Fatty liver is a common disease in Taiwan. In this study, we tried to evaluate the validity of predicting the presence of fatty liver from clinical data instead of liver biopsy or sonography. From a community survey in Putai, a total of 873 adults older than 30 years and quantified as to triglyceride level, body height, body weight, and the results of the oral glucose tolerance test and upper abdominal sonography were recruited for analysis. Using a receiver operating characteristic (ROC) curve, the best 'cutoff values' for determination of fatty liver were predicted from body weight index and serum triglyceride level in 8 clusters grouped by sex, age and presence or absence of diabetes mellitus. The best cutoff values of triglyceride in the 8 clusters varied from 100 to 170 mg/dl with worse validity. Most of the values were 130 and 140 mg/dl. The cutoff values of body weight index were constant in all clusters and showed greater validity than those for triglyceride. They were 115% or 120%. Their accuracy for the prediction of fatty liver was positively correlated with its prevalence. However, their accuracy was lower than 70% in non-diabetic females. We conclude that body weight index is a good parameter for prediction of fatty liver especially in the high risk groups and we recommend that health-determining cutoff values of serum triglyceride and body weight index should be set at 130 mg/dl and 115%, respectively.