Exploring the role of place visitation big data on small area health measure estimation

Abstract

Population-level health measures at small geographic scales (e.g., census tracts), including health conditions, preventive behaviors, risk factors, and overall health status, are crucial for guiding effective health planning and policymaking. It has been well established that demographic and social determinants of health (SDOH) factors contribute to health disparities and thus are usually applied to predict health measure estimation. However, demographic and SDOH indicators are often static and fail to account for the dynamic aspects of daily life. This study explores the role of resident routine activity derived from place visitation big data in estimating health measures at the census tract level in the United States and tests this hypothesis across both urban and rural regions. Hierarchical regression analysis was conducted using demographic and SDOH data (12 variables from the 2019 American Community Survey), and smartphone-based place visitation data (visitation rates to 120 categories of places from SafeGraph Patterns). We analyzed 22 health measures from the CDC’s Population-Level Analysis and Community Estimates (PLACES) dataset, classifying census tracts as urban or rural using the USDA’s Rural-Urban Commuting Area (RUCA) codes. The results showed that incorporating place visitation data significantly contributes to health measure estimation beyond traditional demographic and SDOH variables (mean R² increased by 7.5%). This improvement varied substantially between urban (7.6%) and rural areas (12.5%). Among all health measures, Binge drinking exhibited the greatest predictive gains in the urban analysis (38.8%) and Depression has the highest predictive gains in the rural area (48.9%), with Drinking Places (Alcoholic Beverages) consistently identified as the place category contributing most to model predictions. This study suggests that place visitation big data can be useful auxiliary data source for small-area health measure estimation, complementing traditional demographic and SDOH variables.

Supplementary Information

The online version contains supplementary material available at 10.1007/s43762-026-00287-0.

Introduction

Population-level health measures at small-area levels, such as census tract-level health outcomes, surveillance of health risk behaviors, and overall health status assessment, are essential for effective public health planning and policy development. These measures provide spatially granular insights into community health needs and disparities, which enable effective allocation of resources and targeted interventions. Also, by identifying tracts (neighborhoods hereafter for an intuitive understanding) with the greatest health burdens or behavioral risks, small-area health measure estimation helps to guide public health strategies to maximize impact and efficiency. Also, it helps to identify specific local health disparities, particularly among disadvantaged or underserved communities.

Demographic and social determinants of health (SDOH), including economic stability, education access, health care access, neighborhood and built environment, and social and community context, could be critical predictors of population health outcomes and health disparities (Office of Disease Prevention and Health Promotion, 2025). Demographic and SDOH indicators have been widely adopted in population health research because of their ability to capture persistent, place-based exposures that shape long-term health outcomes. Numerous studies have demonstrated the relevance of the demographic and SDOH variables in predicting various health outcomes and informing public health planning (Alberta, 2024; Amrollahi et al., 2022; Rangachari et al., 2022; Vo et al., 2023).

However, demographic and SDOH indicators are often static, reflecting long term structural conditions while failing to account for the dynamic aspects of daily life. They do not incorporate information about where residents visit, what services they use, or how they interact with their environment. In contrast, patterns of daily mobility, such as routines, place visitation, and access to places, have been associated with individual and neighborhood health outcomes. A growing body of research highlights how visitations to specific places, including fitness centers, parks, pharmacies, hospitals, and drinking establishments, are significantly associated with health outcomes. For instance, frequent visits to fitness centers, recreational facilities, and parks can improve mental health, reduce risks of chronic diseases (e.g., cardiovascular diseases, and enhance overall physical well-being (Tasci et al., 2019; Zhuo et al., 2021; Emery et al., 2022; Fabris & Sinagra, 2022; Cabanas-Sánchez et al., 2024; Jing et al., 2024; Kunutsor & Laukkanen, 2024; Patwary et al., 2024; Akinboyewa et al., 2025). Regular visits to fast-food outlets are associated with unhealthy dietary patterns and higher risks of obesity (Liu et al., 2020; Van Der Velde et al., 2022; Zhou et al., 2022; Xu et al., 2023). Routine visits to healthcare facilities, particularly primary care providers, support early detection and management of chronic conditions, thereby improving long-term health outcomes (Hostetter et al., 2020; Paul, 2024). By omitting routine activity patterns reflected in place visitations, current models may overlook information that could improve the prediction of health measures, especially in communities whose behaviors are not fully captured by demographic and SDOH variables alone.

