposure, Raven's is +0.046 (ns), coding +0.034 (ns), and the pooled cognitive index +0.136 (ns), against the marginal +0.110 (p<0.10) reported at 15 months.