Participants and procedures
Participants for the present analyses were part of the Teen Environment and Neighborhood (TEAN) Study examining the relation of built environmental factors with physical activity and dietary behaviors . Healthy adolescents between the ages of 12–16 years and one of their parents were recruited from the Baltimore, MD-Washington, DC and the Seattle-King County, WA metropolitan areas in 2009–2011. Participants were recruited from 447 census block groups and were evenly stratified across census block groups representing high or low neighborhood walkability and high or low median household income, resulting in four (2 × 2) study design quadrants . Measurement occurred only during the school year and assessments of participants were balanced by season across study quadrants.
Households were contacted by phone, and eligible participants were invited to participate in the study. The overall participation rate was 36%, which was comparable across study quadrants. Participants were phone screened by the research staff and subsequently excluded from the study if they had any physical, medical, or cognitive impairments that would limit their physical activity, affect their dietary behaviors, or impact their ability to complete measures. Eligible participants were mailed consent and assent forms, which were followed up with a phone call by a research assistant to answer questions. After informed consent and assent forms were received, adolescent participants were asked to wear an accelerometer and Global Positioning System (GPS) tracker for 7 days during waking hours. Participants earned US $40 financial compensation for completing study procedures. The study was approved by the Institutional Review Board at San Diego State University and the University of California, San Diego.
A total of 928 adolescents participated in the TEAN study. Analyses in the present paper involving physical activity/sedentary variables excluded participants who did not receive a GPS tracker or never recorded any data (n = 130), whose home address was not available in the geocoding database (n = 29), or who did not provide their school’s name/address or were homeschooled (n = 93). Adolescents who did not wear both the GPS and accelerometer devices together for ≥1 valid school day and ≥ 1 valid weekend day (n = 204) were also excluded to obtain estimates of behaviors on both weekday (school) and weekend days. Valid days were defined as those with ≥8 h of concurrent data from both devices, and school days were defined as weekdays during which the participant spent ≥200 min at their school as measured by the GPS. Analysis involving dietary variables excluded participants who did not complete dietary recalls on ≥1 school day and ≥ 1 weekend day (n = 139). Therefore, the current sample comprised 472 adolescents for physical activity and sedentary variable analyses, and 789 adolescents for dietary analyses.
Adolescents self-reported demographic information including their age, sex, and race/ethnicity (dichotomized as white non-Hispanic versus non-white or Hispanic). Parents reported the highest level of education attained by any adult in the household (dichotomized as college degree versus less education), their marital status (married/living with partner versus other), and their approximate annual household income.
Adolescents self-reported their height and weight. Participants were asked to take the measurements at home (with instructions provided) or use doctor or school measurements taken within the last month.
GPS tracking and location assignment
Participants wore a GlobalSat DG-100 GPS tracker (GlobalSat, New Taipei City, Taiwan), with latitude and longitude collected every 30 s when a signal was attainable. The DG-100 tracker is considered an accurate device for measuring daily location patterns and has good spatial accuracy; the metrics of the DG-100 GPS metrics have been reported in other epidemiological studies . Each participant’s home and school addresses were geocoded and incorporated into ArcGIS (ESRI, Inc., Redlands, CA) to create a 50-m circular buffer around the point resulting from geocoded the home address and a 15-m buffer around the geocoded school parcel. These buffer sizes were selected to minimize misclassification while considering potential errors caused be satellite interference, and are generally consistent with previous physical activity GPS studies [13, 34, 35]. Spatial analyses (methods published previously in ), were performed in PostgreSQL (PostgreSQL Global Development Group, Berkeley, CA) to classify each GPS point by location: at home (within the home or surrounding buffer), at school (within the school parcel or surrounding buffer), or all “other” locations (i.e., any location other than home and school). In this case, transport/trips would be classified in the “other” location.
Physical activity and sedentary time
Hip-worn ActiGraph accelerometers were used (models: 7164, 87.8%; GT1M, 8.0%; GT3X, 3.4%). Non-wear periods were defined as 30+ minute bouts of consecutive epochs with 0 accelerometer counts and subsequently excluded from analyses. GPS and accelerometer data were integrated to 30-s epochs and merged based on timestamps using a nearest neighbor approach (within up to a 30 s difference), and epochs with periods of missing GPS data or accelerometer non-wear time were removed from the dataset.
The Evenson cut points , which have excellent classification accuracy for measuring physical activity and sedentary time , were applied to physical activity counts within each 30-s epoch to classify MVPA and to 60-s epochs to classify sedentary time based on vertical axis accelerometer counts. The shorter epoch period for classifying MVPA has been shown to have greater validity than using 60-s epochs in youth . The combined GPS and accelerometer data were then processed to create location-specific physical activity time, sedentary time, and sedentary bout patterns. For the bout pattern scoring, each epoch was first defined as occurring at home, at school, or at “other” locations. Sedentary bouts were defined as periods of sedentary time lasting ≥1 min. A sedentary bout ended when the epoch had an accelerometer count > 100 (no tolerance) or when the location (home, school, or other) changed for ≥2 consecutive GPS epochs. Next, location-specific sedentary bout pattern variables were scored, including: 1) total sedentary time; 2) prolonged time in sedentary bouts lasting ≥30 min; 3) mean bout duration, which represents the mean duration of all sedentary bouts; 4) period, the average duration between the end of a sedentary bout and the start of another within the wear time period spent at a location (calculated using the same methods described above for sedentary bouts); and 5) alpha, which represents an individual’s distribution/slope of sedentary bout lengths based on a power law function [39, 40]. Alpha is unit-less, with lower values reflecting more time in prolonged (longer) sedentary bout lengths [39, 40].