Importantly, the association between these patterns of daily mobility and health measures may vary substantially between urban and rural communities. Differences in population density, transportation infrastructure, service availability, and built environment influence how individuals engage with their surroundings and, consequently, how mobility patterns relate to health measures. For instance, urban areas often feature greater accessibility to healthcare facilities, fitness centers, and healthy food outlets, promoting more frequent and diverse place visitations that can positively influence health (Losada-Rojas et al., 2021; Jiang et al., 2022; Liu et al., 2022; Henning-Smith et al., 2023; Iamtrakul et al., 2024). In contrast, rural residents may face longer travel distances, fewer service options, and limited transportation, which can constrain visitation patterns and reduce access to health-promoting resources (Cattaneo et al. 2021a; Lechowski and Jasion 2021; Nolan-Isles et al. 2021; Mseke et al. 2024).

Place visitation big data derived from anonymized mobile devices provides an opportunity to access these health-related behaviors at the neighborhood level. These data capture aggregate place visitation patterns at the neighborhood level and provide insight into real-world behaviors that affect health. Recent studies have demonstrated the utilization of such data in public health research. Zhou et al. (2022) demonstrated improvements in obesity prediction through visitation to fast-food outlets, fitness centers, and parks. Jing et al. (2024) used visitation patterns derived from place visitation big data to assess mental health outcomes like depression based on recreational and health service access. Li et al. (2024a, b) revealed spatial and racial disparities in healthcare utilization during the COVID-19 pandemic using place visitation big data. Additionally, Xu et al. (2023) showed that visitation-based metrics such as the Retail Food Activity Index outperform traditional access measures in predicting cardiometabolic risk.

Other studies have extended the use of visitation data to assess behavioral health service utilization and risk exposures. Jing et al. (2023) estimated mental health service visits and visit-to-need ratios across the United States (US) neighborhoods using mobile phone data, and Chang et al. (2022) linked neighborhood-level alcohol outlet visitation derived from mobile phone-based visitation data to domestic violence risk. Additionally, Yu et al. (2024) leveraged smartphone mobility data to estimate actual travel times to mental health facilities across the US. This growth of the body of research highlights the potential of place visitation to complement traditional health modeling approaches by capturing dynamic and behavior-driven exposures at a fine spatial scale.

Therefore, this study aims to investigate whether incorporating routine activity information from place visitation data plays a significant role in health measure estimation in small-population areas (e.g., census tracts), and whether these improvements vary between urban and rural communities. Specifically, the following research questions were addressed: (1) Does the inclusion of routine activity data, derived from place visitation big data, improve the estimation of health measures in small-population areas in the US? (2) How does the impact of routine activity information on health measure estimation vary between urban and rural areas? Our study contributes to the body of knowledge in three ways. Firstly, it addresses a key limitation of demographic and SDOH predictors - by introducing dynamic routine information obtained from place visitation data, to enhance the estimation of health measures. Secondly, while research has demonstrated the importance of place visitation data, most existing studies have focused on single health condition, such as obesity, and depression. This study extends the literature by evaluating the usefulness of visitation data across multiple health measures (N = 22) simultaneously (including health outcomes, health behaviors, and health services utilization), which helps to assess its broader applicability in the context of public health. Thirdly, by examining differences in predictive performance across urban and rural areas, the study offers insight into geographic variability in how visitation data contribute to health measure estimation. This is crucial in guiding more focused and balanced public health strategies, particularly in small-population areas.

Methods

Measures

Demographic and SDOH indicators

In this study, demographic and SDOH were represented by 12 variables derived from 2019 American Community Survey (ACS) 5-year estimates at the census tract level, provided by the US Census Bureau (US Census Bureau, 2025). These variables capture key dimensions of social, economic, and demographic characteristics which align with the description of SDOH given by the Office of Disease Prevention and Health Promotion and were chosen based on their relevance in prior public health literature (Grant et al., 2022; Qiao et al., 2022; Zhou et al., 2022; Liu et al., 2024). A detailed description of these variables is provided in Table 1.

Table 1.

