County-to-county migration modeling in the United States: the effects of data source and model selection

Internal migration plays a critical role in shaping demographic and economic landscapes, yet the ability to model migration flows accurately remains a methodological challenge. This study evaluates the performance of different migration models applied to three key U.S. data sources: the Internal Revenue Service (IRS) migration data, the American Community Survey (ACS), and the Census long-form data. While these datasets provide valuable insights into county-to-county migration, they differ in temporal coverage, flow suppression thresholds, and demographic granularity, each introducing unique challenges to migration modeling. Using a comparative framework, this study assesses the impact of data source selection on the accuracy and bias of widely used migration models, including the gravity model, Poisson regression, and the radiation model. Our findings highlight the trade-offs inherent in each dataset, demonstrating that IRS data yield lower prediction errors in aggregate flow estimates but lack demographic specificity, whereas ACS and Census data offer richer demographic detail and capture a larger number of distinct migration streams, though they may introduce noise due to small-flow estimates and suppression thresholds for confidentiality. The results underscore the importance of aligning data selection with research objectives and contribute to broader discussions on best practices for migration modeling.

Impact/Purpose

This article lays a foundation for assessing future impacts of climate change on human health. The results describe differences between results of statistical models used to project county-to-county migration that emerge solely as a function of the choice of input data source. A suite of migration models were derived from publicly available migration data sources from the Census, American Community Survey (ACS), and Internal Revenue Service (IRS). The results show that models estimated using the IRS data generally perform better as measured by cross-validation, however that benefit is shown to be a function of the censoring steps taking by IRS which exclude small (<20 individuals) county-to-county migration flows in the interest of privacy. Excluding those small flows from Census and ACS improves the respective cross-validation results substantially. Going forward, migration studies will benefit from these results, leveraging this work to better understand the implication of selecting a particular data source and to draw more robust conclusions regarding accuracy of future projections. Migration scholars from academia, NGO, and government may all benefit from this work.

Citation

Morefield, Philip E. AND T. Leslie. County-to-county migration modeling in the United States: the effects of data source and model selection. Springer, Heidelberg, GERMANY, 27(3):455-472, (2025). [DOI: 10.1007/s10109-025-00470-7]

DOI: County-to-county migration modeling in the United States: the effects of data source and model selection