Study design and setting
This cross-sectional study was situated in four medium-sized cities in the Southern part of The Netherlands. The number of inhabitants ranged from 77,450 to 201,259 and the population density ranged from 727 to 1,716 citizens per km2. Although one city was somewhat smaller and less urbanized compared to the other cities, they were comparable regarding the demography of their population such as the percentage of non-Western immigrants (range: 9.9 - 13.4%) and the percentage of inhabitants aged 0–14 years (range: 16.7-17.6%). The selection procedures and characteristics of the participating cities are described in more detail elsewhere [19].
Data on physical activity behavior of the children were obtained by means of a cross-sectional survey consisting of a written questionnaire for parents between September 2007 and January 2008. The data on neighborhood characteristics were collected approximately one year later (between October and December 2008) by means of standardized neighborhood observations (audits) by trained observers. Based on postal code (six positions: four numbers, two letters available from both the questionnaire among parents and from municipal data describing which postal codes fall within one neighborhood), the data from these two study parts were combined for the analyses of this paper. Both study parts will be described in more detail below.
Survey among parents
The study was targeted at primary school children aged 4–12 years. In the Netherlands, children in this age group attend primary school, which, in most cases, is close to or within the area of residence. Initially, all regular primary schools in the four cities (n = 149), except those already participating in other (research) projects aimed at physical activity among children (n = 34) were invited by letter, followed up by a telephone call to participate in the survey. Of the invited schools (n = 115), approximately one third agreed to participate (n = 42). As outlined elsewhere [19], the schools in our study were representative for the total population of schools in the participating cities in terms of school size, socioeconomic status and type of neighborhood.
At each school enrolled in the study, all grades and classes were included in the survey. Because no medical or physical measurements were conducted and considering the negligible (psychological) burden to fill in the questionnaire, no ethics approval was required according to the Dutch Central Committee on Research Investigating Human Subjects. Parents were given written information about the study and by returning the questionnaire they gave consent for the inclusion of their data in the study. In total parents of 11,094 children were provided with a questionnaire. Parents that had more than one child attending the same school, were provided with a questionnaire for each individual child. Response rate was 60%, resulting in 6,624 returned questionnaires. During data entry, 12 questionnaires could not be read and 11 questionnaires were removed because they were completely empty, leaving 6,601 completed and returned questionnaires.
Parents were asked to report the frequency (number of school days and number of days per weekend) their child was involved in outdoor play, considering a typical week in the past month. Parents were also asked to report the duration of outdoor play during week and weekend days (less than 30 minutes per day, 30 minutes to one hour per day, one to two hours per day, more than two hours per day). Furthermore, the questionnaire included items on age and gender of the child and parental education level and net household income per month. Based on parental report of weight and height of their child, BMI was calculated and percentage overweight and obesity (as determined by age and gender specific cut off points provided by Cole et al. [20]) was determined. Because parents were also asked to report their postal code in the questionnaire, the survey data could be coupled to the neighborhood observation data described in the next paragraph.
Neighborhood observations
Neighborhoods were selected for observation based on 1) the number of respondents included in the survey living in the neighborhood in order to maximize the number of respondents in the analyses and, 2) physical neighborhood characteristics (based on a neighborhood typology score from the Ministry of Housing, Spatial Planning and the Environment which classifies neighborhoods into the following six categories: city centre, city non-centre, city green, town centre, rural area, and work area [21]) and social neighborhood characteristics (based on the status score from the Netherlands Institute for Social Research which is based on the percentage of immigrants, percentage of people with low education and percentage of low income households per postal code area [22]) in order to maximize the variance in neighborhood characteristics included in the analyses. In total, 57.6% of the parents that filled in a questionnaire during the survey, were living in one of the 33 observed neighborhoods. Hence, combining the data from the survey among parents and the data from the neighborhood observations, resulted in 3,805 individual respondents for the analyses described in this paper.
Data on neighborhood characteristics (the independent variables) were collected by two trained research assistants by means of neighborhood observations in 33 neighborhoods. The observers were not part of the research team to enhance unbiased collection of the data. The two research assistants observed the neighborhoods using a checklist which they completed by mutual agreement. The checklist was based on the Neighborhood Environment Walkability Scale (NEWS) [23], but was specifically adapted for screening Dutch neighborhoods on environmental characteristics related to children’s physical activity [16]. The inter-rater reliability of the checklist was evaluated as good (percentage of agreement = 77%) in a previous Dutch neighborhood observation study [17]. The scoring form included the following seven main topics: 1) buildings (residential density, land use mix, presence of unoccupied houses and maintenance of buildings), 2) formal outdoor play facilities (number and quality of play grounds, school yards, paved play grounds, and half pipe or skating track), 3) public space (presence and quality of green space and water), 4) street pattern (presence and quality of sidewalks and bike lanes) 5) traffic safety (traffic infrastructure and traffic volume and speed), 6) neighborhood characteristics related to the social environment (e.g. street hygiene such as a litter basket for dog waste, graffiti and vandalism (indicating area deprivation) and the presence of a dog walkers area or adequate street lighting which may contribute to social safety) and 7) general impression of the activity-friendliness of the neighborhood for children.
