We conducted a cross-sectional ecological study at the census tract level across nine of the 10 largest U.S. metropolitan statistical areas (MSAs; hereafter referred to as metropolitan areas or “metros”). MSAs were defined using the U.S. Census Bureau's 2020 delineation of metropolitan and micropolitan areas. Miami was excluded due to the absence of tract-level health estimates in CDC PLACES. For each included metro, county Federal Information Processing Standards (FIPS) codes were obtained from the Census Bureau, and all constituent census tracts were selected. The number of counties per metro ranged from two (Phoenix, Los Angeles) to 22 (New York City), yielding a total of 13,152 census tracts across the nine study metros.

We compiled a comprehensive dataset by integrating information from multiple national sources that provide small area estimates or spatially continuous data (Table 1). These included health indicators, environmental exposures, social vulnerability indices, and demographic composition for each MSA at the census tract level. All data are publicly available, de-identified and contained no personally identifiable information and as such, institutional review board (IRB) approval and informed consent were not required.

Publicly available census tract-level health data were obtained from the CDC PLACES Project, which provides modeled estimates of chronic disease risk factors, health behaviors, and outcomes for U.S. census tracts. PLACES estimates are derived using small-area estimation methods that combine survey data from the Behavioral Risk Factor Surveillance System (BRFSS) with population demographics from the Census Bureau. The outcome was the prevalence of adults reporting ≥14 days of poor mental health in the past month, denoted as frequent mental distress in adults, a validated indicator for poor mental health used in population-level health surveillance.

Data on greenspace visits were obtained from Advan Research mobility traces (20), monthly and aggregated to the whole years of 2021. This dataset aggregates anonymized smartphone location data from millions of devices in the U.S. and provides visit counts to points-of-interest (POIs). Greenspace POIs were defined using the U.S. Park Serve database and OpenStreetMap land-use categories, restricted to public-access parks and recreational spaces. Greenspace visitation data for a given census tract were defined as the total number of visits to greenspace locations made by mobile devices whose owners were residents of that tract, divided by the total number of resident devices in the tract, defined as:

Device residency was determined using Advance Patterns' algorithm, which assigns a device's home census tract based on its long-term nighttime location. This normalization adjusts for variation in the density of devices observed by the mobility panel across geographic areas.

Environmental and social covariates were selected to reflect contextual factors that may influence neighborhood mental health. Normalized Difference Vegetation Index (NDVI) was derived from 2021 VIIRS satellite imagery obtained through NASA's EarthData platform. 12 monthly raster files (8th day of each month in 2021) were downloaded, stacked, and averaged to create an annual mean index. Tract-level data was calculated using spatial overlays: census tract polygons (2010 TIGER/Line shapefiles) were used to extract mean NDVI values from the raster via zonal statistics. Values less than zero (representing water or blue spaces) were set to zero, resulting in our dataset that reflects census level vegetation presence across all our metropolitan areas. Finally, since the obtained data from 2021 primarily relied on 2020 census boundaries, we cross walked them to reassign 2020 bounded data to 2010 geographic areas.

Social vulnerability obtained from the CDC/ATSDR Social Vulnerability Index ((19) release), incorporates 15 census variables are grouped into four themes: socioeconomic status, household composition, racial/ethnic minority status and language, and housing/transportation. For this analysis, we included each of the four theme-specific indices. SVI values range from 0 to 1, with higher values indicating poorer social vulnerability.

The study also adjusted for demographics data, obtained from the American Community Survey (ACS, 2010–2014). Using tract-level data, we extracted three key variables: percentage of female residents, percentage of Black residents, and median age by census tract. These variables were queried and processed and merged into a complete dataset. All datasets were harmonized to the tract geography using 2010 U.S. Census boundaries and merged using GEOID identifiers. The final data frame, after merging all our individual datasets retained coverage for more than 95% of all tracts across the nine MSAs (Table 2). While data from Advan Research was obtained under a commercial license, all other datasets are publicly available. As all data were aggregated at the census-tract level and contained no personally identifiable information, institutional review board (IRB) approval and informed consent were not required.

