We investigated the degree of generalizability of relative risk (RR) estimates for a sample from a well-defined population to a general population using actual data. A total of 45,452 men aged 40-69 years who completed a self-administered questionnaire were followed from 1990-1994 to the end of 1999-2000 in the Japan Public Health Center-based Prospective Study. We considered those who responded to the self-administered questionnaire (45,452 men) as representing the general population (population), and those who underwent a health check-up as representing a sample from a well- defined population (sample) (12,162 men). Exposure distributions, mortality rates, and the confounder-adjusted RRs of all-cause mortality according to cigarette smoking or body mass index (BMI) were compared between the sample and the population using empirical sampling distributions from the population. The sample had significantly different prevalences of exposures and lower mortality rates than the population. Adjusted RRs were significantly higher in current smokers (RR = 1.83) and in subjects who smoked > or =40 cigarettes per day (RR = 2.67) in the sample than the respective values in the corresponding categories (RR = 1.48 and 1.62, respectively) of the population. The adjusted RR of those with the lowest BMI in the sample (RR = 1.30) were also significantly lower than those of the corresponding category in the population (RR = 2.06). Our results suggest that even for RRs, the extrapolation of estimates for a sample from a well-defined population to a general population may not be possible.