'What if' predictions, when compared with responses of treated units, form the basis for all program impact evaluations. This paper explains the statistical concept of strong ignorability and proposes it as a framework for addressing the internal validity of all impact assessments. After identifying assignment mechanisms that are strongly ignorable, the paper discusses what can be done when assignment mechanism are not ignorable, with attention to using instrumental variable techniques. Also examined are research designs that make assignment mechanisms strongly ignorable. The discussion concludes that strongly ignorable mechanisms are desirable, if not mandatory. Over 20 references. (Author abstract modified)