Each year, around 6 million children in West and Central Africa do not receive any lifesaving vaccines, making them ‘zero-dose children.’ To find and serve these children—who are often living in remote, inaccessible areas—UNICEF’s Reach the Unreached (RtU) initiative equips countries with innovative datasets and actionable geospatial tools, leveraging frontier data techniques such as probability models and AI to produce geolocated vaccination estimates.
However, incorporating probability models and non-traditional data sources bring new challenges, such as potential algorithmic biases and inconsistencies. A recent pilot has provided insights to better interpret and address these issues in vaccination coverage estimates.
The challenge: finding zero-dose children
When immunization administrative data are incomplete, immunization planning and targeting rely on national estimates, such as the WHO/UNICEF Estimates of National Immunization Coverage (WUENIC) data, and population projections based on the most recent census. However, national estimates cannot represent subnational realities, and when the latest census is outdated, projections can distort reality. This sometimes leads to underestimated population counts and vaccination rates exceeding 100 per cent.
While strengthening civil registration and health information systems is critical to enhance the reliability of administrative data, data from household surveys can help fill gaps in the interim, while also allowing for further analysis of inequalities in coverage. To address these gaps, RtU was launched in five pilot countries (Cameroon, Chad, Côte d’Ivoire, Guinea and Mali), enhancing survey data through machine learning and geostatistical models to better characterize geographic inequities. This approach has helped locate over 1.3 million zero-dose children, with efforts ongoing to reach even more.
Comparing estimation models
Recognizing the potential biases posed by these innovative methodologies, RtU team members engaged Frontier Data Network (FDN) researchers and data scientists to help compare different estimation models. Together, they co-defined the problem and secured additional expertise through a partnership with the Massachusetts Institute of Technology, evaluating model designs to ensure RtU developed the most reliable country-specific models. They also analysed how zero-dose prevalence correlates with other indicators of child rights deprivation and explored ways to improve data aggregation at the health facility level. Using WUENIC as a benchmark, they confirmed that estimates from the probabilistic models aligned with official country-reported figures.
Future implications
Complementing UNICEF’s support to governments in strengthening civil registration and administrative data systems, the RtU initiative integrates frontier datasets to provide additional information and enhance vaccination coverage estimates for planning, targeting, and service delivery. This model comparison work highlights the need for rigorous validation of probabilistic models to ensure countries use the best possible approaches to identify zero-dose children.
By thoroughly reviewing different model designs, benchmarking against WUENIC and assessing alignment with child rights indicators, RtU provides a replicable framework for improving data-driven decision-making. Refining these methods can help ensure every child—no matter how remote their community—receives lifesaving vaccines and has their fundamental rights upheld.
