BACKGROUND AND OBJECTIVES: Previous cohort studies reported that a single measure of physical activity (PA) assessed at baseline was associated with lower Parkinson's disease (PD) incidence, but a meta-analysis suggested that this association was restricted to men. Due to the long prodromal phase of the disease, reverse causation could not be excluded as a potential explanation. Our objective was to study the association between time-varying PA and PD using lagged analyses to address the potential for reverse causation, and to compare PA trajectories in patients prior to diagnosis and matched controls. METHODS: We used data from E3N (1990-2018), a cohort study of women affiliated with a national health insurance plan for persons working in education. PA was self-reported in six questionnaires over the follow-up. As questions changed across questionnaires, we created a time-varying latent PA (LPA) variable using latent-process mixed models. PD was ascertained using a multistep validation process based on medical records, or a validated algorithm based on drug claims. We set-up a nested case-control study to examine differences in LPA trajectories using multivariable linear mixed models with a retrospective time scale. Cox proportional hazards models with age as the timescale and adjusted for confounders were used to estimate the association between time-varying LPA and PD incidence. Our main analysis used a 10y-lag to account for reverse causation; sensitivity analyses used 5y, 15y, and 20y-lags. RESULTS: Analyses of trajectories (1,196 cases, 23,879 controls) showed that LPA was significantly lower in cases than in controls throughout the follow-up, including 29y before diagnosis; the difference between cases and controls started to increase ∼10y before diagnosis (P-interaction=0.003). In our main survival analysis, of 95,354 women free of PD in 2000, 1,074 women developed PD over a mean follow-up of 17.2y. PD incidence decreased with increasing LPA (P-trend=0.001), with 25% lower incidence in those in the highest quartile compared to the lowest (adjusted hazard ratio=0.75, 95% confidence interval=0.63-0.89). Using longer lags yielded similar conclusions. DISCUSSION: Higher PA level is associated with lower PD incidence in women, not explained by reverse causation. These results are important for planning interventions for PD prevention.