Panel surveys interview the same individuals more than once over a period of time. Attrition from the survey occurs when those individuals die, refuse to be interviewed again, or, for some other reason, cannot be contacted. If the original sample was representative of a specific population, then survey analysis may provide misleading conclusions about changes in population characteristics over time if these individuals leave the sample in a nonrandom way. Therefore, it is important to identify the characteristics of individuals who leave the survey for various reasons. This article explores the extent of and reasons for attrition in the New Beneficiary Survey (NBS) between the first interview in 1982 and the followup interview in 1991. Presented is a comparison of the characteristics of survivors (the reinterviewed sample) with attriters (those in the sample not reinterviewed) from the retired-worker and disabled-worker samples. The article explores a variety of potential determinants of attrition to the probability of attrition. These determinants are examined alone and in a multivariate framework. The NBS sample population is drawn from and linked to Social Security Administrative records, which have exact matched data on mortality as a cause of attrition. These data do not depend on survey-reported reasons for attrition; hence, it allows the examination of the differences in the patterns and predictors of attrition due to death and due to other reasons, primarily, the refusal to be interviewed. Attrition due to death must be identified precisely because misidentification of death as refusal to be interviewed may lead researchers to infer more selective attrition than might be the case. Different patterns of attrition are evident in the comparison of attrition levels and the determinants of attrition for the retired and disabled samples, both composed of persons with relatively high mortality risk. In particular, individuals' health, health insurance coverage, and level of education have different impacts on their likelihood of attrition. In general, it appears that refusal to be interviewed is more evenly spread across populations and characteristics than is death. The analysis shows that attrition due to death and attrition due to refusal are quite different processes, even though health conditions play a role in both processes. The results suggest that because attrition patterns (including death) may be quite different across population samples, sample-specific attrition patterns must be analyzed over the lifetime of any panel study. Long-term studies of panel attrition are necessary to provide researchers analyzing the data with information on potential biases due to nonrandom attrition.