This blog is coauthored with Marina Bassi and Ana Reynoso
Preschool education can be a powerful equalizer. The first five years of life is the period when the brain develops most rapidly and when nutrition, stimulation, and learning experiences shape the child’s readiness for a future trajectory. Yet in many low-income countries, children grow up in environments marked by deprivation—malnutrition, limited parental literacy, and a near absence of early learning opportunities. In such settings, preschool can do more than prepare children for school; it can offset the disadvantages of poverty by providing a structured, nurturing space for early stimulation, play, and social interaction.
When we began evaluating Mozambique’s Early Childhood Development Project (DICIPE), only 3.5% of children aged 3–5 was enrolled in preschool. Poverty was widespread, infrastructure scarce, and few rural families had ever seen an early learning center. Against this backdrop, the Government of Mozambique—with support from the World Bank and the Global Partnership for Education—launched DICIPE to build and staff community preschools (“escolinhas”) in deprived rural areas. Each escolinha was located near a primary school, served by locally hired instructors, and equipped with low-cost, recycled learning materials. The program also organized caregiver sessions on parenting, nutrition, and early stimulation. Between 2017 and 2019, 350 preschools were built across five provinces.
In our paper, we present the results from the impact evaluation of the DICIPE program. The evaluation covered 218 rural communities, randomly assigning 110 to receive preschool infrastructure and services and 108 to serve as controls. Baseline data revealed extreme deprivation: 81% of caregivers were illiterate, nearly all children were stunted, and only 2% had attended preschool before the intervention.
Three years later, DICIPE had transformed the primary education outcomes for the children in the treated communities. Preschool enrollment reached 75% from a baseline near zero (no preschool service was available in those areas), one of the largest increases ever documented in a preschool experimental evaluation. Children in treated communities were 6 percentage points (pp) more likely to enroll in primary school, 3 pp less likely to repeat a grade, and almost 6 pp more likely to be in the appropriate grade for their age. The program also improved children’s readiness to learn: assessments showed higher scores in early literacy, numeracy, fine motor skills, and interest in reading, resulting in a 0.16 standard deviation increase in school-readiness skills. These effects are comparable to those achieved by high-quality preschool programs in middle-income countries, but in a context of far greater vulnerability.
The impacts extended beyond target children. Parents in treated communities were more likely to meet with their child’s school principal and provided more supportive home environments. Younger siblings were 62 pp more likely to attend preschool, showing that new infrastructure benefited subsequent cohorts. A cost–benefit analysis revealed strong economic returns, with conservative benefit-to-cost ratios between 6.4 and 33.8.
These results, discussed in more detail in our VoxDev blog post, confirmed that preschool construction in extremely poor settings can yield durable human-capital gains. In this blog, we focus on another question that we explored in the paper: why did it work so well?
From “does it work?” to “how does it work?”
Knowing that a program works is essential but understanding how it works is key to help policymakers replicate success. In DICIPE, two components were implemented in parallel: the construction of preschools and the implementation of parenting sessions. Both could plausibly influence child outcomes but through different mechanisms. We wanted to know which of these channels mattered most for later schooling.
To do so, we turned to mediation analysis—a statistical approach that decomposes the total impact of a program into the pathways through which it operates. Conducting this analysis became a learning process in itself, revealing both the promise and the practical challenges of applying such methods in large-scale randomized evaluations. This blog focuses on what we discovered—about the program and about the process of uncovering the mechanisms behind it.
Why mediation analysis?
Most impact evaluations estimate total effects: the average difference between treatment and control groups. Yet many programs, especially those with multiple components, operate through several intertwined channels. In DICIPE, the program could affect children’s outcomes in primary school by providing access to preschool services (increasing preschool enrollment) or by changing how caregivers support children at home. Mediation analysis helps us ask a sharper question: how much of the total impact on children’s primary school enrollment and progression occurred because children attended preschool, and how much occurred through other channels, such as improved parenting practices?
This distinction is not merely academic. For governments deciding where to allocate scarce resources, whether to build more preschools, strengthen teacher training, or expand parenting programs, knowing which channel drives impact can directly inform the next round of investment.