All analyses were conducted at the census tract level within each of the nine metropolitan statistical areas (MSAs), relying on the official county-to-MSA crosswalk published by the Office of Management and Budget (OMB) and disseminated through the U.S. Bureau of Labor Statistics (BLS) (13). All analysis were conducted on R Studio 2025.05.0 Build 496. Prior to modeling, descriptive statistics were generated for all study variables to examine distributions and assess tract-level coverage across datasets.

For each MSA, an OLS regression model estimated the association between poor mental health prevalence and greenspace visits, adjusting for NDVI, SVI theme scores and demographic variables. Scatterplots were generated to visualize the associations.

Formally, the global OLS model can be expressed as:

Where Yi denotes the prevalence of poor mental health in census tract I, Gsi represents greenspace visits, NDVIi captures the tract-level vegetation presence, SVIi represents SVI themes from the CDC/ATSDR, Xi denotes the demographic covariates, and εi is the error term. Two supplementary OLS models were additionally conducted to evaluate the relative contributions of NDVI and greenspace visits by alternatively including and excluding each predictor alone and in combination while holding all other covariates constant. In addition, an overall OLS model pooling data from all nine MSAs was conducted. To explore spatial heterogeneity in these associations, we applied Multiscale geographically weighted regression (MGWR) for each MSA using the same model specification. MGWR relaxes the assumptions of traditional geographically weighted regression models (MGWR) by allowing regression coefficients to vary locally and at predictor-specific spatial scales. Adaptive bandwidths were selected for greenspace visitations automatically via cross-validation to balance local bias and variance, thereby defining the neighborhood contributing to each local estimate in a data-driven manner. Local coefficients of association (β) were estimated and visualized to examine spatial variation in the relationship between greenspace visits and poor mental health prevalence, using the formulation:

Together, the OLS and MGWR analyses provided a comprehensive view of both global (city-wide) and local (neighborhood-specific) relationships between greenspace engagement and mental health, supporting inference at scales relevant to urban public health and environmental planning.

The 10 largest U.S. metropolitan statistical areas by population were initially selected for analysis. However, Miami was excluded due to the absence of census tract-level mental health estimates in CDC PLACES. The analytic sample included 13,152 census tracts across the nine MSAs, without the inclusion of Miami (Table 2). Tract counts ranged from 948 (Atlanta) to 3,951 (New York City). Figure 1 displays the distribution of poor mental health prevalence and greenspace visits across metros. Median prevalence of poor mental health ranged from 12.9% in Washington, DC to 18.7% in Houston, while median greenspace visits ranged from 11.5 in Houston to 33.2 in DC. Notably, these distributions appeared inversely patterned. Cities such as Houston showed both the highest prevalence of poor mental health and the lowest greenspace visits, whereas Washington, DC displayed the opposite trend. Other metros generally fell between these extremes, suggesting a consistent ecological association between greenspace engagement and mental health that is further explored in regression models.

Adjusted ordinary least squares (OLS) models revealed significant negative associations between greenspace visits and poor mental health prevalence in most metros (Table 3, Figure 2). Higher levels of greenspace visitation were consistently associated with lower prevalence of poor mental health after adjusting for NDVI, and socioeconomic context, with effect sizes and model fit varying across MSAs. Model performance was strongest in several large metros, including Dallas, Washington DC, and Houston, where adjusted R² values exceeded 0.70, indicating that our model captured a substantial proportion of tract-level variation in mental health outcomes.

To distinguish the effects of actual greenspace engagement (i.e., visitation) vs. greenspace availability (i.e., NDVI-derived coverage), we conducted sensitivity analyses in which models included each exposure separately and jointly (Supplementary Table S1). When modeled alone, greenspace visitation was consistently and inversely associated with poor mental health across most MSAs. In contrast, NDVI-only models showed substantial heterogeneity across metropolitan areas, with greater variability in both the magnitude and direction of associations and generally less stable model fit. When both measures were included simultaneously, model performance improved modestly in some MSAs but remained unchanged or decreased in others. Overall, these findings suggest that greenspace visitation captures much of the variation relevant to the mental health benefits of greenspace, beyond what is explained by the greenspace availability measure alone.

