The study was conducted between February and April 2013 in the municipality of Winterthur, which is the sixth largest city in Switzerland with a population of over 100,000 residents. Winterthur is located in the German-speaking part of Switzerland and represents an urban area which is representative in relation to demographic values such as unemployment rate, foreign nationals or social assistant rate [24]. All teachers from a public school in Winterthur in charge of a sixth grade class (n = 47) were invited to participate in the present study. If they agreed to participate, classes were visited and briefed about the study. Participation was voluntary and children were provided with an information letter for their parents. Parental written informed consent was obtained along with child consent. The ability to engage in usual everyday PA and the absence of any severe disabilities were the only inclusion criteria. The study was approved by the ethics committee of ETH Zurich (EK 2012-N-62).

PA was measured every ten seconds with a tri-axial accelerometer (GT3X, Actigraph, Pensacola, FL, USA). Given that cut-off points for different activity levels are based on the vertical axis, only data from this axis was included in the analysis. A short sampling interval was used to accurately capture children’s PA pattern, which is characterized by short, intermittent bursts of activity [25]. Simultaneously, a GPS receiver (BT-Q1000XT, QStarz, Taipei, Taiwan) recorded the geographical location at ten second intervals whenever there was sufficient satellite signal (Circular Error Probability CEP (50%) < 3m). A recent study by Schipperijn and colleagues [26] showed that this device has an acceptable dynamic spatial accuracy for use in larger population studies with a data collection period of seven or more days. In their experiment, 79% of data fell within 10m of the expected location.

All participating classes were visited during a regular school hour one day before the start of the measurement to distribute the devices among children and to provide them with updated information. Each child received an elastic belt equipped with an accelerometer and a GPS sensor. Participants were asked to wear the belt around the waist from waking time to bedtime on seven consecutive days starting the next day. Each child was given a charger for the GPS device and was instructed to charge the device during night when asleep. Children were asked to complete a small diary throughout the week reporting times they woke up and went to bed and times and reasons when monitors were not worn. In addition, age, gender, nationality and home address were assessed and height and weight were measured using a measurement tape and a digital scale (Beurer GS 12, Beurer GmbH, Ulm, Germany) to the nearest 0.5cm and 0.1kg, respectively. Nationality was defined as the number of parents with a Swiss nationality and subdivided into three categories (e.g. none, one or two of the parents with Swiss nationality).

Each participant’s GPS and accelerometer data was manually reviewed to ensure that both files contained adequate data. The GPS and accelerometer files were then matched by time using existing software (Actilife 6.5.2, Actigraph, Pensacola, FL, USA) producing a measure of activity and location for each recorded GPS point.

The processing of the matched data was performed using MATLAB R2012a (MathWorks, Massachusetts, USA) and R v2.15.2 (R Development Core Team, Vienna, Austria). Intervals with > 60 minutes of consecutive zero activity counts were classified as non-wear time and excluded from analysis [27]. Activity records > 5461 counts per ten seconds were identified as outliers [28] and replaced with the mean of the previous and the following value. Each data point was then classified into one of the four intensity categories sedentary (< 101 counts per minute (CPM)), light (101–2295 CPM), moderate (2296–4011 CPM) or vigorous (≥ 4012 CPM), based on age-appropriate cut-off points [29, 30]. Furthermore, the data was processed by visual observation as well as automatic identification of invalid GPS data points using extreme changes in distance and invalid values of altitude and removing them from data.

The setting-based categorisation of the matched data points was conducted in ArcGIS 10 (ESRI, Redlands, CA, USA) using further geospatial data (land use data, street network and points-of-interest-file), which was provided by the Land Surveying Office of Winterthur. Based on prior studies [15, 19, 31] and the ability to clearly assign each recorded GPS point to a precise location by using ArcGIS, we chose to define seven activity settings: home, own school, other school, recreational facility, street, other, and outside. The definitions and methods of generating the seven settings are presented in Fig 1.

Each participant’s data points were imported into ArcGIS and plotted on a separate layer. To assign a data point to one of the seven activity settings, we used the point-in-polygon overlay in which each participant’s point layer was overlaid on five different polygon layers defined in Fig 1. This overlay was carried out in a hierarchical order by firstly assigning points that fell outside Winterthur, followed by points that fell within the home buffer, school polygons, recreational facility polygons and street polygons. Lastly, all remaining data points were categorized as other. For school, recreational facility and street polygons, we chose a buffer zone of ten meters to take the measurement error of the GPS devices into account [23, 26, 32].

Only matched wear-time data was used for analysis. To best capture all PA settings where children spent time and because no clear standard criteria for the amount of measurement time per observation day exists, we determined an average value from a criterion of one day with at least one hour [19] and at least three days with a minimum of four hours [28]. Thus, the children had to provide at least one day with four full hours of matched data to be included in the analysis.

As outcome measures, total time per week (in hours) and MVPA per week (in minutes) were calculated for each setting and participant. Moreover, the proportion of time spent in MVPA (in %) out of the total time spent in each setting was calculated.

