Social-Infrastructure Time Use (STU)

• Social-Infrastructure Time Use (STU) is a spatially resolved framework that quantifies how neighborhoods engage with venues like grocery stores, sports centers, and cultural sites.
• The methodology combines anonymized mobile-device data with census information to capture key metrics, including time spent per user, visit depth, activity diversity, and spatial inequality.
• Empirical findings reveal spatial clustering, temporal growth trends, and demographic disparities in engagement, providing actionable insights for urban planning, public health, and social equity.

Searching arXiv for the specified paper to ground the article in the cited source. Social-infrastructure Time Use measures (STU) are a spatially resolved framework for measuring how neighborhoods use and experience social infrastructure places over time, with an emphasis on length and depth of engagement, activity diversity, and spatial inequality (Wang et al., 18 Aug 2025). In this formulation, social infrastructure places—such as grocery stores and food markets, sports and recreation venues, dining places, arts and entertainment sites, religious organizations, and other places that support social interaction—are treated not merely as destinations but as settings in which people spend time, interact, and build social ties. STU was developed to fill a gap between national time-use surveys and neighborhood-scale spatial analysis by providing a national, week-by-week, geographically detailed dataset spanning January 2019 through May 2024 across 49 continental U.S. states plus Washington, D.C. (Wang et al., 18 Aug 2025).

1. Conceptual basis and motivation

STU was introduced in response to the limited spatial resolution of existing national time-use sources and the difficulty of scaling self-reported survey approaches. The framework is explicitly motivated by the claim that time spent at social infrastructure places provides a useful lens on urban life, public health, and neighborhood inequality, because time is a scarce resource and because time spent in these places is tied to social connectedness, neighborhood well-being, and equity (Wang et al., 18 Aug 2025).

The paper positions STU as a bridge between time geography, mobility research, urban planning, and public health (Wang et al., 18 Aug 2025). Existing national time-use sources such as the American Time Use Survey (ATUS) are described as powerful for understanding how people allocate time across activities, but not spatially resolved at the neighborhood level and therefore unable to show how time use varies across places. Built-environment and health studies are described as often relying on self-reported surveys that are hard to scale and may suffer from recall bias. On the accessibility side, the paper notes a growing literature on travel-time accessibility, but much less work on dwell time or actual time spent once people arrive at places (Wang et al., 18 Aug 2025).

Within this framing, the central analytic move is to treat census tracts as neighborhood proxies and to construct measures that capture four distinct questions: how long residents spend at social infrastructure places, how deeply they engage with those places, how diverse their time use is across types of social infrastructure, and how unequal that time use is across neighborhoods within larger regions (Wang et al., 18 Aug 2025). This suggests a shift from purely locational or access-based analysis toward a use-based account of neighborhood social infrastructure.

2. Scope of social infrastructure and dataset architecture

The STU dataset is based on anonymized and aggregated mobile-device foot traffic data from Advan, accessed via Dewey Data, collected continuously from January 2019 onward and updated monthly (Wang et al., 18 Aug 2025). The panel includes about 32 million devices, roughly 15% of U.S. phone users. Geographic coverage includes 73,642 census tracts, 18,773 county subdivisions, 3,242 counties, and 939 metropolitan areas, with temporal coverage organized as weekly time series from January 2019 through May 2024 (Wang et al., 18 Aug 2025).

Neighborhood sociodemographic characteristics are drawn from American Community Survey (ACS) 1-year estimates for 2019–2022. Because geographic boundaries changed over time, the study used NHGIS geographic crosswalks and interpolation weights to align post-2020 ACS data to 2019 vintage boundaries (Wang et al., 18 Aug 2025). This boundary harmonization is integral to longitudinal comparability at neighborhood scale.

Social infrastructure places are identified using six-digit NAICS codes corresponding to venues that support social and leisure activities and are matched to ATUS/GHD activity categories (Wang et al., 18 Aug 2025). The major activity categories described are grocery shopping, physical exercise/sports/recreation, attending sporting or recreational events, eating and drinking, arts and entertainment, religious activities, and consumer goods purchases (Wang et al., 18 Aug 2025). The stated rationale is that these places function as “third places,” social gathering spaces, or culturally meaningful venues that support exposure to others, social capital, and community connection.

