Inferring fine-grained migration patterns across the United States

Abstract

Fine-grained migration data illuminate demographic, environmental, and health phenomena. However, United States migration data have serious drawbacks: public data lack spatial granularity, and higher-resolution proprietary data suffer from multiple biases. To address this, we develop a method that fuses high-resolution proprietary data with coarse Census data to create MIGRATE: annual migration matrices capturing flows between 47.4 billion US Census Block Group pairs-approximately four thousand times the spatial resolution of current public data. Our estimates are highly correlated with external ground-truth datasets and improve accuracy relative to raw proprietary data. We use MIGRATE to analyze national and local migration patterns. Nationally, we document demographic and temporal variation in homophily, upward mobility, and moving distance-for example, rising moves into top-income-quartile block groups and racial disparities in upward mobility. Locally, MIGRATE reveals patterns such as wildfire-driven out-migration that are invisible in coarser previous data. We release MIGRATE as a resource for migration researchers.

Fig. 1. MIGRATE estimates.

We estimate annual migration flows between all pairs of Census Block Groups (CBGs) from 2010 to 2019. a Average MIGRATE estimates of out-migration rates across the entire United States. b, c MIGRATE estimates of out-migration rates within New York City. MIGRATE estimates reveal granular spatial patterns invisible in publicly available county-to-county data (inset plot b). Out-migration rates for CBGs with fewer than 100 people are omitted.

Fig. 2. Validating the MIGRATE estimates.

a–c MIGRATE estimates (y-axis) are highly correlated with Census data (x-axis), including a Census populations at the Census Tract and Census Block Group (CBG) level, b movers between each pair of states and each pair of counties (excluding people who remain within the same state or county), and c state and county in-migration rates (i.e., the number of people moving into an area as a fraction of the area’s population). d–f MIGRATE estimates increase agreement with Census datasets relative to raw Infutor data for population counts, movers between states and counties, and in-migration rate, respectively. We compute root mean squared error (RMSE) between (1) MIGRATE estimates and Census data and (2) Infutor data and Census data, and report the reduction in RMSE from using MIGRATE estimates. Bars show the mean reduction in RMSE across all data release years (n = 5 for 5-year population and county-level migration datasets, n = 9 for 1-year state-level migration datasets); error bars plot standard deviation across years. To compare our 1-year MIGRATE estimates to 5-year ACS estimates of population and county-level migration, we average MIGRATE estimates across the same 5-year period each ACS data product covers, using only ACS data products whose time period completely overlaps with the 2010–2019 MIGRATE range. For in-migration rates, all metrics are weighted by state or county population, and points are sized by population.

Fig. 3. Assessment of demographic bias in the raw Infutor data and the MIGRATE estimates.

Infutor data displays biases that MIGRATE estimates greatly reduce. a Average errors in county populations in Infutor data relative to Census data; orange denotes counties where Infutor underrepresents the population, and purple denotes counties where Infutor overrepresents it. MIGRATE estimates remove all county-level errors by construction. b Spearman correlation between county demographics (x-axis) and error in Infutor estimates (y-axis). Infutor’s error is correlated with racial, socioeconomic, and other demographic characteristics. c Comparison of demographic bias in Infutor (purple) and MIGRATE (green). MIGRATE greatly reduces biases for all demographic subgroups. Bars compare demographic subpopulations estimated from Infutor or MIGRATE to ground-truth Census data, averaged over n = 5 population data releases (American Community Survey 5-year estimates, 2015 through 2018). Error bars represent standard deviations across these releases.

Fig. 4. National migration statistics.

a Flows between ten types of Census Block Groups (CBGs)—plurality white, Asian, Black, and Hispanic; urban versus rural; and bottom, second, third, and top income quartile. Rows correspond to the origin CBG, and columns to the destination CBG; for example, the top left entry indicates that 90% of movers from plurality white CBGs move to plurality white CBGs. The final two rows report the proportion of all movers moving to CBGs of each type, and the population share living in CBGs of each type. We report averages across all years. b Probability of moving to a higher-median-income CBG, conditional on income decile of origin CBG, and plurality race of origin CBG. c Distance moved stratified by CBG type.

Fig. 5. Migration in response to California wildfires.

a Out-migration following the Camp fire (2018; top) and Tubbs fire (2017; bottom). Red boundaries plot fire perimeters; black lines plot county boundaries; Census Block Groups (CBGs) are colored by domestic out-migration rate in MIGRATE. Out-migration rates exceed 50% in many CBGs within the fire perimeters. b Out-migration rates in different groups of CBGs over time according to MIGRATE estimates. Out-migration rates in the year after the fire are higher in CBGs within the fire perimeter (red line) than in groups of CBGs outside the fire perimeter (other lines), including those neighboring the fire perimeter, those in affected counties, or those within California. c Out-migration rates in the American Community Survey (ACS) 5-year county-to-county data remain relatively constant over time.

Fig. 6. Flowchart detailing the process of mapping addresses to Census Block Groups (CBGs).

We start with all Infutor addresses and classify them into five address types. We remove addresses which lie within US territories but not US states (around 0.26% of the addresses). Incomplete addresses—those that do not have a precise street address line—as well as PO boxes and rural routes are mapped probabilistically to CBGs according to their ZIP code. The vast majority (90.4%) of addresses are clean, complete street addresses, which are sent to the Census geocoder, which is able to map 81.26% of these addresses to a Census Block Group. If the Census geocoder fails to provide a match, we reattempt matching using ESRI's ArcGIS geocoder; this achieves a lower match rate of 58.24%, in part because the sample it is applied to is more difficult to parse. In total, we are able to map 99.21% of all addresses in the original Infutor data: 83.33% to a precise latitude and longitude, and 15.89% to a ZIP code.