This study was based on a cross-sectional study conducted within the AmeriSpeak panel (https://www.amerispeak.org), a probability-based survey designed to represent the US population. The AmeriSpeak panel utilized a stratified two-stage sampling design (details in supplement) [20]. Data collection was completed between October and November 2019 by the National Opinion Research Center (NORC) at the University of Chicago [21]. ACT24 recalls were administered online using a link sent via email or text message. The present study included participants aged 20 to 75 who completed a short online survey and a previous-day recall on a randomly selected day. Participants who completed the first survey were invited to a second recall on a different randomly selected day, 1–2 weeks later. Our analytic sample included participants who completed the survey and provided at least one valid ACT24 recall. The NORC research ethics review board approved the study protocol, and all participants provided written informed consent. Participants received up to $30 for completing study activities.

From the total 2877 participants contributing a total of 4575 recalls, we excluded recalls based on any one of the following criteria: (1) participants reported more than 1 h per day of unknown time (gaps; private unreported time; n = 257 recalls), (2) more than 2 h per day of overlapping time (n = 93 recalls), and (3) recalls 0 active hours of PA (n = 87 recalls) [22]. After exclusions, the survey sample included 2625 participants with at least one valid recall (n = 2226 first recalls, n = 1912 second recalls).

The AmeriSpeak panel utilized a stratified two-stage sampling design [20]. Briefly, the first stage involved selecting National Frame Areas as primary sampling units and census tracts or block groups as secondary sampling units [20]. To approximate a nationally representative sample of the US population, sample weights were developed for each recall based on the final weights of AmeriSpeak panelists, which adjusted for panel selection probabilities, non-response, and population coverage. The final weights were adjusted to external population totals based on age, gender, race/ethnicity, housing tenure, telephone status, and Census Division [22]. The study-specific sampling weights were calculated for each recall day, aiming to address potential variations in PA levels throughout the week. Lastly, the weights were normalized to ensure that each day of the week contributed equally to the analyses. This sampling strategy allowed us to obtain a representative sample of US adults and estimate PA levels in the population across different days of the week [20].

Data were collected using the self-administered ACT24 previous-day recall, which asked participants to report their PA via smartphone, tablet, or computer for the previous day (midnight-midnight). Recalls with less than one hour of missing or unknown activity (e.g., gap time, private, prefer not to say time) and at least 22 h of total reported time were considered valid. ACT24 has shown greater accuracy and higher internal correlations than a traditional questionnaire in large epidemiologic studies. Participants could select from 170 individual activities grouped into 14 major categories linked to the Compendium of Physical Activities [23, 24]. Each activity was assigned a score to estimate energy expenditure in metabolic equivalent of tasks (METs) mins/day. Reportable PA behaviors included light (1.8–2.9 METs), moderate (3.0–5.9 METs), and vigorous intensity (≥ 6.0 METs) activities [25]. Subsequently, three PA outcomes (hours/day [h/d]) were created: LPA, MVPA, and total active time (including LPA and MVPA). Consistent with the American Time-use Survey [26], time-use PA outcomes were calculated for seven domains-specific PA (h/d), including leisure, work, household, transportation, personal, and other activities.

We utilized a widely accepted segregation index to measure the degree to which different racial and ethnic groups are isolated from one another within their county. The calculated isolation index values were merged with the respective participant’s Federal Information Processing Standards (FIPS) codes at the county level.

$$\text= _^\left[\frac_}\right]\times \left[\frac_}_}\right]$$

where xi is the number of residents who are NH Black/Hispanic individuals, the number of residents who are NH Black/Hispanic in census tract i, ti is the total population in tract i. X is the number of NH Blacks/Hispanic people in the county. The isolation index measures the degree to which members of a minority group are primarily exposed to other members of the same group, based on residence in a specific area, as opposed to members of other racial or ethnic groups [27, 28]. In this study, we calculated the index based on the county-level of residence. A higher index value indicates a greater probability of contact between members of the same minority group (high segregation) and a lower probability of contact between members of that group and individuals from other racial or ethnic groups. This analysis includes the following racial/ethnic isolation segregation indices: NH Black versus others (including all other race/ethnicity groups such as NH White, Hispanic, Asian, and others), and Hispanic versus others (including all other race/ethnicity groups such as NH White, NH Black, Asian, and other racial and ethnic groups).

Our analysis incorporated individual-level covariates that could influence the associations between the segregation indices and the PA outcomes. Demographic covariates included age, gender (male/female), educational attainment (high school or less, some college/associate degree, bachelor’s degree, or graduate degree), and household income: “ < 50,000,” “50,000–99,000,” “100,000–149,000,” and “150,000 + ” [22]. BMI was coded as a continuous variable.

Additional neighborhood-level variables could be associated with the exposure and outcome variables. County-level poverty was indicated as the percentage of households living below the poverty line within each county (expressed in quartiles), with a higher quartile indicating higher poverty levels. Region included four census regions: Northeast, Midwest, South, and West.

Participants’ characteristics were summarized for the overall sample and stratified by gender. Each isolation segregation index was modeled as a continuous variable. For continuous variables, we calculated weighted means and standard deviations (SD). To describe PA behavior, we used data from one recall, up to two recalls per person, focusing on PA behavior for days rather than individuals. We reported the mean PA outcomes and domain-specific PA for the overall population and stratified by gender.

We conducted weighted multivariate linear regression to examine the association between isolation segregation indices and PA outcomes (h/d, including LPA, MVPA, and total active time) stratified by racial and ethnic groups. We set two-tailed p-values < 0.05 as the threshold for statistical significance across all analyses. Furthermore, we investigated whether the association is moderated by gender for NH Black and Hispanic groups separately. Lastly, we separately examined the associations between segregation and each domain-specific PA for NH Black and Hispanic groups. All analyses were adjusted for covariates. Additionally, we calculate NH Black and Hispanic isolation indices using NH White adults as the reference group for the sensitivity analysis. To account for the complex sampling design, we performed analyses using the “Survey” package in the R statistical software (version 4.2.2; www.r-project.org).