Description of the demographic and SDOH variables used in this study

Variable name	Description
%White	Proportion of White population
%Black	Proportion of Black or African American population
%Hispanic	Proportion of Hispanic population
%Asian	Proportion of Asian Population
%Senior	Proportion of population aged 60 years and above
%Poverty	Proportion of individuals living below the federal poverty level
%Uninsured	Proportion of population without health insurance coverage
%Unemployed	Proportion of population that is unemployed
%High School	Proportion of population aged 25 years and over with at least a high school diploma or equivalent
%Rent Housing	Proportion of occupied housing units that are renter-occupied
%Living alone	Proportion of households where someone lives alone
Median household income	Median income of all households in a census tract

Health measures

The health measure variables in this study consist of 22 measures, each representing the estimated prevalence – that is, the percentage of adults aged 18 years and older in a census tract who report having a specific health condition, engaging in a health-related behavior, or utilizing preventive care services. These measures were obtained from the Population Level Analysis and Community Estimates (PLACES) dataset, developed by the Centers for Disease Control and Prevention (CDC) (PLACES, 2025). The selected measures encompass a wide range of physical health conditions (e.g., obesity, diabetes), mental health outcomes (e.g., depression, frequent mental distress), health-related behaviors (e.g., binge drinking), and preventive care utilization (e.g., routine medical checkups, cholesterol screening). These prevalence estimates are model-based and provided at the census tract level for all US tracts.

Routine activity information (place visitation)

Routine activity information consisted of per capita visits from census tracts to 120 categories of places (e.g., fast food restaurants, parks, fitness centers, grocery stores, and recreational centers) for the year 2019. The routine activity information was derived from SafeGraph Patterns data (SafeGraph, 2019), now Advan Patterns (Barry, 2023). The Patterns data provides anonymized mobile device data derived from GPS-enabled mobile devices collected from millions of users across the US. Patterns are aggregated at the census block group and tract levels, offering detailed information about visits to approximately 7 million points of interest (POIs) such as parks, restaurants, gyms, healthcare facilities, and religious centers (Jing et al., 2024). The data includes the home neighborhoods of visitors to POIs, which enables the identification of the census tracts of origin of visitors across the US. Using this information, we calculated the per capita visits from each census tract to the selected POI categories.

For each census tract T, the per capita visits to a specific POI category were calculated using the following equation:

where represents the number of visits from tract  T to POI ( n in total) reported in SafeGraph Patterns, and is the number of resident devices in tract T. Rather than treating  V as a population-representative per-capita visitation rate, we interpret it as a relative measure of tract-level routine activity. It is worth noting that SafeGraph Patterns is derived from a non-random panel of GPS-enabled mobile devices with a partial sampling rate that averages approximately 7.5% nationally and ranges from roughly 4.5% to 14.5% across space, time, and population groups (Li et al. 2024a, b). The panel, therefore, does not sample residents uniformly. Because N counts resident devices rather than the resident population, dividing visit counts by adjusts for between-tract differences in panel coverage and yields a measure of average visits per sampled device. The resulting  V can then be interpreted as a behaviorally informative signal of where tract residents tend to visit.

The year 2019 was selected purposely to ensure that the measures reflect normal conditions rather than pandemic-induced behavioral disruptions. Given the WHO declared the COVID-19 as a pandemic on March 11, 2020, the 2019 data is the most recent dataset well-suited to our study aims, as it reflects stable, routine activity patterns and health behaviors.

To ensure comparability across models and avoid biased estimates due to missingness, we restricted the analysis to contiguous US census tracts with complete data across the ACS, CDC PLACES, and SafeGraph Patterns datasets. Census tracts were retained only if: (1) the CDC PLACES dataset reported non-suppressed prevalence estimates for all 22 health measures, (2) the SafeGraph Patterns dataset reported at least one resident device observation required to compute per-capita visitation rates, and (3) all 12 ACS-derived demographic and SDOH variables were non-missing. These criteria were applied uniformly nationwide prior to urban-rural stratification based on 2010 RUCA codes. After applying the completeness and merging criteria, the final sample consisted 69,424 census tracts (56,608 urban; 12,816 rural).