The publicly released repository contains weekly tabular datasets, Python code for processing and reproducing the measures, panel coverage rates, and interactive maps (Wang et al., 18 Aug 2025). Each table includes common attributes such as GEOID, Timestamp, Per_User_STU, Per_Visit_STU, Diversity, and Gini. The data are also broken out by activity type, with measures for “all” social infrastructure and for each ATUS category, using suffixes like _sector or category-specific names (Wang et al., 18 Aug 2025).

3. Measurement framework and formal definitions

The conceptual structure of STU is organized around three neighborhood dimensions: length and depth of engagement, activity diversity, and spatial inequality (Wang et al., 18 Aug 2025). These dimensions are operationalized through four main measures and a set of aggregated derivatives.

Measure	Interpretation	Dimension
Per User STU	Average weekly time spent by neighborhood residents at social infrastructure places, normalized by the number of mobile devices in the panel	Length of engagement
Per Visit STU	Average dwell time per visit at social infrastructure places	Depth of engagement
Shannon diversity of STU	Diversity of neighborhood time use across social infrastructure activity types	Activity diversity
Gini index of STU (GSTU)	Inequality in STU across neighborhoods within a larger geographic unit	Spatial inequality

For neighborhood $i$ and activity type $k$, weekly time spent is denoted $T_{i,k}$ and is computed by aggregation over all POIs $j$ of type $k$:

$T_{i,k} = \sum_{j \in N_k} V_{i,j} \times L_j$

where $V_{i,j}$ is total visits from neighborhood $i$ to POI $j$, $N_k$ is the set of POIs of activity type $k$0, and $k$1 is estimated dwell time at POI $k$2 based on bucketed dwell-time distributions (Wang et al., 18 Aug 2025).

Per User STU for neighborhood $k$3 is defined as the average weekly time spent at all social infrastructure places, normalized by the number of devices $k$4:

$k$5

The paper states conceptually that

$k$6

where $k$7 is the set of activity categories (Wang et al., 18 Aug 2025). This is the foundational measure of length of engagement.

Per Visit STU is the average time spent per visit across social infrastructure places:

$k$8

where $k$9 is the number of visits from neighborhood $T_{i,k}$0 to category $T_{i,k}$1, and $T_{i,k}$2 is the total number of visits from neighborhood $T_{i,k}$3 across all categories (Wang et al., 18 Aug 2025). This is the foundational measure of depth of engagement.

To measure diversity across activity sectors, the paper adapts Shannon’s Diversity Index:

$T_{i,k}$4

where $T_{i,k}$5 is the number of distinct social infrastructure sectors visited by residents of neighborhood $T_{i,k}$6, and $T_{i,k}$7 is the proportion of total STU time spent in sector $T_{i,k}$8 relative to all sectors (Wang et al., 18 Aug 2025). Higher values indicate both more sectors used and a more even distribution of time across sectors.

To quantify inequality across neighborhoods within a larger region, the study computes a Gini-like measure:

$T_{i,k}$9

where $j$0 is the number of neighborhoods in the larger area, $j$1 is population share of neighborhood $j$2, $j$3 is per-user STU for neighborhood $j$4, $j$5 is cumulative share of per-user STU up to neighborhood $j$6, and neighborhoods are ordered by increasing $j$7 (Wang et al., 18 Aug 2025). Higher GSTU is interpreted as greater inequality in STU distribution.

4. Foot-traffic methodology and preprocessing

The foot-traffic methodology is based on the Advan Weekly Patterns product (Wang et al., 18 Aug 2025). Each point of interest has a brand/place identifier, NAICS code, street address, latitude/longitude, census geographic code, and open and close dates. A visit is recorded when a device ping falls inside the geometric polygon of a POI, and the duration of a visit is inferred from the length of time the device remains within the POI polygon (Wang et al., 18 Aug 2025).

