Bayesian reconstruction of under-five small area populations: correcting for undercount bias and source uncertainty

The National Institute of Child Health and Human Development awarded Emory University $415,711 on August 26, 2026, under the Child Health and Human Development Extramural Research program (CFDA 93.865) to develop a Bayesian hierarchical estimation framework for reconstructing under-five population counts at the small-area level, correcting for persistent undercount bias and multi-source uncertainty in Census data.

The research addresses systematic undercounting of children under age five in the decennial Census—5.4 percent in 2020, representing approximately 1.3 million missed young children. Current Census adjustment methods relying on demographic analysis, post-enumeration surveys, and population estimates program projections lack mechanisms to account for bias and uncertainty in component inputs (births, deaths, net migration), rendering them unreliable in small or data-sparse counties where undercount variation is most pronounced and poorly characterized. The proposed work develops tailored methodologies capturing both systematic patterns of undercount and inherent data uncertainty from administrative and survey sources, with emphasis on small counties and rural areas where undercount has been least evaluated. Accurate granular enumeration at small-area levels is necessary for effective targeting of public health funding, resource planning, and monitoring of under-five disease and mortality trends.

Emory University Office of Sponsored Programs is the recipient and prime performer, with place of performance in Atlanta, Georgia. The period of performance runs through July 31, 2028. The assistance type is Project Grant.

Abstract

Accurate and granular enumeration of numbers of young children under age five (U5) is critical to effectively target public health funding, resource planning, and accurate monitoring of trends in U5 disease and mortality. However, the united states census bureau (uscb) decennial census data has persistently faced challenges in accurately counting the U5 population, and for decades has systematically undercounted this population. Incomplete enumeration of the U5 population in the decennial census has severe consequences for federal funding of various programs that serve children's needs. In the 2020 census, children under age five had the highest net undercount of any age group at 5.4%, or approximately 1.3 million missed young children. Recent research underscores that this undercount is not random but deeply patterned by subpopulations, household structure, and geography. A severe limitation of current findings is the lack of evaluation of how U5 undercount varies in small counties and rural areas. Additionally, previous efforts to identify U5 undercounts have compared census counts to benchmarks such as demographic analysis (DA), the post-enumeration survey (PES), and population estimates program (PEP) projections. While descriptively useful, these approaches do not adjust for bias or uncertainty in U5 inputs-births, deaths, and net migrations-limiting their reliability in small or data-sparse counties where undercount and multi-source error are most pronounced. To produce accurate and robust small-area U5 population estimates, IT is essential to develop rigorous, tailored methodologies that capture both the systematic patterns of U5 undercount and the uncertainty inherent in administrative and survey-based data sources. Our proposed approach develops a flexible bayesian hierarchical U5 estimation framework to produce small-area estimates of the U5 population that are robust to source-specific bias and data sparsity. The proposed model fuses multiple uscb data sources-including the decennial census, PEP, and american community survey (ACS) microdata with vital statistics on births and deaths, using a data fusion likelihood that accounts for definitional inconsistencies, measurement error, and underreporting. We incorporate a novel age-specific net migration sub-model that improves estimation of noisy U5 net migration by borrowing strength from older, more reliably measured age-groups. We will pursue two specific aims. Aim 1 will develop the full bayesian hierarchical U5 estimation framework to obtain posterior distributions of bias-adjusted county-level U5 populations and associated uncertainties. Aim 2 will evaluate model accuracy, utility, and scalability using simulation studies under varying levels of undercount and data quality. This proof-of-concept study will be conducted in four southeastern states (georgia, alabama, north carolina, and south carolina) for years 2010-2025 to evaluate feasibility for nationwide expansion. This work directly responds to longstanding calls by federal statistical agencies to improve demographic estimation for young children, particularly at subnational levels.