Statistical analysis

Hierarchical regression analyses were conducted to evaluate the contribution of routine activity information (place visitation) to the estimation of health measures at the census tract level. In the first stage, we fitted a baseline model where each health measure was estimated based on the demographic and SDOH variables only. In the second stage, another model was fitted by adding place visitation variables in addition to the demographic and SDOH variables, enabling us to assess the additional explanatory power provided by residents’ routine activity information. Predictors were standardized using z-score normalization prior to model fitting. For each health measure, we employed Extreme Gradient Boosting (XGBoost), which was selected because tree-based ensemble learning methods have demonstrated strong predictive performance in high-dimensional health and epidemiological datasets and are capable of capturing nonlinear relationships and feature interactions without requiring strict parametric assumptions (Chen & Guestrin, 2016; Atias et al., 2025). To assess model generalization and reduce the risk of overfitting, we employed 5-fold cross-validation. Census tracts were randomly partitioned into five mutually exclusive folds. In each iteration, four folds were used for model training, and the remaining fold was used for validation. The XGBoost regressor was used with its default hyperparameters (n_estimators = 100, max_depth = 6, learning_rate = 0.3), which provided baseline performance across all 22 health measures. Model performance was evaluated using out-of-sample obtained from 5-fold cross-validation. Performance metrics were averaged across folds to estimate out-of-sample model performance. was selected as the primary evaluation metric because it directly quantifies the additional variance explained by the inclusion of visitation data across health measures and geographic contexts. Improvements were quantified by comparing mean values between the baseline (demographic and SDOH-only) models and the enhanced (demographic and SDOH plus visitation) models. Subgroup analyses were conducted by stratifying census tracts into urban and rural categories according to the 2010 Census Urban-Rural Classification, to assess variation in the impact of place visitation data across urbanicity contexts.

Sensitivity analysis was also performed by randomly subsampling the dataset at various fractions (e.g., 20%, 40%, 60%, and 80%) and evaluating the stability of model improvements in order to test the robustness of the model improvement with respect to sample size.

To further evaluate the impact of place visitation data, we conducted a focused analysis on Binge drinking and Depression. Binge drinking was selected because it exhibited the largest improvement in model performance among all health measures, while Depression was included due to its observed urban-rural difference in model improvement. We further examined the spatial distribution of absolute residuals for binge drinking and depression predictions across both urban and rural census tracts. To further assess the spatial structure of model errors, Global Moran’s I spatial autocorrelation analysis was conducted on the residuals of both the SDOH-only and SDOH plus place visitation models for Binge drinking and Depression in urban and rural census tracts. Moran’s I was used to quantify the degree of spatial clustering in model residuals, where higher positive values indicate stronger clustering of prediction errors across neighboring census tracts. Spatial autocorrelation analysis was performed using queen contiguity-based spatial relationships.

Results

Model performance results for health measure prediction for entire US

Table 2 presents the average performance across these 22 measures. The mean (averaged across the 22 measures) increased from 0.814 to 0.875 (7.5% improvement), and the average root mean square error (RMSE) decreased from 2.082 to 1.679, indicating greater overall prediction accuracy with the inclusion of routine activity data.

Table 2.

Average model performance for predicting 22 health measures across all US census tracts. Model 1 uses demographic and SDOH variables only, while model 2 incorporates routine activity variables derived from place visitation data

Model	Predictors	R 2	RMSE
1	Demographic and SDOH indicators	0.814	2.082
2	Demographic and SDOH + Place routine activity variables	0.875	1.679
	Improvement	7.5%

Table 3 presents health measures that obtained above 10% improvement in predictive performance after the inclusion of routine activity. Among the 22 health measures analyzed (Supplementary Table 1), several outcomes benefited from the addition of routine activity information. Binge drinking prevalence exhibited the largest improvement, with improving by 36.85%. Other health measures showing substantial improvements ( improvement greater than 10%) included Visits to doctors for routine checkups (24.47%), Depression (22.35%), Current asthma (12.58%), Obesity (12.03%), High cholesterol (10.96%), and Cholesterol screening (10.32%).

Table 3.

Health measures showing above 10% improvement in predictive performance () following the inclusion of routine activity obtained from place visitation data using the XGBoost model. Model 1 includes demographic and SDOH data only, while Model 2 incorporates both demographic and SDOH plus place visitation data

Health measure	Model 1 R 2	Model 2 R 2	Improvement (%)
Binge drinking	0.552	0.756	36.85
Visits to doctor for routine checkup within the past year	0.671	0.835	24.47
Depression	0.658	0.806	22.35
Current asthma	0.768	0.865	12.58
Obesity	0.787	0.881	12.03
High cholesterol	0.751	0.833	10.96
Cholesterol screening	0.766	0.845	10.32

