This paper deals with abductive reasoning on knowledge bases that are expressed at different levels of abstraction, but are not necessarily organized as a set of increasingly more abstract models, each one giving a complete (even if abstracted) description of a domain. We claim that the search for abductive explanations in such a context and, in particular, the choice of the “right” level at which explanations have to be determined, should be driven by the available observations in such a way that explanations involving low-level phenomena are allowed only if there are specific observations related to them, or higher-level explanations cannot be found. We present formal definitions following this principle and we discuss how explanations can be computed according to the definition.