For the school location, variables were derived for school days only (e.g., average minutes/day of MVPA across school days). For the home and “other” locations, variables were derived for a “weighted week”, calculated as ([mean daily values across school days*5] + [mean daily values across non-school days*2]) ÷ 7, similar to previous protocols . This was done to provide a better representation of a full week when the number of actual wear days on weekdays and weekends was imbalanced. Participants were required to have spent an average of ≥30 min/day in a location category (i.e., home, school, or other) as measured by the combined accelerometer wear time and GPS location data, to calculate each location-specific physical activity and sedentary time variable. Location-specific variables were recoded as missing for participants who did not spend ≥30 min/day in that location on average across days. These location wear time inclusion criteria were stricter than what were used in our previous analyses of location-based MVPA . The stricter inclusion criteria were selected to better capture a representation of typical behavior in a location and eliminate momentary changes in behavior as adolescents transitioned across locations. It specifically aimed to improve estimates of the percent of time in the home location that was spent sedentary but was also likely to improve estimates of MVPA.
Trained interviewers attempted three 24-h dietary recalls (two weekdays and one weekend day) via telephone for each participant on unannounced, random, non-consecutive days during the school year. Dietary data were collected and analyzed using the Nutrient Data System for Research software (version 2010) developed by the University of Minnesota Nutrition Coordinating Center . The computer-assisted, interviewer-administered recalls were facilitated by an automated, four-stage, multi-pass technique [42, 43], which has acceptable validity for assessing dietary intake in youth as young as 8 years old  and has been implemented in national US dietary surveys . Both the interviewers and participants had access to posters that displayed two-dimensional illustrations of cups, spoons, bowls, and other common food shapes to assist with portion size estimation.
The dietary behaviors of interest included daily energy intake (kcal), as well as other daily diet quality indicators such as added sugar (g), sodium (mg), fruits and vegetable (servings), high calorie beverages (number), sweet and savory snacks (number), whole grains (servings), fiber (g), fat calories (kcal), and saturated fat calories (kcal) . Percent of fat and saturated fat were also calculated (fat kcal ÷ total kcal from models not per 100 kcal of energy intake*100). During recalls, participants were asked to report the location in which they consumed food/drinks during each eating episode (regardless of where it was purchased or prepared), which was used to derive location-specific dietary behaviors for at home, at school, and at “other” locations. For example, school-based dietary behaviors were defined as any food consumed at school regardless of where it was bought or made. The “other” location comprised all other locations, for example at work, after school programs, deli/take-out/store, restaurant/cafeteria/fast food, friend’s home, and party/reception/sporting event locations. For the school location, variables were derived for school days only, whereas for the home and “other” locations, variables were derived for a weighted week using the same equations that were used to calculate the weighted physical activity variables.
All models were mixed-effects linear regression models, fitted with the “MIXED” command in SPSS version 24 (IBM SPSS Statistics, IBM Corporation), and included a random census block group intercept to account for the nesting of participants within census block groups. All models were adjusted for neighborhood walkability (low vs high), census-based median household income (low vs high), and adolescent and household characteristics including age, sex, race/ethnicity, and highest parental education level. All physical activity and sedentary models were adjusted for ActiGraph model, number of school and weekend days the participant wore the accelerometer, and average minutes/day of wear time. Models investigating dietary behaviors were additionally adjusted for participant height and weight and number of days of dietary recall.
To determine the magnitude of differences in physical activity and dietary behaviors across locations (i.e., home, school, “other”; Aim 1), location was entered as a categorical repeated-effects independent variable, and separate models were investigated for each behavior dependent variable using the aforementioned covariates. A second set of physical activity and sedentary time models evaluated differences in the proportion of time in each location that was spent in MVPA, total sedentary time, and time in 30+ minute sedentary bouts standardized per 60 min of wear time. Such proportional variables were not calculated for mean bout duration, period, and alpha since these variables are less likely to be affected by total time spent in a location. A second set of dietary behavior models evaluated the proportional differences in dietary behaviors as a function of energy intake in each location using nutrient densities (i.e., dietary value per 100 kcal of energy intake in each location). All Aim 1 models produced adjusted mean estimates of obesogenic behaviors in each location, as well as standard errors for these estimates. Post-hoc multiple comparisons were computed to determine if there were significant differences in these estimates and the 95% confidence intervals were reported. A more conservative p-value of p < .01 was used to provide evidence for significant differences between locations due to the large number of tests.
Associations of each behavior between each pair of locations were tested to investigate the extent to which a participant with more healthy behaviors in one location engaged in more healthy behaviors in other locations, relative to other participants (Aim 2). Each behavioral variable at “other” locations (e.g., “other” MVPA) was regressed on the same behavioral variable at home (e.g., home MVPA) and at school (e.g., school MVPA), and each behavioral variable at home was regressed on the same behavioral variable at school in mixed models. Both the independent and dependent behavioral variables were standardized to have a mean of zero and standard deviation of one to derive standardized regression coefficients. The magnitude of the standardized associations of behaviors between locations were interpreted as small (< .3) medium (.3–.5), and large (>.5) .