Performance improvement in health measure prediction for urban and rural tracts

Place visitation data increased the prediction performance of all health measures with significant urban-rural differences (mean increased by 7.6% for urban tracts: tracts = 56,608; and 12.5% for rural tracts: tracts = 12,816) (Supplementary Table 2). The performance improvements vary among health measures. As Fig. 1 presents, the health measures that benefited with more than 10% performance improvements (i.e. improvement) in urban tracts include Binge Drinking (38.81%), Visits to doctor for routine checkup (24.36%), Depression (22.20%), Obesity (12.25%), Current asthma (12.03%), High cholesterol who have been screened in the past 5 years (11.98%), and Cholesterol screening (10.17%). For rural areas, 11 health measures benefited with more than 10% performance improvements with the top seven including Depression (48.91%), Binge drinking (44.48%), Visits to Doctor for Routine Checkup (33.30%), Obesity (26.62%), Current asthma (20.78%), High cholesterol who have been screened in the past 5 years (17.74%), and Cholesterol screening (16.93%) (for more details on the result, refer to Supplementary Tables 3–4).

Fig. 1.

Performance improvement in health measure prediction for urban and rural tracts

To assess whether these improvements are sensitive to sample size differences between urban and rural tracts, a sensitivity analysis was conducted. As shown in Fig. 2, the model performance remained stable across varying sample fractions, suggesting that the observed performance improvement was not driven by differences in the number of census tracts analyzed.

Fig. 2.

Sensitivity analysis result

Focused analysis on Binge drinking and depression

Focused analysis on Binge drinking and Depression suggests that the integration of routine activity information substantially enhanced model accuracy in both urban and rural contexts. For Binge drinking, the urban SDOH-only model achieved an of 0.55, which increased to 0.76 after incorporating place visitation data (Fig. 3a), while the rural model improved from 0.49 to 0.71 (Fig. 3b). Similarly, Depression prediction also shows improvements after incorporating visitation data. In urban analysis, the increased from 0.67 to 0.82 (Fig. 3c), while in rural analysis, the improved from 0.47 to 0.71 (Fig. 3d).

Fig. 3.

Observed and predicted prevalence of Binge drinking and depression in urban and rural tracts

To further evaluate the impact of place visitation data, we examined the spatial distribution and spatial autocorrelation of absolute residuals for both Binge drinking (Fig. 4) and Depression (Fig. 5) predictions across census tracts. For Binge drinking, the inclusion of place visitation data reduced both the magnitude and spatial clustering of residuals in urban and rural tracts. In urban tracts, Moran’s I decreased from 0.672 in the SDOH-only model (Fig. 4a) to 0.358 after incorporating visitation data (Fig. 4b), while rural tracts showed a decrease from 0.548 (Fig. 4c) to 0.334 (Fig. 4d). Spatially, the SDOH-only models exhibited stronger clustering of high residuals across portions of the Midwest, South, and Western United States, whereas the SDOH plus visitation models showed more fragmented and less spatially concentrated residual patterns.

Fig. 4.

Spatial distribution of absolute residuals and Global Moran’s I statistics for Binge drinking across urban and rural census tracts in the US. Dark blue and red indicate lower and higher prediction error respectively

Fig. 5.

Spatial distribution of absolute residuals and Global Moran’s I statistics for Depression across urban and rural census tracts in the US. Dark blue and red indicate lower and higher prediction error respectively

Similar patterns were observed for Depression. In urban tracts, Moran’s I decreased from 0.692 to 0.399 after incorporating visitation data (Fig. 5a-b), while rural tracts showed a reduction from 0.617 to 0.405 (Fig. 5c-d). The SDOH-only models exhibited stronger residual clustering across portions of the Pacific Northwest, Midwest, and Great Plains. After incorporating place visitation data, residual errors became less spatially clustered and more dispersed across regions.

Figure 6 presents the most influential place visitation categories associated with Binge drinking and Depression prediction, based on SHAP (SHapley Additive exPlanations) (SHAP, 2025) value analysis. SHAP is a model interpretability method that quantifies the contribution of each feature to a specific prediction. Positive SHAP values indicate that a feature increases the predicted prevalence of a health outcome, while negative SHAP values indicate a decreasing contribution to the model output (Lundberg & Lee, 2017).

Fig. 6.

