Patterns of everyday movement, captured anonymously through smartphones, can reveal more about a neighborhood’s health than income, education, and other conventional indicators alone. A new study published in Computational Urban Science reports that adding place-visitation data to standard demographic and social determinants of health variables measurably improved the estimation of 22 health measures across nearly 70,000 US census tracts.
The research team, led by Temitope Akinboyewa and colleagues, drew on three large national datasets. Demographic and social determinants came from 12 variables in the 2019 American Community Survey five-year estimates. Health measures came from the Centers for Disease Control and Prevention’s PLACES dataset, which provides model-based estimates of the percentage of adults aged 18 and older in each tract who report specific conditions, behaviors, or preventive care use. The behavioral layer came from SafeGraph Patterns, now Advan Patterns, which aggregates anonymized GPS-enabled mobile device visits to roughly seven million points of interest, from parks and gyms to fast-food restaurants and healthcare facilities.
From these records, the researchers calculated per capita visits from each census tract to 120 categories of places for the year 2019, chosen because it predates the COVID-19 pandemic and therefore reflects stable, routine activity. Dividing visit counts by the number of resident devices adjusts for differences in panel coverage between tracts.
After restricting the analysis to contiguous US tracts with complete data across all three sources, the final sample comprised 69,424 census tracts, of which 56,608 were classified as urban and 12,816 as rural using the USDA’s Rural-Urban Commuting Area codes. For each of the 22 health measures, the team fitted hierarchical regression models in two stages: a baseline using demographic and social determinants only, and an enhanced model adding the visitation variables. Both used Extreme Gradient Boosting, a tree-based machine learning method well suited to capturing nonlinear relationships and feature interactions, with five-fold cross-validation to assess out-of-sample performance and guard against overfitting.
The gains were not uniform across health measures. A sensitivity analysis that randomly subsampled the data at 20, 40, 60, and 80 percent fractions confirmed the improvements remained stable and were not an artifact of unequal sample sizes.
The urban-rural contrast proved particularly striking. Although visitation data improved prediction for all 22 measures in both settings, the mean R² gain was 7.6 percent in urban tracts compared with 12.5 percent in rural tracts. In rural areas, 11 health measures improved by more than 10 percent, led by depression at 48.91 percent and binge drinking at 44.48 percent. In urban tracts, binge drinking led with a 38.81 percent gain. In a focused analysis, the urban binge drinking model’s R² rose from 0.55 to 0.76 with visitation data, while the rural model improved from 0.49 to 0.71; depression prediction in rural tracts climbed from 0.47 to 0.71.
Spatial analysis reinforced the pattern. Using Global Moran’s I statistics on model residuals, the researchers found that adding visitation data reduced the geographic clustering of prediction errors for both binge drinking and depression. Urban binge drinking residuals dropped from a Moran’s I of 0.672 to 0.358, and rural residuals from 0.548 to 0.334; depression showed similar reductions. Models relying only on demographic and social variables left strongly clustered errors across parts of the Midwest, South, and West, suggesting the visitation layer captured localized behavioral and environmental variation those variables miss. Some regions, notably Utah and to a lesser extent Louisiana, retained persistent residuals in the binge drinking models, which the authors link to distinctive regional alcohol regulations and cultural contexts.
To interpret which places mattered most, the team applied SHAP, a model interpretability technique that quantifies each feature’s contribution to individual predictions. Drinking Places (Alcoholic Beverages) consistently ranked as the strongest contributor to binge drinking predictions in both urban and rural tracts, with higher visitation generally associated with higher predicted prevalence. In urban areas, Religious Organizations was the second strongest predictor, with high visitation generally correlating with lower predicted binge drinking. In rural tracts, convenience-oriented destinations such as limited-service restaurants and gasoline stations with convenience stores featured prominently. For depression, urban predictions were driven largely by entertainment and commercial destinations including casinos, jewelry stores, and general merchandise stores, while rural predictions leaned on convenience stores, limited-service restaurants, hotels, and tobacco stores.
The authors are careful about what these findings do and do not mean. SHAP values describe the model’s output, not causal effects, and the associations operate at the census tract level, so inferring individual behavior from them would constitute an ecological fallacy. The data are cross-sectional, all variables are aggregated, and visitation, demographic, and social determinants likely share unmeasured confounders. The device panel also does not sample residents uniformly, and although normalization by resident device counts mitigates between-tract coverage differences, it does not fully correct unequal sampling within tracts, so visitation measures describe patterns in the device-bearing population rather than the full resident population. The authors also note their interpretations of urban-rural differences, such as concentrated rural activity around fewer destinations, should be treated as hypotheses for future investigation rather than established explanations.
Within those limits, the study makes a case for place-visitation data as a complementary input to small-area health estimation and surveillance, particularly in rural communities where static socioeconomic indicators may leave more variation unexplained. Because the analysis covered 22 measures simultaneously, it extends earlier work that focused on single outcomes such as obesity or depression. The authors suggest that mobility-derived indicators could help public health agencies identify localized behavioral contexts and allocate prevention resources more spatially, while cautioning that testing whether the observed associations reflect genuine behavioral or environmental mechanisms will require longitudinal designs, individual-level data, or causal-inference frameworks, and that future work should examine whether the gains hold at coarser geographic scales such as counties or states.
Subject of Research: Using smartphone place-visitation big data to improve small-area health measure estimation at the census tract level in the United States
Article Title: Exploring the role of place visitation big data on small area health measure estimation
Article References: Akinboyewa, T., Ning, H., Li, Z., Lessani, M. N., Li, X., & Qiao, S. (2026). Exploring the role of place visitation big data on small area health measure estimation. Computational Urban Science, 6(1), Article 52. https://doi.org/10.1007/s43762-026-00287-0
Image Credits: AI Generated
DOI: 10.1007/s43762-026-00287-0
Keywords: place visitation data, small-area estimation, census tracts, CDC PLACES, SafeGraph Patterns, XGBoost, social determinants of health, urban-rural differences, binge drinking, depression, predictive modeling, SHAP
News Source: Phoebe Ingram. (October 29, 2026). Phone Location Data Sharpens Neighborhood Health Estimates Across the United States. Scienmag.
