Bootstrap Inference for Local Populations

The randomized available case study, in which a nonrandom set of cases (patients, animals, laboratory runs) is randomized among two or more treatments, is a staple of biomedical research. Traditionally, such studies have been analyzed as though the cases were a random sample from an infinitely large population (1). The resulting statistical inferences address incorrect populations. More importantly, in the presence of response measurement error these inferences are inappropriate for the correct populations, understating the differential impact of treatment (2). In this paper I develop and illustrate a nonparametric bootstrap approach to inference in such studies, an approach that is faithful to the local origins of the randomized cases and can account for the influence of measurement error.