Heterogeneous Effects of Pre-primary Education on Child Development Outcomes in Bangladesh: Using Project Data and Causal Forest Analysis

• In Event: Strengthening ECD Monitoring: Psychometrics, Population Surveys, and Integrated Data Platforms

Proposal

Pre-primary education (PPE) is increasingly recognized as a critical foundation for children’s cognitive, social, and emotional development, laying the groundwork for lifelong learning and future labor market success. A robust body of evidence, particularly from high-income countries, has documented the long-term benefits of high-quality PPE. Classic intervention studies such as the Perry Preschool Project in the United States and the nationwide PPE expansion in Azerbaijan illustrate that PPE is one of the most cost-effective public investments, producing substantial returns in education, health, and social outcomes. Importantly, these studies highlight that children from disadvantaged backgrounds often benefit disproportionately, thereby making PPE an essential instrument for promoting equity.

Among low- and middle-income countries (LMICs), Bangladesh stands out as a country that has made notable strides toward expanding PPE access through ambitious policy reforms. The Government of Bangladesh has progressively integrated PPE into its broader education sector strategies, moving toward a model of universal, tuition-free access. This effort has been facilitated by development partners such as the World Bank, Asian Development Bank (ADB), UNICEF, UNESCO, JICA, and others, who have adopted a wide sector approach to support education reforms through successive phases of the Primary Education Development Program (PEDP II, PEDP III, and PEDP IV). These programs provided coordinated support for the development of textbooks, teacher training and salaries, curriculum reforms, and institutional guidelines. In particular, the commitment to provide one year of free PPE for all four- to five-year-olds reflects Bangladesh’s recognition of early education as a national priority.

While policy momentum has accelerated, empirical evidence on the causal effects of PPE in LMICs remains scarce. Unlike high-income contexts where experimental and quasi-experimental studies abound, research in LMICs has largely relied on observational data and correlational analyses. For example, Pimenta (2023) examined multi-country data to assess associations between PPE attendance and child outcomes, documenting heterogeneous effects moderated by family wealth, parental education, and rural–urban divides. However, without causal inference methods, these studies cannot disentangle the true impact of PPE from confounding factors such as parental motivation or community characteristics. This gap in the literature leaves policymakers in LMICs with limited evidence on whether and how PPE can reduce inequalities in early child development.

Study objective and contribution
This study seeks to address this critical research gap by leveraging data from a randomized controlled trial (RCT) of the Early Years Preschool Program (EYPP) conducted in Meherpur district, Bangladesh. The EYPP was jointly implemented by the World Bank, the American Institutes for Research (AIR), and Save the Children as a pilot initiative designed to test the feasibility and impact of offering an additional year of preschool prior to the nationwide rollout of free one-year PPE. The program emphasized play-based, child-centered pedagogy, and included revised curricula and teacher training. By analyzing RCT data, this study contributes robust causal evidence on the heterogeneous impacts of PPE in a low-income setting, going beyond average treatment effects to explore which groups of children benefit most.

The study uses panel data from approximately 1,800 children, with a notably low attrition rate across the study period, thereby preserving statistical power and internal validity. To estimate the impacts of an additional year of PPE, we first calculate the average treatment effect of EYPP participation across key developmental domains: literacy, numeracy, social-emotional skills, and motor skills. Beyond this, we employ machine learning methods—specifically the causal forest approach—to estimate Conditional Average Treatment Effects (CATEs).

The causal forest approach offers several advantages over traditional econometric methods. Conventional approaches, such as subgroup analysis, difference-in-difference-in-differences (DDD), or quantile regression, often require researchers to specify interaction terms ex ante, limiting flexibility when exploring complex, multidimensional heterogeneity. In contrast, causal forests leverage recursive partitioning and ensemble methods to automatically detect heterogeneous treatment effects across high-dimensional covariate spaces. This allows us to more accurately estimate how child- and family-level characteristics—such as socioeconomic status, home learning environment, or baseline skills—condition the effectiveness of PPE.

Results reveal consistent evidence that disadvantaged children derive the greatest benefits from participation in EYPP. Across all developmental domains, children from poorer households and those with fewer home learning resources exhibited larger gains in literacy, numeracy, social-emotional, and motor outcomes compared to their more advantaged peers. This pattern aligns with international evidence that PPE can serve as an equalizing force when implemented with quality standards.

More granular insights emerge from the CATE estimation based on children’s baseline skill levels. Children with mid- to low-level baseline cognitive skills (−1 < SD < 0) experienced the largest gains in both literacy and numeracy domains. Conversely, children with extremely low baseline skills (SD < −1) benefited less, potentially reflecting challenges in bridging wide developmental gaps within a one-year timeframe. Similar patterns were observed in social-emotional and motor development, where children with lower starting skills nonetheless gained more than those at higher baselines. Taken together, these results suggest that PPE, when delivered with sufficient quality and teacher preparation, holds significant potential to mitigate early developmental disparities in Bangladesh.

The interpretation of these findings, however, requires caution with respect to external validity. The EYPP pilot was implemented under conditions of enhanced teacher training and curriculum revision, emphasizing child-centered, play-based pedagogy. This differs markedly from the prevailing mode of PPE delivery in Bangladesh, where classrooms often remain teacher-centered and academically oriented. Consequently, the impacts observed in Meherpur may not be fully replicable in nationwide settings unless comparable quality standards are scaled up.

Furthermore, while the causal forest approach provides nuanced insights into treatment heterogeneity, the method is not immune to issues such as overfitting or sensitivity to hyperparameter tuning. Robustness checks, including cross-validation and comparison with more traditional subgroup analyses, were conducted to ensure reliability of estimates. Nonetheless, future research could complement this study by examining long-term outcomes and assessing how the impacts of PPE persist as children progress through primary school.

The findings carry important implications for policymakers in Bangladesh and other LMICs. First, they underscore the value of investing in quality PPE not merely as a universal entitlement, but as a strategic tool for reducing developmental inequalities. The evidence that children from disadvantaged backgrounds gain disproportionately suggests that PPE can play a central role in advancing equity goals and achieving Sustainable Development Goal (SDG) 4.2, which calls for universal access to quality early childhood development, care, and pre-primary education.

Second, the results highlight the need to ensure that scaling up PPE does not come at the expense of quality. Teacher preparation, curriculum reform, and monitoring of classroom practices must accompany expansion efforts to replicate the gains observed in the EYPP. Without such safeguards, PPE risks becoming a missed opportunity, particularly for the very children who stand to benefit most.

Finally, the study contributes methodologically by demonstrating the utility of machine learning approaches in education research. By applying causal forests, we provide richer insights into treatment heterogeneity than would have been possible with conventional econometric techniques, thereby advancing the methodological toolkit available for evaluating complex education interventions.

In sum, this study offers new causal evidence on the heterogeneous effects of PPE in Bangladesh, showing that disadvantaged children and those with mid- to low-level baseline skills benefit the most from participation in a well-prepared preschool program. While external validity remains a challenge, the findings provide a compelling case for integrating equity considerations into PPE policy design and underscore the importance of ensuring quality alongside access. More broadly, the study demonstrates the promise of combining experimental data with cutting-edge machine learning methods to generate actionable insights for education policy in LMICs.

Author

• Kohei Uno, Kobe University