Direct standardization of incidence rates in the presence of incomplete data

When comparing the disease incidence rates for several subpopulations, epidemiologists often use direct standardization to adjust for potential confounding variables. In population-based studies, however, the data are often incompletely classified with respect to membership in the subpopulations of interest. In such a situation, one often assumes that the cases with missing data have the same distribution as the complete cases, that is the data are missing completely at random. In this setting, we derive variance estimates for the directly standardized rates which account for the use of incomplete data. We illustrate the use of these methods with data from a study of the incidence of gastrointestinal cancer by immigrant status where birthplace data are often incomplete.