An analysis of data from the Joint Canada/United States Survey of Health (JCUSH), allows us to compare prevalence estimates that result from four different question sets designed to assess disability from a group of respondents residing in either Canada or the United States. Depending upon the question set used and the coding applied to the responses, age-standardized prevalence estimates varied widely in both countries. In the U.S. noninstitutionalized adult population, disability prevalence estimates ranged from as low as 15.3% to as high as 36.4%, while in Canada the estimates ranged from 13.4% to 37.3%. Concordance and discordance in identification as disabled among these question sets were also examined. In both countries, less than 20% of those identified as disabled by any question set were identified as disabled on all four question sets when using conservative response coding to define disability. Concordance in answers to these questions was also found to be associated with older age, single marital status, low education and low income in both countries. Discordance between question set pairs was similar across both countries whether among measures based on the same domains of disability or different domains of disability. The theory, methods and future of disability measurement in health surveys are discussed in light of these findings. We conclude that understanding and interpreting national prevalence estimates requires more thoughtful attention to the purposes for which data are being collected, the specific definition and operationalizations of disability for those purposes, the methodology used in the data collection and analysis process and the areas of both commonality and difference in the populations identified by each question set. In terms of cross-cultural comparisons, the use of a common set of questions and answer categories and similar survey methodologies provides much more robust results.