A1052
Title: Assessing balance of functional covariates via dynamic time warping distances
Authors: Mireya Diaz - Case Western Reserve University (United States) [presenting]
Abstract: The dynamic time warping (DTW) distance is widely used in computational geometry. Despite its widespread use, little is known about its distributional properties and, thus, its utility for statistical inference, particularly in assessing the balance of baseline functional covariates, an open question. A simulation study examined such properties empirically under null and alternative scenarios for four families of functions: linear, quadratic, sinusoidal, and exponential. Alternative scenarios evaluated time transformations. All scenarios were tested under three levels of noise and two series lengths. Three implementations of DTW were assessed: The original DTW, the constrained Sakoe-Chiba, and a linear weighted version. For the three implementations, most alternative scenarios generate a DTW distribution that differs substantially from the null, marking the 0.1 ASMD threshold. That is not the case for the exponential function when the time transformation generates differences still close to the 0.1 threshold; in these cases, the distributions overlap. This is more patent in series with fewer observations. The scenarios revealed too that the three DTWs are susceptible to noise under the null but not under the alternative. Beta-type distributions fit well many of the DTWs distances, as does the Frechet distance, another ordered statistic of the Euclidean distance.