Neighborhood boundaries were defined by local data bases from the municipal organization, so that the results of the study could be easily interpreted by local policy makers. In general, these boundaries correspond with what people perceive as “their neighborhood” and boundaries often coincide with physical “boundaries” such as a railway, busy road, channel or tunnel. In the Netherlands, parents are free to choose a primary school for their child, according to their own opinion and beliefs. Due to practical considerations, many parents choose a primary school close to their home. Hence, in the majority of cases, both the residence and the school of the children will fall within the same neighborhood. Indeed, from our data sample it appeared that 76.5% of the children included in the neighborhood observations, attended school in the same neighborhood as they live in. In cases where the school did not fall within the neighborhood observation area, children were likely to attend school in the adjacent neighborhood.
Similar to another Dutch neighborhood observation protocol developed by Van Lenthe et al. [24], before the start of the actual data collection, a random sample of ten percent of the streets within each neighborhood was selected for observation by foot, based on a list of all streets per neighborhood. Thereafter, all remaining streets in the neighborhood were observed per bicycle, so that all streets were included in the observation. All observations were carried out during normal school days after school time and before dark, to mimic best the real conditions under which children are usually involved in outdoor play in their neighborhood.
Measures
The dependent variable in all analyses was outdoor play in minutes per week which was calculated by multiplying the number of school days and weekend days the parents reported their child was involved in outdoor play by the average minutes per day the child was involved in outdoor play during school days and weekend days (which was recoded as follows: less than 30 minutes per day = 15 minutes per day, 30 minutes to one hour per day = 45 minutes per day, one to two hours per day = 90 minutes per day, more than two hours per day = 150 minutes per day). Finally, to calculate the total minutes of outdoor play per week, minutes spent on outdoor play during school days and during weekend days were summed.
As stated in the previous paragraph, the neighborhood observation checklist included seven main topics, which yielded in total 33 independent variables which will be described briefly here. A detailed description of all variables included in the analyses is given in Additional file 1: Appendix A. Residential density was estimated by weighting and summing nine items on type of residences in the neighborhood, with a higher sum score representing higher residential density. Land use mix was defined as the proportion of enterprises to residences (range 0-100%). Presence of unoccupied houses was measured on a five-point scale (none-all) and maintenance of buildings was measured on a three-point scale (bad, moderate, good). The total number of play grounds, school yards, paved play grounds, and half pipe or skating track per km2 was calculated and summed for each neighborhood, resulting in one score for number of formal outdoor play facilities per km2 per neighborhood. Quality of play grounds, school yards, paved playgrounds, and half pipe or skating track was defined on a scale from 0.00 to 1.00, according to ten quality aspects mentioned in Additional file 1: Appendix A. For each type of outdoor play facility separately, 0.10 point per quality aspect was awarded whenever applicable. Hence the quality score per outdoor play facility could range from 0.00 to 1.00, higher scores represent better quality. A mean quality score for all outdoor play facilities per neighborhood was then calculated. Presence of green space, water, sidewalks and bike lanes were each measured on a four-point scale (none-many). Quality of green space, water, sidewalks and bike lanes were defined on a scale from 0.00 to 1.00 (see Additional file 1: Appendix A for specification of quality aspects). Traffic infrastructure included the following single-item variables each measured on a four-point scale (none-many): pedestrian crossings without traffic lights, pedestrian crossings with traffic lights, traffic lights, refuges / safety islands, parallel parking places, parking lots (grouped), speed bumps, home zones, 30 km/ hour zones, roundabouts, and intersections. Traffic volume and speed was calculated as a sum score of 6 items each measured on a four-point scale (none-many) (Cronbach’s alpha = 0.898), with a higher score representing higher traffic volume and speed. Presence of a dog walking area was a dichotomous item, as was the presence of a litter basket for dog waste and the presence of street lighting. The presence of graffiti, vandalism and dark spaces were each measured on a four-point scale (none – many). General impression of the activity-friendliness of the neighborhood for children was estimated by a score ranging from 1–10, with a higher score representing a more favorable impression. Two items were removed from the analyses due to lack of variation among neighborhoods: the presence of parking garages and the presence of low-traffic / car-free zones. Except for the number of play facilities (which were counted in each neighborhood), the research assistants that performed the neighborhood observations had to give one overall neighborhood score for each of the measures.
