From Distribution Shift to Kernel Methods: A Study of Empirical Phenomena in Machine Learning