Top 10 place visitation categories ranked by contribution to the prediction of Binge drinking (a, b) and Depression in urban and rural areas (c, d)

For Binge drinking, it was observed that Drinking Places (Alcoholic Beverages) and Religious Organizations were the top two predictors contributing to the prevalence of binge drinking in urban tracts. As shown in Fig. 6a, high visitation to Drinking Places (Alcoholic Beverages) was generally associated with positive SHAP values, indicating that greater visitation to these locations tends to increase the model-predicted prevalence of Binge drinking. Also, Religious Organizations show a mixed contribution (i.e., both high and low feature values span negative and positive SHAP ranges), but high visitation generally correlates with a reduction in Binge drinking prediction (negative SHAP values with red dots). SHAP values reflect each feature’s contribution to the model’s predictions and should not be interpreted as causal effects; this association is observed at the census tract level and does not imply a protective effect at the individual level.

In rural areas (Fig. 6b), Drinking Places (Alcoholic Beverages) again demonstrated the highest SHAP impact, with higher visitation levels generally corresponding to positive SHAP values and higher predicted binge drinking prevalence. Limited-Service Restaurants shows a negative SHAP value trend for high visitation, suggesting that visitation to these places predicted a lower Binge drinking prevalence in rural areas.

For Depression, distinct urban-rural differences were observed in the types of visitation categories contributing to model predictions. In urban tracts (Fig. 6c), the most influential predictors were largely entertainment- and commercial-service-related POIs, including Casinos (except Casino Hotels), Jewelry Stores, All Other General Merchandise Stores, and Beer, Wine, and Liquor Stores. Higher visitation to Casinos (except Casino Hotels), Jewelry Stores, and General Merchandise Stores generally corresponded to positive SHAP values and increased predicted depression prevalence. In contrast, categories such as Caterers, Casino Hotels, Art Dealers, and Child Day Care Services generally showed negative SHAP trends for higher visitation levels. In rural tracts (Fig. 6d), the most influential predictors were primarily convenience-oriented and service-access-related destinations, including Convenience Stores, Limited-Service Restaurants, Hotels and Motels, Gasoline Stations with Convenience Stores, and Tobacco Stores. Several of these categories showed positive SHAP trends for higher visitation levels, indicating associations with increased predicted depression prevalence.

Discussion

Estimating health measures plays a crucial role in public health planning and decision-making. However, existing models have relied heavily on demographic and SDOH variables, which do not capture the dynamic nature of residents’ routine activities. In this study, we examined whether integrating routine activity information derived from place visitation data improves the estimation of health measures at the census tract level in the US. The findings indicate that adding these dynamic behavioral indicators to traditional demographic and SDOH indicators yields measurable improvements in predictive performance for a range of health outcomes.

Among the 22 health measures analyzed and for entire US tracts, Binge drinking exhibits the largest relative improvement in model performance, with 36.85% increase in . Other health measures that exhibited notable improvements include Visits to doctor routine checkup within the past year (24.47%), Depression (22.35%), Current asthma (12.58%), Obesity (12.03%), High cholesterol (10.96%), and Cholesterol screening (10.32%). These percentages represent the relative increase in the compared to baseline models that included only SDOH, indicating the additional explanatory power gained by integrating routine activity data into the prediction models.

Sensitivity analyses confirmed that these improvements were robust to potential biases due to unequal sample sizes between urban and rural tracts. It was observed from the results that model improvement remained stable across different sample sizes, indicating minimal sensitivity to changes in sample fraction. This suggests that the observed enhancement in explanatory power from incorporating routine activity data is robust and not merely an artifact of unequal tract representation.

The spatial autocorrelation analysis further demonstrated that incorporating place visitation data significantly reduced the geographic clustering of model residuals for both Binge drinking and Depression. The reduction in Moran’s I values suggests that routine activity information captures additional spatially behavioral and environmental variation that may not be fully represented by demographic and SDOH variables alone. The reduced residual spatial clustering also suggests that the models were better able to account for localized contextual differences associated with these health outcomes. We also observed that certain regions, particularly Utah and to a lesser extent Louisiana, exhibited more persistent residual patterns in the Binge drinking models even after incorporating place visitation data. This suggests that these areas may reflect distinctive regional behavioral, cultural, or policy-related contexts associated with alcohol consumption that are not fully captured by current demographic, SDOH, and visitation variables. For example, Utah has consistently reported among the lowest binge drinking prevalence rates in the US and is also characterized by unique alcohol regulations and social environments influenced in part by the strong presence of the LDS population within the state (Bohm et al., 2021; America’s Health Rankings, 2024; Utah, 2024). Our observation suggests that some geographically specific behavioral contexts may remain difficult to capture using generalized national-scale models and may require additional region-specific contextual variables in future work.

