Although physical activity has been considered as an important modifiable risk factor for obesity, the empirical evidence on the relationship between physical activity and obesity is mixed. Observational studies in the public health literature fail to account for time-invariant unobserved heterogeneity and dynamics of weight, leading to biased estimation of the effect of physical activity on obesity. To overcome this limitation, we propose dynamic fixed-effects models to account for unobserved heterogeneity bias and the dynamics of obesity. We use nationally representative longitudinal data on the cohort of adults aged 18–50 years in 1994/95 from Canada's National Population Health Survey and followed them over 16 years. Obesity is measured by BMI (body mass index). After controlling for a wide range of socio-economic factors, the impact of four alternative measures of leisure-time physical activity (LTPA) and work-related physical activity (WRPA) are analyzed. The results show that each measure of LTPA exerts a negative effect on BMI and the effects are larger for females. Our key results show that participation in LTPA exceeding 1.5 kcal/kg per day (i.e., at least 30 min of walking) reduces BMI by about 0.11–0.14 points in males and 0.20 points in females relative to physically inactive counterparts. Compared to those who are inactive at workplace, being able to stand or walk at work is associated with a reduction in BMI in the range of 0.16–0.19 points in males and 0.24–0.28 points in females. Lifting loads at workplace is associated with a reduction in BMI by 0.2–0.3 points in males and 0.3–0.4 points in females relative to those who are reported sedentary. Policies aimed at promotion of LTPA combined with WRPA like walking or climbing stairs daily would help reduce adult obesity risks.