In pooled analyses across all nine metros, Model A (without Metro fixed effects) indicated that each additional greenspace visit was associated with a 0.037% lower prevalence of poor mental health (β = −0.037, p < 0.001l Supplementary Figure S1), with the model explaining just over half of the variance (R² = 0.53; Supplementary Figure S2). When Metro fixed effects were added in Model B, model fit improved substantially (R² = 0.62; Supplementary Table S2), reflecting city-level context accounted for additional variability. Relative to Atlanta (reference), Phoenix (+2.7 pp), Houston (+1.3), and Philadelphia (+1.2) exhibited higher adjusted baselines of poor mental health prevalence, while Washington, DC (−1.1) and Chicago (−0.7) were lower. In this specification, the association between greenspace visits and poor mental health remained robust though slightly attenuated (β = −0.029, p < 0.001). Taken together, these pooled models confirm higher greenspace engagement is consistently linked with better mental health outcomes across metropolitan areas, while also highlighting meaningful differences in baseline prevalence of both poor mental health and greenspace visits between cities (Supplementary Figures S3, S4). In our within-metro models, the strongest association were observed in Houston (β = −0.113, p < 0.001) and Atlanta (β = −0.096, p < 0.001), where each additional greenspace visit was associated with over a 0.1% lower prevalence of poor mental health. In contrast, Philadelphia (β = 0.003, p = 0.29) and Washington DC (β = −0.07; p < 0.001) showed weak or inconsistent associations. Model explanatory power varied, with R² values ranging from 0.55 in NYC to 0.86 in Dallas, indicating robust model fit overall.

While OLS models identified metro-wide associations between greenspace visits and poor mental health prevalence, these estimates assume spatial homogeneity. To capture potential neighborhood-level variation, we applied multi-scale GWRs. The results of our MGWRs allowed to visualize the multiple spatial scales within each MSA. Our results revealed that the strength and direction of the greenspace-mental health association varied substantially across neighborhoods, with distinct patterns emerging in each MSAs (Figures 3, 4).

Because our aim was to test whether greater greenspace engagement corresponds with lower prevalence of poor mental health, inverse relationships (negative coefficients) are interpreted as favorable i.e., more greenspace visits is associated with improved mental health among the population. Conversely, positive coefficients suggest the opposite pattern, which we interpret as unfavorable. Across metropolitan areas, spatial heterogeneity followed a consistent contextual pattern rather than a purely geographic one. Stronger inverse associations between greenspace visits and poor mental health were concentrated in dense, higher socially vulnerability urban neighborhoods, whereas lower-density, lower-vulnerability suburban and exurban tracts generally exhibited weaker to mild or inconsistent relationships.

In Atlanta, the strongest beneficial relationships were seen in the southern tracts like Fulton and DeKalb Counties, areas that are characterized by historically higher social vulnerability, while northern suburban tracts with lower vulnerabilities showed weaker or no associations (14). Houston exhibited a similar vulnerability-stratified pattern with strong inverse associations concentrated in the urban core and weaker associations in outer suburban tracts (Montgomery and Fort Bend Counties). In Dallas, favorable associations were concentrated in socially vulnerable urban tracts, while outer suburban areas exhibited weaker effects. Los Angeles and Chicago showed more mixed patterns.

Patterns in Los Angeles and Chicago were more heterogeneous but followed similar contextual gradients. Tracts with greater access to large, centrally located public parks showed stronger associations, whereas lower-density coastal and valley areas showed minimal effects. In Chicago, tracts along the lakefront and in the central city In Chicago, tracts along the lakefront and in the central city.

New York City showed strong inverse associations in dense tracts surrounding major public greenspaces, particularly in parts of Manhattan, Brooklyn, and Queens, while outer borough tracts exhibited weaker or unfavorable associations. Philadelphia represented a notable exception in the overall association pattern of our analysis. Although some tracts near Fairmount Park showed favorable relationships, many surrounding neighborhoods exhibited weak or unfavorable associations, consistent with the null findings observed in metro-wide OLS models. In Washington, DC, associations were generally weak across the district, with modest localized effects near major parks like Rock Creek but minimal effect in the Eastern neighborhoods.

