In order to estimate the expected number of accidents for an "entity" (intersection, road section, driver), the entity must be assigned membership in some population of similar entities. Concepts like "bias-by-selection" have no meaning unless they refer to selection from some population. In this article, three problems in defining such populations, with particular reference to populations of road sections, are discussed. The problems are: (1) how to define an adequate "reference" population for various road safety measures; (2) problems created by the "moving" definition of road sections with an abnormal accident count; and (3) problems in treating road sections of different length as members of the same population of sections. It is concluded that only entities which are "similar" in some respect may constitute a population. There is no satisfactory way of defining a population of moving road sections. Likewise, the problem of whether, e.g., a four-kilometer section should be seen as a sample of one unit from a population of four-kilometer sections, or as a sample of, e.g., four units from a population of one-kilometer sections is at present unsolved. A theory must be built in order to solve these difficulties of defining populations, in particular populations of road sections.