Timely data on educational attainment at granular geographic levels remains scarce in many countries, limiting evidence-based policy-making. Recent advances in machine learning have enabled the use of non-traditional data sources like satellite imagery and mobile phone records to measure development indicators. While these approaches have been successful in predicting outcomes such as wealth, poverty, or population density, previous attempts to predict educational attainment have achieved only modest accuracy. A key challenge is finding data sources that directly reflect human capital rather than its economic consequences. Here we show that language patterns and user behavior in social media can explain up to 70 percent of the variance in regional educational attainment. Our machine learning framework leverages linguistic features, user behavior, and network characteristics from 25 million geolocated tweets from the United States and Mexico. It performs particularly well in predicting higher education levels and maintains a good performance even with limited data collection periods. These results show that digital communication patterns can serve as reliable proxies for human capital. In light of the rapid expansion of social media use around the globe, this represents a promising approach to tracking educational outcomes in regions lacking granular and timely survey data.

Participatory teaching methods have been shown to be more successful than traditional rote learning in high-income countries. It is, however, less clear if they can help address the learning crisis in low- and middle-income countries, where classes tend to be large and teachers have fewer resources at their disposal. Based on a field experiment with 440 teachers from 220 schools in Tanzania, we use official standardized student examinations to assess the impact of a pedagogy-centered intervention. A five-day in-service teacher training on participatory and practice-based methods improved students test scores 18 months later by 0.15σ. The additional provision of laptops with a learning software allowing teachers to refresh their content knowledge did not yield further learning gains for students. Complementary results from qualitative surveys and interviews suggest that the program was highly appreciated by different stakeholders, but that participants are unable to assess its impact along different dimensions, giving equally positive evaluations of its successful and its less successful elements.

Good teachers are the backbone of a successful education system. Yet, in developing countries, teachers' content knowledge is often inadequate. This study documents that primary school math teachers in the department of Morazán in El Salvador only master 47 percent of the curriculum they teach. In a randomized controlled trial with 175 teachers, we further evaluate a computer-assisted learning (CAL) approach to address this shortcoming. After a five months in-service training combining CAL-based self-studying with monthly workshops, participating teachers outperformed their peers from the control group by 0.29σ, but this effect depreciated by 72 percent within one year. Our simulations show that the program is unlikely to be as cost-effective as CAL interventions directly targeting students.

This study provides novel evidence on the relative effectiveness of computer-assisted learning (CAL) software and traditional teaching. Based on a randomized controlled trial in Salvadoran primary schools, we evaluate three interventions that aim to improve learning outcomes in mathematics: (i) teacher-led classes, (ii) CAL classes monitored by a technical supervisor, and (iii) CAL classes instructed by a teacher. We find that CAL lessons lead to larger improvements in students' mathematics skills than traditional teacher-centered classes. In addition, teachers add little to the effectiveness of learning software. Our results highlight the value of CAL approaches in an environment with poorly qualified teachers.

Previous research on educational aspirations and educational decision-making has mostly focused on high-income countries and thus on a relatively homogeneous socio-economic context. However, educational decision-making may be sensitive to contextual factors such as economic deprivation, a dysfunctional welfare state or poor access to credit markets – characteristics shared by most low- and middle-income countries. To better understand how economically disadvantaged individuals in developing countries make their educational choices, we conducted a survey based on a random sample with high school students in the rural department Morazán in El Salvador, a lower middle-income country in Latin America. Our results show that regardless of the social background, almost all students aspire to pursue tertiary education, probably due to the high tertiary degree premium in earnings and the high social benefits. However, the lack of possibilities to finance their studies generally prevents the realisation of these aspirations for lower social background students. While in high-income countries, cost factors are not very important in the decision-making process, the burden of costs explains around 45 percent of the social background effect in El Salvador.

Education is one of the key resources in the fight against poverty. While substantial progress has been made in terms of school enrollment, evidence suggests that educational quality is still alarmingly low in many developing countries. Various explanations have been suggested, but one very obvious factor in the educational production function has received surprisingly little attention: the content knowledge of teachers. For this study, we administered an exam-type assessment to a representative sample of 224 primary school teachers in Morazán, El Salvador. The average teacher scored 47% correct answers on 50 questions covering the official math curriculum for second to sixth graders. Overall, our results point to an even more worrying situation than suggested by previous findings based on indirect measures of content-related teacher skills in several African countries.

Despite billions spent annually on teacher training, rigorous evidence on standalone in-service programs remains scarce, as most evaluated programs bundle training with curriculum or material reforms. We address this gap through a large-scale randomized controlled trial with 338 schools and over 6,000 students in El Salvador. Teachers are randomly assigned to either a control group or one of three training programs focusing on (i) content knowledge, (ii) pedagogical knowledge, or (iii) a combination of both inputs. We find lasting effects on teacher content and pedagogical knowledge of up to 0.3σ and 0.5σ respectively one year after program end. Yet, this only changes teachers' classroom practices in the short-run and does not translate into significant student learnings. The data most closely aligns with a setting where teachers face a dual challenge: introducing new ideas in a rigid environment while navigating the significant learning gaps present among students in later grades.