Local models were estimated for each census tract using an adaptive spatial kernel, such that the number of neighboring tracts contributing to each local regression varied according to tract density. The optimal bandwidth was selected automatically by minimizing the cross-validation (CV) score, balancing local bias and variance.

To complement the metro-specific models, we also conducted pooled analyses across all nine MSAs (Supplementary Table S2). In Model A, without accounting for variations between MSAs, greenspace visits were inversely associated with poor mental health (β = −0.037, p < 0.001; R² = 0.53). After adding Metro fixed effects in Model B, the association persisted but was modestly attenuated (β = −0.029, p < 0.001), while model fit improved considerably (R² = 0.62), underscoring the importance of accounting for city-level differences. These pooled results confirm the robustness of the overall relationship, even as the metro-specific models highlight substantial heterogeneity in both magnitude and direction of effects.

Across metros, a consistent urban-suburban gradient was observed. MGWR local fit surfaces were computed using a k-nearest-neighbor (kNN) approach, where local R² values were estimated by comparing observed and fitted outcomes within a fixed neighborhood (k = 200) around each census tract. The R² maps showed consistently high explanatory power across all metros (local R² ≥ 0.80), with stronger fit in dense urban neighborhoods and slightly lower but still robust fit in peripheral areas (Figure 4). This pattern suggests that greenspace engagement may be particularly salient for mental health in high-density areas, where exposure to stressors and social vulnerability is greater but public park access is more frequent and well connected. In contrast, suburban and exurban contexts may have larger quantities of greenspace, but less public transit, more fragmented accessibility, and a reliance on private yards, which may limit the degree to which greenspace engagement translates into measurable mental health benefit. Taken together, these findings indicate that the mental health benefits associated with greenspace engagement are highly context-dependent, with the strongest associations occurring in dense, socially vulnerable urban neighborhoods patterns that are obscured in city-wide average estimates.

This study offers several strengths that advance understanding of greenspace and mental health. By analyzing nine of the largest U.S. metropolitan areas, we provide one of the few multi-city perspectives on neighborhood-level associations, capturing both commonalities and differences across diverse urban-suburban contexts. The use of mobility traces to measure greenspace visits represents an important innovation, moving beyond static measures of greenness toward actual engagement, more closely tied to realized health benefits. Finally, the application of MGWR revealed spatial heterogeneity within each city, that global models cannot detect, offering insight into the highly localized nature of greenspace-health relationships. Together, these strengths provide a robust and scalable framework for future research at the intersection of urban design, behavioral data, and population health.

A few limitations must be noted. First, the cross-sectional design precludes causal inference, and results should be interpreted as associations rather than effects. Second, analyses were conducted at the census-tract level, raising the possibility of ecological fallacy when inferring individual outcomes; moreover, survey-based health estimates may themselves be biased if participation or reporting varies across neighborhoods, which may help explain unfavorable associations in some areas. Third, mobility data are subject to smoothing and potential underrepresentation of certain populations, which could bias estimates of greenspace use. Additionally, our dataset consists of cumulative annual exposures and outcomes, based on CDC estimates and Advan research, which may not be perfectly aligned. Future work should examine multi-year mobility trends. Finally, measures of greenspace quality, amenities, and safety were not included, factors that as previously mentioned could account for some of the spatial heterogeneity observed.

Future research should adopt longitudinal designs to assess temporal dynamics in greenspace use and mental health. For example, seasonality considering weather and post-COVID19 effects may impact the patterns of outdoor activities and social vulnerability factors. Incorporating finer measures of greenspace quality and accessibility, as well as equity-focused audits, would strengthen the evidence base for interventions. Linking mobility data with health services, such as nature-based prescription programs could further clarify how greenspace can be operationalized in public health practice.

In summary, this study provides new evidence that greenspace engagement is associated with mental health outcomes in U.S. metropolitan areas, and the strength of associations vary across space and are strongly conditioned by social vulnerability. These results highlight the importance of moving beyond measures of greenness to consider actual use, and of pairing greenspace planning that address the social and structural contexts shaping mental health.