Visitor home locations are available only at the census block group level for privacy reasons. Home location is inferred using six weeks of nighttime data, defined as 6 pm to 7 am, together with a minimum evidence threshold (Wang et al., 18 Aug 2025). For tract-level work, block-group measures are aggregated up to tracts using FIPS-based matching. This privacy-preserving architecture constrains residential geolocation precision while still permitting tract-level aggregation.

The raw data provide bucketed dwell time rather than exact duration. The buckets are 0–4 minutes, 5–10 minutes, 11–20 minutes, 21–60 minutes, 61–120 minutes, 121–240 minutes, and $j$8 minutes (Wang et al., 18 Aug 2025). For each POI, total time is estimated from visitation frequency multiplied by the median dwell time for each bucket. The paper notes that work-related and non-work-related behavior cannot be fully distinguished in the longest bucket, but states that this is unlikely to materially affect the analysis because it represents about 10% of overall visitation frequency on average (Wang et al., 18 Aug 2025).

Two major preprocessing issues are addressed. First, co-located POIs can generate duplicated records across overlapping or co-located POIs, so neighborhood total STU is evenly distributed across them. Second, because the device panel evolves over time, STU measures are normalized by monthly panel size to reduce bias from changes in sampling rate (Wang et al., 18 Aug 2025). These steps are methodological safeguards against duplication artifacts and nonstationary panel coverage.

The dataset is also explicitly multi-scale. Neighborhood-level measures are aggregated to county subdivisions, counties, and metropolitan areas. Per User STU is aggregated using device counts $j$9, Per Visit STU is aggregated using visit counts $k$0, and Shannon diversity is also aggregated using device counts (Wang et al., 18 Aug 2025). This weighting scheme preserves the scale-specific interpretation of each measure under aggregation.

5. Empirical patterns across time, space, and neighborhood characteristics

The paper reports strong spatial clustering for the foundational measures. Global Moran’s I equals 0.359 for Per User STU and 0.256 for Per Visit STU, with $k$1 in both cases (Wang et al., 18 Aug 2025). The reported interpretation is that neighborhoods with similar levels of STU are geographically clustered rather than randomly distributed, reflecting regional disparities across the United States.

The diversity and inequality maps show an inverse relationship. Eastern and northeastern neighborhoods tend to show higher spatial inequality and lower STU diversity, whereas southeastern neighborhoods tend to show higher STU diversity and lower inequality (Wang et al., 18 Aug 2025). The authors interpret this as suggesting that access to a more diverse set of social infrastructure places may support more equitable time use patterns. This suggests that diversity of use and equity of distribution may be analytically linked, although the paper states this as an interpretation rather than as a causal identification result.

The dataset supports analysis at weekly, monthly, and yearly scales. Per User STU across neighborhoods increased over time from 2019 to 2023, and the fitted log-normal distribution’s scale parameter rises from about 86.01 in 2019 to 271.21 in 2023 (Wang et al., 18 Aug 2025). The weekly panel is described as enabling interrupted time series or event-based analysis of shocks such as pandemics, hurricanes, policy changes, and infrastructure developments.

Neighborhood-level Per User STU is reported as best approximated by a log-normal distribution among candidate distributions tested. The paper compares normal, log-normal, Weibull, exponential, Gamma, power-law, and chi-squared distributions using the Kolmogorov–Smirnov (K-S) test, and reports the log-normal fit as strongest, with KS statistics roughly ranging from 0.169 to 0.280 in the yearly plots (Wang et al., 18 Aug 2025). At the activity-category level, art and entertainment, eating and drinking, and consumer activities show higher median time use, while art and entertainment, events, and sports show greater skewness, meaning time use is concentrated in a small subset of neighborhoods (Wang et al., 18 Aug 2025). Routine categories are described as having lower skewness and more even distributions.

