The Use and Misuse of Computers in Education: Evidence from a Randomized Experiment in Colombia

97 schools (from 100 randomly assigned; three dropped out) and 5,201 children surveyed at follow-up in grades three through nine at baseline; 426 teachers surveyed

Students in grades 3-9 (at 2006 baseline) in Colombian public schools, largely rural, participating in the government's Computadores para Educar (Computers for Education) programme

ACCESS intervention WITH a teacher-training component, which makes the null more informative than a bare device drop. The programme placed privately-donated refurbished computers in schools and ran a 20-month in-school teacher training run by the Universidad de Antioquia, aimed explicitly at integrating computers into the teaching of Spanish. 100 interested schools randomly assigned in August 2006, baseline survey 2006, follow-up May-June 2008; 97 schools in the analysis. Outcome is a shortened version of the Colombian national examination (SABER) — four grade-banded forms, normalized within grade against the control distribution. That is a standardized, externally built measure, which matters given how much of this literature runs on implementer-designed tests. THE VALUABLE FINDING IS THE IMPLEMENTATION DATA, and it is the mechanism behind roughly half the nulls in this cluster. The treatment was DELIVERED — 96% of treatment schools treated, 4% of controls; 13.4 computers per treatment school against 5.1 in controls; 95% of treatment teachers reported receiving the training against 8% of controls. And then almost nothing happened in the classroom. Only 42-44% of treated teachers reported using a computer at all in the previous week (control 17%). Teachers reported using computers about half a day a week more, but reported NO increase in using them for in-class activities — the authors infer preparation rather than instruction. School-level computer use rose only in grades three and four (+20pp and +16pp) and was insignificant and under 10pp in every higher grade. Most damning, students reported the extra computer use occurred in COMPUTER SCIENCE (+29.4pp, p<0.01) and not in Spanish (+0.02, ns) — the subject the whole programme targeted. Learning to use a computer displaced learning with a computer. Weaknesses to record: attrition to follow-up was 37%, driven by rural migration, but it is closely balanced across arms and estimates with and without baseline controls are virtually identical, which is the standard argument that internal validity survives. Power is adequate but not generous — the design can detect effects of about 0.125 SD at the 10% level, and rejects effects above 0.2 SD in Spanish. A World Bank Policy Research Working Paper (Impact Evaluation Series No. 29); the evaluation was designed and authored by a World Bank economist and a Columbia academic, independent of the programme operator (CPE) and of the private donors of the hardware.

Two years, 97 randomized schools, 5,201 children, computers plus a 20-month teacher training aimed at the Spanish curriculum, and the result was nothing: +0.077 SD in Spanish (SE 0.076) and +0.088 SD in math (SE 0.109), with the design able to reject Spanish effects above 0.2 SD. No effect in any sub-competency, either gender, or (bar two offsetting grade-level outliers) any grade. The paper's contribution is not the null but the reason for it, which it measured directly: the machines arrived, the teachers were trained, and the computers never entered Spanish instruction. Students' extra computer exposure showed up as +29 percentage points in computer science class and +2 points (ns) in Spanish. Teachers reported more computer use overall but no increase in in-class use, and the effect was confined to grades three and four. This is the canonical documentation of the "delivered but not integrated" failure mode.

Minimal (school-level randomization; balanced attrition; estimates unchanged by baseline controls).

The mechanism this paper isolates — hardware delivered, teachers trained, computers never entering the target subject's instruction — is the most reliably reproduced finding in the whole access literature. Cueto et al. (2025) found the identical pattern in rural Peru a decade later on a ten-year horizon (teachers trained, no gain in teacher digital skills, under an hour a week of extra classroom use, no learning effect). Banerjee et al. (2007) recorded the same counterfactual from the other side — Indian comparison schools already had computers and were never observed using them for instruction. Angrist & Lavy (2002) is the earlier null of the same shape.