Statistical analyses
From the 3,805 individual respondents in this study, 52 questionnaires were excluded from further analyses because of missing values on the outcome measure outdoor play, and 91 additional questionnaires were removed because of missing values on potential confounders: age or gender of the child (n = 6) and parental education (n = 85). Furthermore, questionnaires of children living more than three days per week on another address than the address described in the questionnaire (n = 18) were removed. Since some questionnaires had to be removed because of more than one exclusion criterion, the final data base for the analyses on outdoor play encompassed 3,651 respondents.
Because different environmental characteristics are expected to be associated with outdoor play between boys and girls and children of different age groups [4], analyses were conducted separately for boys and girls and in age groups 4–6, 7–9, and 10–12 years. Descriptive analyses were conducted with SPSS 17.0 (Chicago, Illinois). ANOVA and chi-square tests were performed to assess differences (p < 0.05) in characteristics between boys and girls in each age group for continuous and categorical variables respectively. Likewise, t-tests and chi-square tests were performed to asses differences (p < 0.05) between respondents that were included in a neighborhood observation (n = 3,805) and the original sample of parents derived from the questionnaire (n = 6,601).
To quantify the association between neighborhood characteristics and children’s outdoor play, multilevel GEE analyses were conducted with SAS 9.1 (Cary, North Carolina). Because most of the independent variables were collected at the neighborhood level, but the dependent variable was collected at the individual level, multi-level analyses with neighborhood as a clustering variable were applied in order to correct for the multi-level structure of the data. Because data were collected completely anonymously, there was no information available on the number of children per household, and hence family-related clustering effects cannot be accounted for in the multi-level analyses. To estimate the possible design effects associated with family clusters, design effect calculations were performed, using the following equation: design effect = 1 + ICC*(n – 1), whereby n is the median number of children within one family. Although municipal data showed that the median number of children within one family was one (which always yields a design effect of 1.00), the calculations were also performed for a median number of two children per family. A pedometer based study by Jacobi et al. [25] has shown that the ICC for physical activity of siblings within one family can be as high as 0.3, hence an ICC of 0.3 is used for the design effect calculations.
Because of the non-normal distribution of the dependent variable outdoor play and its error terms (as assessed by histograms and normal probability plots, data not shown) and since this outcome measure is a count variable (number of minutes outdoor play per week), a Poisson distribution was applied [26, 27]. As a consequence, exponents of the original regression coefficient estimates were calculated and interpreted as relative rates (RR). The RR can be interpreted as estimated proportional difference in the amount of outdoor play. For example, an RR of 1.10 indicates 10% longer outdoor play for each additional unit in predictor variables. Due to the Poisson analysis, the proportion of explained variance cannot be reported.
The first step in the analyses focused on environmental characteristics within each of the seven main topics included in the neighborhood observations: buildings, formal outdoor play facilities, public spaces, street pattern, traffic safety, neighborhood characteristics related to the social environment and general impression of the activity-friendliness of the neighborhood for children. All independent variables of one topic were entered simultaneously into a separate model (so one model per topic), which was adjusted for age of the child and parental education level, as indicated by highest completed education of the parent who filled in the questionnaire (it was assumed that this person was the primary caregiver, in the majority of cases this was either the biological mother or the biological father, 81.8% and 11.6% respectively). Parental education level is considered a good indicator for socio-economic status in The Netherlands [28] and is preferred when statistically controlling for socio-economic status in a regression model [29]. Quantitative (i.e. presence or amount) and qualitative aspects of neighborhood characteristics were entered simultaneously in each step of the analyses.
In order to quantify the association between the environmental characteristics and outdoor play when adjusted for the environmental characteristics from other topics, multivariate regression analyses were also performed. In these analyses, all significant (p-value < 0.05) variables from the analyses per topic were entered into a multivariate model, which was also adjusted for age of the child and parental education level. Non-significant variables (p-value > 0.05) were removed one-by-one from the multivariate models (backward elimination procedure), until all variables were statistically significant (p-value < 0.05, except for the potential confounders age of the child and parental education level which were forced into the multivariate model irrespective of significance). In order to check whether choosing a more liberal p-value of 0.10 as a decision criterion for this backward elimination procedure would have influenced the final multivariate models, additionally, all analyses were re-run with a more liberal p-value of 0.10 (same procedure as stated above, only with a p-value of 0.10 instead of 0.05. These additional analyses were performed to check whether potentially important variables were excluded too easily from the multivariate models when applying a p-value of 0.05.