In the Dutch Longitudinal Study among the Elderly, a probability sample of 3149 persons from the population of the Netherlands, initial ages 65-99 years, was followed from 1955 for 28 years. Vital status was determined in 1983 for 84% of the original sample. Multivariate regression models were used to predict the realized probability of dying (RPD), a measure of longevity based on actual survival time, sex and age at baseline. Predictor variables included physical, mental and social indicators of health status. The total variance explained was 20.2%. Objective health (rated by the examining physician) showed the strongest relationship with the RPD. Upon removing objective health from the model, 19.5% of the variance remained explained. Separate analyses were performed for (1) bio-medical, physical and mental examination variables, (2) disability and health care use variables, and (3) social and psychological variables; in each case the variance explained was over 11%, demonstrating considerable interdependence among predictors. Across different regression models, bio-medical and disability variables proved to be the most stable predictors of longevity.