The paper also tests whether neighborhood STU reproduces known demographic patterns documented in ATUS and other studies. Urban neighborhoods have higher Per User STU than rural neighborhoods, with a significant urban-rural difference given by K-S statistic 0.531 and $k$2 (Wang et al., 18 Aug 2025). Neighborhoods with higher Black population shares tend to have lower total STU, with Per User STU coefficient $k$3 and Per Visit STU coefficient $k$4, while neighborhoods with higher White population shares show higher STU, with Per User STU coefficient $k$5 and Per Visit STU coefficient $k$6 (Wang et al., 18 Aug 2025). Higher female share is associated with lower total Per User STU, coefficient $k$7, and higher share of residents with a bachelor’s degree or higher is associated with higher STU, with Per User STU coefficient $k$8 and especially strong association for sports/recreation, coefficient $k$9 (Wang et al., 18 Aug 2025). These findings are described as broadly consistent with ATUS-based literature on racial, gender, and education disparities in leisure and physical activity.

6. Validation, analytical uses, and interpretive cautions

Validation proceeds in three main ways: sampling representativeness of the mobile panel, benchmarking against ATUS, and consistency with known demographic patterns (Wang et al., 18 Aug 2025). For representativeness, panel device counts are compared to ACS population size at tract, county, and state levels. Panel coverage varies spatially, with lower representation in the western and northwestern U.S.; correlation between device counts and population is strong at county and state levels, exceeding 90% alignment; at the tract level, alignment is weaker and varies by year, roughly 54% to 72%; and 2019 shows the most variability and should be treated cautiously for tract-level analyses (Wang et al., 18 Aug 2025). The paper therefore recommends normalizing STU by monthly panel size and checking representation bias when using the data.

Benchmarking against ATUS for 2023 focuses on out-of-home activities. The comparison maps ATUS “average time per day per civilian population” to Per User STU and ATUS “average daily time for engaged individuals” to Per Visit STU (Wang et al., 18 Aug 2025). STU values generally fall within the ATUS range for corresponding activity categories, and rankings across activity types are broadly consistent. The paper notes that eating and drinking may be underrepresented in STU because ATUS includes at-home time, while STU captures only out-of-home time at social infrastructure places. It also notes that Per Visit STU is slightly lower than the ATUS “engaged individuals” measure, likely because ATUS sums all activity time across the day, while STU records only a visit-specific duration (Wang et al., 18 Aug 2025).

These validation results bear directly on common interpretive errors. STU is not a substitute for all-purpose time-use accounting, because it is restricted to time at social infrastructure places and derives duration from foot-traffic records rather than diary reconstruction. Nor is it a pure accessibility indicator, because it measures realized time use rather than travel-time accessibility alone. The paper’s treatment of the longest dwell-time bucket also marks a substantive limitation: work-related and non-work-related behavior cannot be fully separated for visits longer than 240 minutes (Wang et al., 18 Aug 2025). A plausible implication is that STU is most robust when interpreted as a behavioral measure of neighborhood engagement with social infrastructure rather than as a direct measure of motivation, preference, or welfare.

The practical applications emphasized in the paper center on public health, environmental and contextual research, and neighborhood well-being and planning (Wang et al., 18 Aug 2025). Because STU captures how neighborhoods spend time in socially meaningful places, it can be linked to mental health, physical activity, social connection, depression prevalence, and health equity. The authors explicitly suggest pairing STU with public-health datasets like CDC PLACES. They also state that STU can be linked to climate vulnerability, neighborhood built environment, urban form, and accessibility to social infrastructure, and note the compatibility of tract-level measures with datasets such as the Climate Vulnerability Index (Wang et al., 18 Aug 2025). In planning and policy contexts, STU is presented as a means to identify neighborhoods with low access to diverse social infrastructure, regions with highly unequal time use, groups or areas underutilizing social infrastructure, and places where targeted interventions could improve social connectedness and well-being.

Taken together, STU is presented as a national, multi-scale, time-resolved measurement system for how people use social infrastructure places in U.S. neighborhoods (Wang et al., 18 Aug 2025). Its contribution lies in combining mobile-device foot-traffic data, activity-based classification of POIs, and neighborhood sociodemographic data to produce Per User STU, Per Visit STU, STU diversity, and STU inequality. The reported result is a dataset that reveals spatial clustering, temporal growth, and demographic disparities in social-infrastructure time use while offering richer spatial resolution and temporal continuity than traditional surveys (Wang et al., 18 Aug 2025).