When examining urban and rural contexts separately, results revealed that rural areas showed a substantially larger average improvement in predictive accuracy compared to urban areas. Specific outcomes such as Binge drinking and Depression exhibited particularly large gains in rural tracts after incorporating place visitation data. In urban areas, residents typically have many nearby options for healthcare, food, recreation, and other services (Cyr et al., 2019). As a result, visitation patterns in urban settings may be more dispersed and heterogeneous, which could make it more difficult to identify consistent tract-level associations using aggregated visitation measures. Moreover, traditional SDOH data, which includes detailed socioeconomic and demographic indicators, may already capture substantial variation across densely populated urban areas. Rural areas may differ in this regard because access to services and amenities is often more limited and routine activities may be concentrated around a smaller number of commonly visited destinations (Vitale Brovarone and Cotella 2020; Cattaneo et al. 2021b). Therefore, aggregate visitation to places such as clinics, religious centers, or convenience stores may provide additional tract-level information beyond what static demographic and SDOH variables convey. Similarly, lower demographic variation across some rural tracts may increase the relative contribution of visitation patterns to model prediction. However, these interpretations should be seen as hypotheses for future investigation rather than causal explanations for the observed predictive gains.

The SHAP analysis provided further insights into specific place visitation categories that contribute to model predictions differently across urban and rural areas. Although Drinking Places (Alcoholic Beverages) consistently exhibited strongest SHAP impact influence on Binge Drinking, the broad distribution of SHAP values observed suggests that the relationship between drinking places visitation and Binge drinking prediction varies across different tract-level contexts. The variability may be due to the interaction that may exist between drinking places visitation with other tract level factors, such as availability of alternative recreational destinations, socioeconomic conditions, or broader neighborhood activity environments. We also identified different place visitation categories associated with Binge drinking. In urban areas, visitation to Religious Organizations had a generally negative association with predicted Binge drinking at the tract level. This pattern was less pronounced in rural areas, where convenience and alcohol related places contributed more strongly to model predictions. One possible interpretation is that religious organizations in urban settings may serve broader social and community-support functions beyond religious participation alone, potentially contributing to social environments associated with drinking behavior. This interpretation is consistent with previous studies that have documented a negative association between religious participation and binge drinking (Guimarães et al., 2018; Rivera et al., 2018; Mohler, 2024). In contrast, rural areas showed stronger associations with Limited-Service Restaurants and Gasoline Stations with Convenience Stores, suggesting that patterns of binge drinking in these areas may be influenced by limited access to recreational facilities and higher exposure to food environments associated with risk behaviors. The scarcity of recreational facilities may limit opportunities for engaging in healthy social activities, potentially leading individuals to seek alternative forms of leisure, such as alcohol consumption. Interestingly, standalone Convenience Stores show a positive association with Binge drinking. This may be because such stores often serve as easily accessible points of alcohol purchase in rural communities, where few alternative retail or entertainment options exist.

Similarly, the Depression SHAP analysis revealed different place visitation environments associated with depression prediction across urban and rural areas. In urban tracts, entertainment- and commercial-service-related destinations such as Casinos (except Casino Hotels), Jewelry Stores, and General Merchandise Stores contributed strongly to Depression prediction. This may be because these destinations reflects areas characterized by more socially and commercially intensive activity environments, which may be associated with stress-related or socially isolating urban lifestyles at the tract level. In contrast, rural areas showed stronger associations with Convenience Stores, Limited-Service Restaurants, Gasoline Stations with Convenience Stores, and Hotels. This may reflect the limited availability of recreational, healthcare, and social infrastructure commonly observed in rural communities, where convenience-oriented destinations often serve as major routine activity spaces.

This study contributes to growing evidence that behavioral data, such as anonymized place visitation records, can enhance population health measures modeling. It demonstrates that the place visitation big data has the potential of providing meaningful behavioral context that can improve the estimation of health outcomes. These behavioral indicators can help in capturing social and environmental exposures not reflected in traditional SDOH data. By reflecting real-world activity patterns, these data improve predictive accuracy and offer a more context-aware understanding of place-based health disparities. Also, the observed urban-rural differences in visitation environments associated with Binge drinking and Depression prediction may also have implications for geographically tailored public health strategies. For example, the stronger association between binge drinking prediction and alcohol-related or convenience-oriented destinations in rural areas may help identify communities where interventions focused on alcohol access, community engagement, or alternative recreational resources could be prioritized. In contrast, the more diverse entertainment- and commercial-service-related visitation patterns observed in urban areas may indicate the need for broader community-level prevention approaches that account for more heterogeneous mobility and social environments. More generally, mobility-derived visitation data may help public health agencies identify localized behavioral contexts associated with health outcomes and support more spatially informed allocation of prevention resources.

