The participants were sampled from four high schools in suburban Philadelphia. A total of 1517 children were identified through class rosters at the beginning of 9th grade. Those with a special classroom placement, and who spoke English as a second language were ineligible to participate. Based on this selection criteria 98% of the adolescents were eligible (n=1,487). Parental consent was obtained on 99% of the eligible adolescents (n=1,478), and 97% of the eligible participants completed the baseline survey (n=1,429). A total of 30 adolescents were absent on the assent/survey day, and a further 19 were not interested in the study. The participants who completed the baseline survey were followed-up every 6 months over a 4 year period. The University of Pennsylvania Institutional Review Board granted ethical approval for the study.
The participants self-reported their height and weight, from which BMI (kg/m2
) was calculated. It has been shown that self-reported and measured BMI’s are highly correlated (8
). The height and weight data were compared to CDC growth charts, to help identify any excessive high and low self-reported values (9
Time spent watching television/videos and playing video games, on a week night during the school-term, was self-reported by the participants (<1hr/day, 1 hr/day, 2 hrs/day, 3 hrs/day, 4 hrs/day, or >5 hrs/day). The questions used to capture screen time have previously been validated (10
Gender, race, and maternal education (marker for socioeconomic status) were included as covariates in the present study. We also adjusted for time spent in moderate to vigorous physical activity (MVPA) and hours of sleep, to determine if any associations between screen time and changes in BMI were independent of MVPA levels and sleep duration.
Longitudinal quantile regression was used to investigate changes in the BMI distribution over time. This statistical approach is an extension of ordinary least square regression and models the effect of predictors across the distribution of continuous outcome variables; in the present study the 10th
BMI quantiles were specified in the models (11
). BMI was modeled as the dependent variable, with study wave (coded: 0, 1, 2…6, and 7) and gender included as independent variables, to describe changes in the BMI distribution over time (model 1). Screen time was added as an independent variable to determine if screen time was associated with changes in the BMI distribution over time (model 2). Race and maternal education (model 3), hours of sleep (model 4), and MVPA (model 5) were also added as independent variables; the purpose of models 3 to 5 was to determine if any association between screen time and changes in BMI remained after adjustment for the covariates. A first order autoregressive correlation structure was used, and 95% confidence intervals were calculated using 500 bootstrap samples, to take into account the repeated measures on individuals (11
). All analyses were conducted using Stata 12.0 (StataCorp LP, College Station, TX).