The view is presented that multilevel methods are just one element in a hypothetical complete analysis of observational data on human subjects. In most contexts several sources of uncertainty, in addition to those captured by a multilevel analysis, are present, and so the confidence placed in the results of a typical multilevel analysis is unrealistically optimistic. A “soft-ware-free” analysis of longitudinal data with rectangular design is outlined. Questions posed by the National Center for Education Statistics and elaborated by de Leeuw and Kreft are briefly discussed.