When schools close during a crisis, the first instinct is often to reach for sophisticated digital learning platforms. Yet in many of the places where educational disruption is most damaging, reliable internet access is precisely what families do not have. A large multi-country experiment suggests that a much simpler technology can make a measurable difference: an ordinary phone call.
A peer-reviewed study in Nature brought together five randomized controlled trials conducted in India, Kenya, Nepal, the Philippines and Uganda during major COVID-19 school disruptions. Across 9,148 children, researchers tested low-tech remote education delivered through weekly text messages, or through text messages combined with roughly 20-minute phone tutorials that targeted instruction to each child’s current numeracy level.
The combined phone and SMS intervention increased learning by 0.321 standard deviations across the five countries, p < 0.001. SMS alone produced a much smaller average effect of 0.078 standard deviations, p = 0.006, and its effectiveness was inconsistent across settings. The phone-based approach also worked when government teachers delivered it, not only when instruction came from non-governmental organisations.
That matters because education interventions frequently weaken when they move from a tightly managed pilot into real public systems. Here, the researchers found evidence that a simple intervention could travel across countries, survive implementation by different organisations and remain comparatively inexpensive.
School closures exposed a problem that internet learning could not solve everywhere
Educational shocks are not rare exceptions. Disease outbreaks, extreme weather, conflict and other emergencies can interrupt schooling for months at a time. The study notes estimates that 222 million children experience regular educational disruption and require education-in-emergencies support. In low-income countries, climate shocks alone are associated with losing about 10% of the academic year on average.
The technological constraint is equally important. The researchers point out that fewer than 15% of households in low-income countries have internet access, while more than 70% have access to mobile phones. That gap changes what a scalable emergency education programme can realistically look like.
Rather than treating the phone as a substitute classroom, the intervention paired accessible technology with a specific teaching principle. Children received foundational numeracy instruction matched to what they could currently do. Regular low-cost assessments helped tutors adjust the next lesson, an approach related to teaching at the right level.
This distinction between platform and pedagogy is central to the study. A phone makes contact possible. It does not, by itself, guarantee effective teaching.
Five randomized trials tested the same basic idea
The trials covered India, Kenya, Nepal, the Philippines and Uganda. Sample sizes were 708 children in India, 1,985 in Kenya, 2,678 in Nepal, 2,469 in the Philippines and 1,308 in Uganda, giving a combined sample of 9,148.
All five countries had experienced schooling disruption associated with the pandemic. The trials were deliberately coordinated so that the researchers could examine a similar intervention across different contexts rather than trying to compare unrelated programmes after the fact.
The core phone intervention typically combined weekly SMS content with approximately 20-minute phone tutorials over eight weeks. Tutors focused on foundational numeracy and adapted instruction to the child’s assessed level. All countries except India also included an SMS-only comparison arm. Nepal and the Philippines included both government and NGO delivery arms, allowing the researchers to test whether the model still worked when embedded in public systems.
Random assignment is an important strength. Because children were allocated to intervention and comparison conditions, differences at the end of the trials can be interpreted much more confidently as effects of the programmes than would be possible in an observational comparison.
Learning was assessed directly with children by phone. The harmonised surveys also collected information on well-being, self-perceptions, non-cognitive skills, parental beliefs and behaviour. The researchers ran multiple robustness checks, including alternative model specifications and adjustments for multiple hypothesis testing.
Phone calls produced the more consistent learning gains
Pooled across all five countries, children receiving phone tutorials plus SMS improved by 0.321 standard deviations relative to controls, with p < 0.001. By comparison, the pooled SMS-only effect was 0.078 standard deviations, p = 0.006.
The difference is not merely statistical. The researchers compare the phone effect with a recent review in which the median effective education intervention produced a gain of about 0.10 standard deviations. On that benchmark, the average phone effect was roughly three times as large.
SMS-only support was more uneven. Detectable statistically significant effects appeared in Uganda and the Philippines, but not consistently across every country. That suggests that sending educational content can help under some circumstances, particularly where alternatives are extremely limited, but a live instructional interaction appears more dependable across settings.
The programme also changed concrete skills. Across the pooled analysis, phone and SMS increased the share of students correctly solving division problems by 13.5 percentage points, p < 0.001. The control mean was 14.6%, so this represented a 92% relative increase. The intervention also raised performance on fractions by 6.4 percentage points even though fractions were not directly taught, suggesting that stronger foundational numeracy may have supported more advanced competencies.
The largest gains appeared where disruption was most severe
The average effect conceals substantial variation. Uganda and the Philippines produced especially large gains, and both had experienced close to two years of school disruption.
In Uganda, students receiving the phone intervention gained 1.193 additional operation levels, p < 0.001. They were also 34.4 percentage points more likely to reach the highest proficiency level measured, p < 0.001. In the Philippines, phone tutorials produced a 0.56-level gain and a 13 percentage-point increase in reaching the highest operation level, both with p < 0.001.
The Ugandan learning trajectory is particularly striking. Among grade 4 children in the control group, about 17% could perform division at baseline, falling to 10% at endline during the prolonged disruption. In the treatment group, 48% could divide at endline. The intervention therefore did more than slow deterioration in that setting. It reversed the observed learning loss and pushed performance beyond the baseline level.
