Adapting for scale: Experimental evidence on computer-aided instruction in India (with Karthik Muralidharan)
Latest version (Oct 2025); NBER Working paper Explainer threads on Twitter(X) and Bluesky
Revisions requested at the Journal of Political Economy
Many interventions that “work” in small-scale trials often fail when scaled. This highlights the need to adapt promising interventions for scalability by addressing constraints that bind at larger scales. We do so in the context of a personalized adaptive learning (PAL) software that was highly effective in a small-scale trial. We adapt the PAL implementation for scalability by integrating it into public school schedules, and experimentally evaluate this adaptation in a more representative sample over 20 times larger than the original study. After 18 months, treated students scored 0.22σ higher in Mathematics and 0.20σ higher in Hindi, a 50–66% productivity increase over the control group. Learning gains were proportional to student time on the platform, providing a simple, low-cost metric for monitoring implementation quality in future scale-ups. The adaptation and its experimental validation have informed scale-ups now reaching over 250,000 students.
Improving Public Sector Management at Scale: Experimental Evidence on School Governance in India (with Karthik Muralidharan)
Latest version, Revisions requested at Journal of Political Economy: Microeconomics (JPE:Micro)
NBER Working Paper (Nov 2020); (Slides)
We present results from a large-scale experimental evaluation of an ambitious attempt to improve management quality in Indian schools (implemented in 1,774 randomly-selected schools). The intervention featured several global “best practices” including comprehensive assessments, detailed school ratings, and customized school improvement plans. It did not, however, change accountability or incentives. We find that the assessments were near-universally completed, and that the ratings were informative, but the intervention had no impact on either school functioning or student outcomes. Yet, the program was scaled up to cover over 600,000 schools nationally. We find using a matched-pair design that the scaled-up program continued to be ineffective at improving student learning in the state we study. We also conduct detailed qualitative interviews with frontline officials and find that the main impact of the program on the ground was to increase required reporting and paperwork. Our results illustrate how ostensibly well-designed programs, that appear effective based on administrative measures of compliance, may be ineffective in practice
Using technology to deliver preschool services at scale: Experimental evidence from India (with Ajinkya Keskar, Mauricio Romero and Karthik Muralidharan) Draft coming soon
We experimentally evaluate two iterations of a large-scale, technology-aided home stimulation program to enhance parent-child interactions and improve preschool quality for children aged 3 to 5. The base intervention, which provided structured learning activities via WhatsApp groups, successfully engaged parents but did not improve children's math and language outcomes. However, an enhanced version with intensified engagement—featuring more frequent, personalized messaging, and structured preschool worker support— increased parental participation beyond the base intervention and improved children’s learning outcomes by 0.12-0.2 s.d. At a total implementation cost of $1.17 (97.11 rupees) per child per year, our findings demonstrate the efficacy of a highly cost-effective and scalable intervention for improving early childhood learning in low- and middle-income countries.
Childhood Learning and Adult Earnings in Four LMIC Cohorts (with Jishnu Das, Yash Dhuldoya, Marta Favara, Amory Gethin and Alan Sanchez)
Researchers and policymakers in low- and middle-income countries (LMICs) have customarily relied on indirect evidence to justify investments in learning, by either linking test scores to economic growth or connecting adult cognitive assessments to earnings. The reliance on indirect evidence is largely because longitudinal data linking school-age learning to adult earnings are scarce. We address this gap by using data from Ethiopia, India, Pakistan, and Peru to estimate associations between foundational learning (FL), measured by test scores around age 12, and human capital and earnings in young adulthood. In Peru, India, and Pakistan, a one-standard-deviation increase in FL is associated with an 8--13 log point increase in adult earnings. Childhood scores also predict 1.1--1.5 additional years of schooling and 9.3--13.5 percentage points higher college or vocational attendance, as well as higher skills in adulthood. Interestingly, a linear-projection decomposition indicates that measured schooling and adult test scores span from 18\% in Peru to 50--57\% of the FL-earnings gradient in India and Pakistan, suggesting that FL may raise earnings over and above the association with schooling and adult scores. Finally, FL can account for about 8-22 percent of the earnings gap between young Ethiopians and Indians in our sample, a number that could rise to 50 percent as returns to learning grow over the life cycle. We interpret these descriptive findings in light of recent causal evidence on the learning-to-earnings pipeline and discuss the implications for policy investments in FL.