To conduct the mediation analysis, we followed a 2-steps process that we describe below.
Step 1: Building the framework
We began by mapping the program’s logic chain. As mentioned before, the intervention created two main potential mediators: children’s preschool enrollment and caregivers’ participation in parenting education sessions. These, in turn, could influence outcomes such as primary school enrollment, grade progression, and overall academic readiness. Our central hypothesis was straightforward: if preschool enrollment explains a large share of the impact on later schooling, then the act of attending preschool—not simply being offered the opportunity—is a critical driver of success.
Step 2: Estimating indirect and direct effects
We applied the framework developed by Attanasio and coauthors (2020, 2022), combining experimental variation from the randomized design with mediation models. We first estimated the total effect of DICIPE on each primary-school outcome. Then we examined how strongly the program affected each mediator, i.e., whether it substantially raised preschool enrollment and whether it increased attendance at parenting sessions. Finally, we assessed how each mediator related to the outcome, controlling for treatment assignment.
This allowed us to decompose the total effect into an indirect effect (the portion explained by each mediator) and a direct effect (the remaining impact operating through other channels). Because preschool enrollment increased dramatically in treated communities, we expected it to play a central role in explaining downstream results.
The results were clear. Preschool enrollment emerged as the dominant mediator linking the DICIPE program to children’s later schooling success. Roughly 60% of the total effect of the intervention on primary school enrollment and grade progression operated through this channel. The direct effect—capturing other pathways such as spillovers, improved parental motivation, or community engagement—was smaller and often statistically insignificant.
In contrast, participation in parenting education sessions did not significantly mediate outcomes, possibly reflecting the limited frequency of meetings and the incomplete implementation of that component. These findings confirm that, in this context, the opportunity to attend preschool itself was the decisive mechanism behind the program’s success.
Understanding this helps reframe the narrative. In many middle-income countries, where preschool coverage is already high, policy debates revolve around improving quality. In Mozambique’s rural areas, by contrast, simply building preschools and ensuring attendance was sufficient to generate substantial improvements in educational trajectories. When the starting point is near zero, access is the mechanism.
Lessons we learned from doing mediation analysis in the field
The process of conducting this analysis offered valuable lessons. Measuring mediators in large-scale field settings requires solid measurement and persistence. Tracking preschool enrollment across remote villages involved cross-checking survey responses with administrative records and community logs to ensure accuracy. The parenting-education component, meanwhile, revealed an important practical insight: when sessions are irregular or weakly documented, their causal role becomes difficult to measure and, more importantly, presumably less consequential for outcomes.
Another lesson was statistical. Even in a well-powered randomized control trial, mediation analysis can have lower precision because it depends on variation in the mediator rather than in the treatment. Carefully pre-specifying the mediators and grounding them in theory helped maintain clarity and interpretability. Conducting this exercise within a government-led, large-scale program reminded us that rigorous causal tools can—and should—be applied beyond small-scale pilots.
What policy implications can we derive from this mediation analysis?
The findings from Mozambique carry clear policy lessons for low-income countries expanding early childhood education. When baseline preschool coverage is close to zero, physical access is not just necessary—it is transformative. Governments should prioritize the construction and staffing of preschools in the most deprived areas, ensuring that every community has a functional, accessible center. Enrollment data should be systematically tracked, as it is a leading indicator of program effectiveness rather than a secondary compliance measure.
At the same time, while the parenting-education component showed limited impact here, it may remain a promising complement if delivered effectively. Future programs could explore lower-cost but higher-frequency formats—such as radio sessions, SMS nudges (see, for example, a parenting program in Jamaica), or structured community groups—to strengthen home learning environments.
Finally, integrating mediation analysis into the evaluation of bundled interventions provides researchers the possibility of better understanding the mechanisms affecting the program’s outcomes. More importantly, it offers governments a way to allocate funds to the levers that yield the greatest returns. For early childhood policy, this means identifying where investments—whether in access, quality, or parental support—produce the highest social payoff.
Join the Conversation