These country differences should not be interpreted as a clean test of why the intervention worked better in one place than another. Multiple conditions changed simultaneously. The researchers therefore describe their mechanism analysis cautiously. Still, two patterns stood out: stronger effects tended to occur where educational alternatives were weakest and where implementation became better at matching instruction to children’s actual learning levels.
Targeting improved as the programme learned how to scale
Scaling successful social programmes often produces a so-called voltage drop, where results weaken as implementation expands beyond the original pilot. This study found a different pattern.
Accuracy in targeting instruction to children’s learning levels rose from 50.9% in Nepal to an average of 81.5% in Uganda. The trials occurred sequentially, and the intervention’s effectiveness also tended to rise in later studies. That does not prove that improved targeting caused the larger effects, because the country contexts differed. It does, however, support the argument that implementation quality matters at least as much as the communication technology.
Engagement remained high as well. Weekly participation ranged from about 70% to 80% across sites. For a programme operating during disruption and in relatively low-resource settings, that level of continued contact is important. A technically impressive platform offers little educational value if households cannot or do not use it.
Government teachers achieved sizeable effects too
Nepal and the Philippines provided another critical test. In these countries, the researchers randomized delivery through government teachers and NGO instructors.
Across the two settings, government teachers improved learning by 0.40 operation levels, while NGO instructors improved it by 0.30 levels. Both effects were statistically significant at p < 0.001.
This result addresses a recurring policy problem. An intervention may work when delivered by a specialised research or nonprofit team but fail when transferred into a large public bureaucracy. The present trials do not show that implementation will succeed in every education system, but they demonstrate that the phone tutoring model did not depend on NGO delivery to produce learning gains in the settings tested.
There were also suggestive spillovers for teachers. In Nepal, government teachers randomized to deliver tutoring became 9.3 percentage points more likely to target feedback to students’ learning levels, p < 0.05. They were also 15.8 percentage points more likely to say they would want to be a teacher again if making the choice anew. The authors caution that these teacher outcomes were self-reported, but they raise the possibility that practising targeted instruction remotely could influence later teaching behaviour.
The economics may be as important as the effect size
Emergency education programmes must often be deployed quickly and under tight budgets. The average estimated cost of the combined phone and SMS intervention was about US$11 per child. Because the model relied on devices families already possessed, the main costs involved content, training, monitoring and airtime rather than purchasing new hardware.
Using learning-adjusted years of schooling, a metric intended to combine the quantity and quality of education, the researchers estimated that the programme generated 3.6 learning-adjusted years of schooling per US$100 spent. That placed it among the top 10 of 150 education interventions in the comparison used by the study.
The authors also describe the result as approximately four years of high-quality instruction per US$100 when rounded for communication. Cost-effectiveness estimates always depend on assumptions about how learning gains translate into equivalent schooling, but the underlying point is clear: the programme produced comparatively large gains using a low-cost delivery channel.
What the study does not establish
Five randomized trials across five countries provide stronger evidence than a single pilot, but they do not make the programme universally transferable. Every trial occurred in the context of pandemic-era disruption, and the magnitude of benefits depended partly on what educational support children would otherwise have received.
That matters for interpretation. The largest gains occurred in Uganda and the Philippines, where disruption was particularly severe and alternative instruction was limited. A phone tutorial added to a functioning full-time school system may produce a different effect from the same tutorial delivered when the alternative is little or no schooling.
The mechanism analysis is also exploratory. Targeting accuracy improved alongside effect sizes, but countries differed on many dimensions at once. The study cannot isolate improved targeting as the sole reason later trials performed better.
Some secondary outcomes, including teacher attitudes and practices, were self-reported and should therefore be treated more cautiously than the randomized child learning outcomes. The intervention also focused heavily on foundational numeracy. Whether similarly structured phone tutoring produces equally strong effects for literacy, secondary-school subjects or more complex curricula requires additional evidence.
Finally, access to a phone is not the same as equal access to tutoring. Households may differ in device availability, network reliability, language, caregiver support and the times at which a shared phone is available. The trials show that the model can reach large numbers of learners in low-resource settings, not that access barriers disappear.
Resilient education may depend on simpler technology used well
The most useful lesson from the study is not that phones outperform every other educational technology. It is that resilience depends on matching the delivery system to the conditions in which learning has to continue.
During a disruption, a low-bandwidth technology already present in most households can have advantages that disappear in a conventional technology comparison. It requires little new infrastructure, can be deployed quickly and allows a teacher to adapt instruction in real time.
But the contrast between SMS and live calls shows why access alone is not enough. Text messages were cheap and sometimes effective, yet their results were smaller and less consistent. The stronger intervention combined accessibility with human interaction, assessment and targeted teaching.
For governments preparing for future school closures caused by climate events, disease outbreaks or other shocks, that combination is consequential. Resilience does not necessarily require recreating an entire classroom online. In these five trials, a short weekly call built around the child’s actual learning level was enough to produce sizeable and repeatable gains.
Source Information
Study: Mobile education builds resilience during shocks in five countries
Authors: Noam Angrist, Micheal Ainomugisha, Sai Pramod Bathena, Peter Bergman, Colin Crossley, Claire Cullen, Thato Letsomo, Moitshepi Matsheng, Rene Marlon Panti, Shwetlena Sabarwal and Tim Sullivan
Journal: Nature
Published: 9 September 2026
DOI: 10.1038/s41586-026-10990-x
