Abstract:
|
Sleep duration is an important factor associated with various health outcomes and continues to be of interest to researchers for its potential investigative benefits. It is often measured through self-report, though self-reported sleep duration is systematically biased. Accelerometer devices have been recommended as an alternative method to provide an objective measure of sleep duration, though commonly used sleep detection methods may be sensitive to types of devices, sampling frequencies, location of placement, and population demographics. Thus, sleep detection algorithms developed in a specific study are often not successfully generalizable. This study develops novel statistical methods for estimating sleep duration that are more robust and consistent, based on parametric and non-parametric models. Our proposed parametric methods are based on two changepoint detection methods: 1) an optimal partitioning method and 2) a fused lasso approach, as well as on cosinor analysis. Furthermore, we conduct non-parametric rhythm analysis, which provides measurements on sleep variability, duration, and timing, helping to further knowledge on impacts of sleep patterns on health outcomes.
|