The Evidence on Teaching

Experimental Evidence on the Effects of Home Computers on Academic Achievement among Schoolchildren

Fairlie, R. W., & Robinson, J. · 2013

grade Brctindependentreplicatednumbers spot-checked
Sample
1,123 students (559 assigned to treatment); 15 middle and high schools across 5 California school districts; two school-year cohorts. All participants lacked a home computer at baseline. Follow-up outcomes come from school administrative records, so attrition on the main outcomes is negligible.
Population
US schoolchildren in grades 6-10 (middle-school students are the large majority), 15 California schools, no home computer at baseline
Design
The cleanest ACCESS experiment in the literature and deliberately so — free refurbished computers were delivered to randomly selected students with NO training, NO software curriculum and NO school-side intervention, precisely to isolate "a computer in the home" from "a technology programme". Individual-level randomization; ITT estimates reported throughout (the authors decline to scale to LATE because two-thirds of control-group computer acquisitions happened late in the year). Main outcomes are school-provided administrative records — quarterly course grades, credits earned, unexcused absences, tardies, end-of-year enrolment — plus California STAR standardized tests in English-language arts and math, which are state-administered and entirely independent of the intervention. A follow-up survey supplies the intermediate inputs (homework time, software use, self-rated computer knowledge). THE FIRST STAGE IS THE POINT: a null with a weak first stage is uninformative, and this one is strong (81% vs 26% ownership at follow-up, +2.5 hours/week of total computer use on a control base of 4.2). So the null is a statement about computers, not about compliance. Precision is unusually good: 95% CIs of [-0.10, 0.06] GPA points on course grades and roughly [-0.16, +0.05] SD on STAR scores rule out even modest effects in either direction. Funded by Computers for Classrooms Inc. (a Microsoft-certified refurbisher that also donated machines), the ZeroDivide Foundation and the NET Institute; the refurbisher supplied hardware but had no role in design, analysis or authorship, and the finding is a null, so funder direction is not a live concern here. Numbers below are from the full text of NBER working paper w19060 (May 2013), which matches the AEJ: Applied published version.
Key findings
Free home computers raised ownership by 55 percentage points and total computer use by 2.5 hours a week, and moved nothing. Course grades: ITT ~0, SE 0.04, 95% CI [-0.10, 0.06] GPA points. STAR standardized scores: point estimates near zero, 95% CIs of [-0.15, 0.05] SD in English and [-0.16, 0.04] SD in math. Credits earned, unexcused absences, tardies, and end-of-year enrolment: all null. A pooled z-score of the three main academic outcomes is -0.05 SD (SE 0.05). The mechanism is visible in the usage split: the extra computer time divided almost evenly between schoolwork (+0.8 h/wk), games (+0.8 h/wk) and social networking (+0.6 h/wk), and homework time did not move in either direction. This is the strongest available evidence that device access alone is neither a lever nor a hazard.
Genetic confound
Minimal (individually randomized).
Replication notes
The core result — a large first stage in computer access with nothing downstream — is the most reliably reproduced finding in this cluster. Malamud & Pop-Eleches (2011, Romania) got a negative on grades from the same intervention; Beuermann et al. (2015, Lima) and Cristia et al. (2017, rural Peru) got nulls on achievement from randomized laptop provision; Cueto et al. (2025) extended the Peruvian null to a ten-year administrative horizon. The direction differs at the margins (null here, negative in Romania), and the plausible reason is recorded in this paper's own data - US treatment children reported no reduction in homework time, whereas Romanian winners did.
DOI / URL
10.1257/app.5.3.211

Effects

OutcomeMetricValueMeasureTimingVsHorizonClass
First stage: home computer ownershippp81% of treatment vs 26% of control reported a computer at follow-up (+55pp); only 11 of 559 treatment students failed to collect their machine, so the true first stage is if anything understated. Home internet service also +25pp (42% vs 17%).self-report surveyend of the school yearbusiness-as-usualend-of-treatmentbehaviour
First stage: total hours of computer use per weekhours/week+2.5 hours/week on a control mean of 4.2 (a ~60% increase). The increase is entirely home use; no reduction in computer use at school or elsewhere, so this is added access rather than relocated access.self-report surveyend of the school yearbusiness-as-usualend-of-treatmentbehaviour
Course grades (numeric GPA equivalent, quarters 3-4, all subjects)GPA pointsITT indistinguishable from zero, SE 0.04; 95% CI [-0.10, 0.06] for all subjects and [-0.05, 0.11] for academic subjects — each half-interval is about one-fourth of a plus/minus grade modifier. Pass rate effect 0.00 (SE 0.01). No effect in any individual subject; quantile treatment effects indistinguishable from zero across the whole distribution.administrativequarters 3 and 4 of the treatment school yearbusiness-as-usualend-of-treatmentdomain-skill
California STAR standardized test, English-language artsSDpoint estimate near zero with or without prior-year score controls; 95% CI [-0.15, 0.05] SD. Proficiency-or-above indicator 95% CI [-0.04, 0.04].standardizedspring of the treatment school yearbusiness-as-usualend-of-treatmentdomain-skill
California STAR standardized test, mathematicsSDpoint estimate near zero; 95% CI [-0.16, 0.04] SD. Proficiency indicator 95% CI [-0.08, 0.04]. Inverse-CDF plots show near-complete overlap of treatment and control distributions, so the mean null is not masking offsetting tails.standardizedspring of the treatment school yearbusiness-as-usualend-of-treatmentdomain-skill
Pooled academic z-score (grades + two test scores)SD-0.05 (SE 0.05); adding the five administrative outcomes gives -0.02 (SE 0.03)mixedend of the treatment school yearbusiness-as-usualend-of-treatmentdomain-skill
Credits earned, unexcused absences, tardies, end-of-year enrolmentadministrative countsno significant effect on any of them; home computers neither improved nor damaged attendance, discipline or grade progressionadministrativeend of the treatment school yearbusiness-as-usualend-of-treatmentbehaviour
DISPLACEMENT: total time spent on homeworkhours/weekno evidence that treatment students reported fewer (or more) hours of homework. This is the load-bearing contrast with Malamud & Pop-Eleches (2011): the Romanian voucher reduced homework and lowered grades; the Californian computer did not reduce homework and did not lower grades. The authors also note that reported homework effort was low in both arms, leaving room for improvement that never materialised.self-report surveyend of the school yearbusiness-as-usualend-of-treatmentbehaviour
What the computer was used FOR: schoolwork vs games vs social networkinghours/weekschoolwork +0.8 (control mean 1.9), games +0.8 (control mean 0.8), social networking +0.6 (control mean 0.6). The productive and unproductive uses grew in near-equal absolute amounts, and the entertainment uses roughly doubled off a low base.self-report surveyend of the school yearbusiness-as-usualend-of-treatmentbehaviour
Intermediate inputs — turning homework in on time, time on last essay/project, receiving help from peers or teachers electronically, software use, self-rated computer knowledgepp / self-report scalesnull on every one. Notably, self-rated computer knowledge did NOT rise, despite the ownership shock — in contrast to Malamud & Pop-Eleches (2011) and Beuermann et al. (2015), where computer skill was the one thing that clearly moved.self-report surveyend of the school yearbusiness-as-usualend-of-treatmentnear-transfer

Cited by