We conducted a prospective study to clarify the relationship between mortality and body mass index (BMI) in a middle-aged and elderly population in Japan. The subjects included 12,649 people (5,686 males and 6,963 females) aged from 40-69 years, who were identified by a mail survey between 1987 and 1989, and thereafter were followed up from the date of the survey until February 28, 1995. Cox's proportional hazards model was used to adjust for age, smoking status, drinking habit and occupation, and hazard ratios were calculated for total mortality and cancer mortality based on levels of BMI, while using a BMI of 22 to approximately 24 as a reference level. Results showed a U shaped relationship between BMI and total mortality in females. Females demonstrated a minimal total mortality in the reference group, but a significantly high hazard ratio in the lowest group of BMI < 20 (hazard ratio = 1.95, p < 0.01) and the highest group of BMI > or= 26 (hazard ratio = 1.71, p < 0.01). On the other hand, males did not demonstrate such a U shaped relationship, but instead showed an L shaped relationship between BMI and total mortality. Males showed a minimal total mortality at a BMI of 24 of approximately 26 (hazard ratio = 0.92), but a significantly high hazard ratio in the lowest BMI group (hazard ratio = 1.57, p < 0.01). In contrast, an elevated hazard ratio was not observed in the highest BMI group (hazard ratio = 1.05). The relationship between BMI and cancer mortality was found to be similar to the relationship between BMI and total mortality in both males and females. Our findings suggest that moderate overweight may be a sign of good health among both middle-aged and elderly men in Japan.