Background To examine how factors from a social ecologic model predict physical activity (PA) among adolescents using a longitudinal analysis. Methods Participants in this longitudinal study were adolescents (ages 10-16 at baseline) and one parent enrolled in the Transdisciplinary Research on Energetics and Cancer-Identifying Determinants of Eating and Activity (TREC-IDEA) and the Etiology of Childhood Obesity (ECHO). Both studies were designed to assess a socio-ecologic model of adolescent obesity risk. PA was collected using ActiGraph activity monitors at two time points 24 months apart. Other measures included objective height and weight, adolescent and parent questionnaires on multilevel psychological, behavioral and social determinants of PA, and a home PA equipment inventory. Analysis was conducted using SAS, including descriptive characteristics, bivariate and stepped multivariate mixed models, using baseline adjustment. Models were stratified by gender. Results There were 578 adolescents with complete data. Results suggest few statistically significant longitudinal associations with physical activity measured as minutes of MVPA or total counts from accelerometers. For boys, greater self-efficacy (B = 0.75, p = 0.01) and baseline MVPA (B = 0.55, p < 0.01) remained significantly associated with MVPA at follow-up. A similar pattern was observed for total counts. For girls, baseline MVPA (B = 0.58, p = 0.01) and barriers (B = -0.32, p = 0.05) significantly predicted MVPA at follow-up in the full model. The full multilevel model explained 30% of the variance in PA among boys and 24% among girls. Conclusions PA change in adolescents is a complex issue that is not easily understood. Our findings suggest early PA habits are the most important predictor of PA levels in adolescence. Intervention may be necessary prior to middle school to maintain PA through adolescence.