Nonetheless, some limitations should be acknowledged. First, a limitation concerns the representativeness of the smartphone-based mobility data used in this study. SafeGraph Patterns is derived from a third-party panel of GPS-enabled mobile devices and panel coverage varies across space, time, and population subgroups as a result of differences in smartphone ownership, application usage, and location-sharing behavior (Li et al. 2024a, b). These differences may be especially pronounced between urban and rural areas, where device ownership patterns, network connectivity, and the density of tracked points of interest differ, as well as across demographic groups whose smartphone use is known to vary. Although normalizing visit counts by the number of resident devices reduces the influence of between-tract differences in panel coverage, this adjustment does not fully correct for unequal sampling probabilities within tracts. The visitation measures should therefore be interpreted as describing patterns observable in the device-bearing population rather than as direct estimates for the full resident population, and should be interpreted with greater caution in tracts with limited panel coverage. Second, this study reports improvements in predictive performance, not causal effects. Several aspects of the design limit causal and individual-level interpretation. This is because the data are cross-sectional, all variables are aggregated to the census tract level, and visitation, demographic, and SDOH variables are likely interrelated and may share unmeasured cofounders. Inferring individual-level behavior or causal mechanisms from tract level associations would constitute an ecological fallacy. SHAP values reported in this study quantify each predictor’s contribution to the model’s output and should be interpreted as descriptions of the model, not as causal effect estimates. Future research using longitudinal designs, individual-level data, or causal-inference frameworks would be needed to assess whether and how the observed associations reflect underlying behavioral mechanisms. Another limitation is that the analyses were conducted exclusively at the census tract level. The contribution of place visitation data to health measure estimation may vary across alternative geographic aggregation levels, such as counties or states, due to differences in spatial heterogeneity and contextual averaging. Future research should evaluate whether the reported relationships and model improvements remain consistent across multiple spatial scales.

Conclusion

This study examined whether incorporating resident routine activity data derived from anonymized mobile phone place visitation records improves health measure estimation at the census tract level in the US. The results indicate that adding place visitation variables to traditional demographic and SDOH predictors yields measurable gains in predictive performance across multiple health measures, with the largest relative gains observed for binge drinking, depression, routine medical checkups, obesity, and current asthma. Predictive gains were larger, on average, in rural tracts than in urban tracts, suggesting that visitation data may carry information not captured by static demographic and SDOH variables in these settings. We emphasize that these findings concern predictive accuracy at the census tract level. They do not establish that visitation to particular place categories causes the health outcomes examined, nor that individual-level behaviors map directly onto the tract-level associations reported here; cross-sectional, aggregated data of this kind cannot support such inferences. With these limits in mind, the results suggest that place visitation data may be a useful auxiliary input for small-area health measure estimation and surveillance, complementing rather than replacing established demographic and SDOH indicators. Examining whether and how the observed associations reflect underlying behavioral or environmental mechanisms will require longitudinal designs, individual-level data, or causal-inference frameworks.

Authors’ contributions

Conceptualization: Li, Akinboyewa, Ning; Methodology: Akinboyewa, Ning, Lessani, Li; Data curation: Akinboyewa, Ning, Lessani; Formal analysis: Akinboyewa, Ning; Visualization: Akinboyewa, Ning; Writing original draft: Akinboyewa, Ning; Writing review and editing: Li, Qiao, Ning, Lessani, X. Li; Supervision: Li; Funding acquisition: Li, Qiao.

Funding

This manuscript is the result of funding in part by the National Institutes of Health (NIH) under Award Number R21MD018666. It is subject to the NIH Public Access Policy. Through acceptance of this federal funding, NIH has been given a right to make this manuscript publicly available in PubMed Central upon the Official Date of Publication, as defined by NIH. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data availability

The dataset with all the variables used in the statistical analysis is included in the Supplementary file.

Declarations

Competing interests

The authors have declared that no competing interests exist.