Mridul Joshi

This study evaluates a digital empowerment intervention designed to improve how college students in India engage with smartphones and social media. Using a randomized controlled trial, we test both a high-touch classroom course and a low-touch text message version. The intervention addresses self-control problems and habit formation, introduces commitment devices, and builds skills for identifying misinformation. We combine survey data, administrative records, and high-frequency smartphone tracking to estimate effects on mental health, academic outcomes, and digital behavior.

This study examines how faculty gender influences both academic and psychological outcomes for STEM students. Leveraging a rare setting in which engineering students in India are randomly assigned to instructors, we estimate the causal effects of exposure to female faculty across multiple undergraduate institutions. The design combines administrative data, standardized assessments, and surveys measuring STEM-related confidence, anxiety, and beliefs about gender and ability—offering new evidence on how representation and identity shape student experiences in higher education.

This study examines the impact of active learning on STEM achievement through a systematic review and new empirical analyses. We assess the methodological quality of existing research and combine this with original data from a large-scale student–instructor random assignment in India and nationally representative samples from China, India, and Russia. Anchored in competing theoretical perspectives from cognitive science, the study investigates when, how, and for whom active learning may, or may not, impact learning outcomes in higher education.

This study tests whether assigning more advanced reading material to stronger third-grade students can improve literacy outcomes in low-cost private schools in Nigeria. In a randomized trial across 60 schools, the intervention modified a structured, tablet-based reading program to better align reading difficulty with student ability. The study combines experimental results with a formal model of learning that incorporates non-convex losses from mistargeted instruction—where learning declines sharply once material exceeds a student’s zone of proximal development. This approach challenges standard assumptions in instructional targeting and offers new insight into the limits of personalized learning at scale.

This paper introduces a framework for estimating heterogeneous treatment effects when the outcome is measured using a psychometric scale, like a test or survey, by leveraging item response theory to detect variation at the item level. Standard approaches ignore the possibility that different items may respond differently to treatment, which can distort findings and underestimate uncertainty. By analyzing over 5 million item responses across 75 datasets, we show that modeling “item-level” heterogeneity not only yields more accurate standard errors and treatment effect estimates but also resolves key identification problems and improves generalizability to new settings.

This paper shows that identical patterns of treatment effect heterogeneity can arise either from differences in student ability or from how specific test items respond to an intervention. Using item response theory within a causal framework, we demonstrate that these sources of variation are indistinguishable from total scores alone. We introduce a model that uses item-level data to separate person- and item-dependent effects and apply it to a reading intervention to illustrate how this distinction matters for interpretation.