Conventional analysis of in vitro assays of motor proteins rests on the assumption that all proteins with the same chemical composition function identically; however molecule-to-molecule variation is often seen even in well-controlled experiments. In an effort to obtain a statistically meaningful set of time traces that simultaneously avoid any experimental artifacts, we performed quantum-dot labeled kinesin experiments on both surface and levitated microtubules. Similar to glassy systems, we found that mean velocities of individual kinesin motors vary widely from one motor to another, the variation of which is greater than that expected from the stochastic variation of stepping times. In the presence of heterogeneity, an ensemble-averaged quantity such as diffusion constant or randomness parameter is ill-defined. We propose to analyze heterogeneous data from single molecule measurements by decomposing them into homogeneous subensembles.