Statistical analyses were performed in R. Descriptive statistics were used to calculate frequency distributions (number (n) and proportion (%)), mean and standard deviation (SD), or median and interquartile range (IQR) to describe the general characteristics of the study population. As the outcome measures were not normally distributed, median and IQR were used to present total time, the amount of MVPA and proportion of time spent in MVPA. We used multilevel analyses accounting for the hierarchical structure of the data in which individuals and schools were included as random effects [33]. Models were transformed by log transformation to fulfil the model assumptions. To adjust for individual differences, wear time, nationality and week day were included as potential confounders, as these parameters may have an influence on children’s activity patterns [34, 35]. A backward elimination algorithm with Akaike’s information criterion (AIC) as goodness-of-fit measure was applied to test the contribution of the entered predictors. To explore interactions between setting and gender, we finally used a single-step-method to calculate contrasts between boys and girls in every setting. Gender differences were provided by the transformed and adjusted p-values.

During the measurement week, all children spent time at home and at own school, on streets and in other places. Four participants (3.4%) did not visit recreational facilities, eight participants (6.7%) did not record any time at other schools and 75 children (63%) did not leave the City of Winterthur.

Children spent 38.8% of their time at home (22.7 hours; IQR: 9.6–32.5), 27.2% at their own school (15.9 hours; IQR: 9.3–21.8) and 15.3% on streets (8.1 hours; IQR: 6.0–11.3). Only 2.3% were recorded in recreational facilities (0.4 hours; IQR: 0.1–1.8) and at other schools (0.4 hours; IQR: 0.1–2.4). The remaining 11.6% (6.1 hours; IQR: 4.2–8.3) were spent in other places such as other’s homes, shopping malls or woodland (Fig 2). Gender did not have a significant influence, neither on total time recorded nor on time spent in any setting (all p > 0.05).

Table 2 provides the results on the amount of MVPA accumulated overall and in the seven activity settings for the total population and separated by gender. Over all settings, children accumulated 296.7 minutes of MVPA during the measurement week. Boys were significantly more active than girls were during the week, accumulating 63.9 more minutes in MVPA than girls (p < 0.001). The most MVPA was accumulated on streets (34.5%) and at own school (26.8%). Street was the only setting in which girls accumulated slightly, albeit not significantly, more MVPA than boys (p > 0.999). In all remaining settings, boys achieved a higher amount of MVPA than their female peers did. However, we only found a significant gender difference at other schools with boys reporting 11.2 more minutes in MVPA (p = 0.035) and a non-significant trend at own school with girls spending 8.4 minutes less in MVPA (p = 0.053).

Overall, the children recorded 8.4% of the time in MVPA, with boys spending only a slightly higher proportion of time in MVPA compared to girls (9.7% vs.7.2%; p = 0.189). The proportion of time spent in MVPA was largest in recreational facilities (19.4%), at other schools (19.2%) and on streets (18.6%). A considerably smaller proportion was classified as MVPA when visiting own school (8.6%), other places (7.1%) and places outside of Winterthur (5.2%). The proportion of time in MVPA at home (3.0%) was even smaller (Table 3). For six out of the seven activity settings, the proportion of time in MVPA was higher among boys than among girls, whereby significant differences were found at own school (3.4%, p = 0.049) and out of study area (6.9%, p = 0.009). Moreover, a non-significant trend was found at other schools with boys being more active than girls were (8.8%, p = 0.072). Street was the only setting where girls showed a slightly higher proportion of time in MVPA than boys (19.1% vs. 17.9%), however, this difference was not statistically significant (p = 0.477).

Although the combination of accelerometry and GPS seems to offer sensitive and accurate measures, there are several limitations. Firstly, known problems with waist-worn accelerometers are associated with the underestimation of activities in which the body centre of gravity is relatively fixed such as cycling or movements of the upper body [47]. Twenty-nine children engaged in sports could not wear the devices during their training or competition. Although only six children reported such an event more than once a week, this loss of recorded PA possibly has resulted in incomplete measures and underestimations, which predominantly affected recreational facilities and the school setting. The choice of processing methods and threshold values to determine MVPA may also have an impact on the recorded PA level [48]. Reactivity to measurement devices could lead to an unnaturally high level of PA. Although our participants did not show an especially high activity level on the first measurement day as reported in previous studies [49], this could have affected their behaviour during the whole measurement week. Furthermore, certain environments such as urban canyons, vegetative cover or indoors within structures impermeable to satellite signals can lead to inaccurate or missing GPS positions [15]. By identifying invalid GPS points and choosing a buffer zone of ten meters around polygons, we tried to reduce issues with spatial inaccuracy, but under- or overrepresentation of certain activity settings due to missing or misclassified GPS points cannot be completely excluded. In particular, the use of buffer zones accounting for the measurement error of the GPS devices might have generated new misclassifications, especially affecting the street setting. Consequently, the development of clear standards, which ensure more consistency in data collection and processing procedures between studies based on accelerometry and GPS is necessary. The current study was conducted during late winter and early spring with rather cold and rainy weather conditions. Bringolf-Isler and colleagues [44] showed that the time spent in different activities such as vigorously playing outdoors or biking significantly decreases with the sum of precipitation, whereas sedentary time spent indoors increases. Bad weather conditions may also be a reason not to participate in outdoor activities during recess [41]. Several studies further demonstrated that children use different activity settings to achieve PA during different seasons [19, 38]. Accordingly, our findings are not representative for the entire year. The study sample derived from one city in the German-speaking part of Switzerland. The included schools were public schools situated in different districts and distributed throughout the whole city. Although the participating children derived from different backgrounds and neighbourhoods, and are rather representative for the population of Winterthur, our results are not generalizable for suburban or rural schools from other Swiss regions.