Twenty-six males (26.5 +/- 6.0 yr; X +/- SD) were studied before and after a fourteen week endurance training program to determine the validity of anthropometric equations for estimating changes in body composition (BC). Anthropometric measures included skinfolds (SF), circumferences, and diameters. Body density (BD) was determined by underwater weighing corrected for residual lung volume. Training resulted in an increase in BD (1.061 +/- 0.002 to 1.067 +/- 0.002 g/ml; X +/- SEM) and decreases in body weight (73.0 +/- 2.1 to 71.4 +/- 2.0 kg), relative fat (16.6 +/- 0.9 to 14.1 +/- 0.8%), fat weight (12.4 +/- 1.0 to 10.2 +/- 0.8 kg), and seventeen of the anthropometric measures (p less than 0.05). Cross-validation of twenty-four equations revealed validity coefficients (r2) and total error in relative fat (RFE) of r2 = 0.40-0.77 and RFE = 2.60-10.15% before training and r2 = 0.14-0.61 and RFE = 2.62-9.45% after training. Linear and base 10 logarithmic (log10) equations using primarily SF measures tended to have higher r2 and lower RFE than equations based on quadratic and natural logarithmic (loge) models and other anthropometric measures. Paired t-tests revealed that of these equations with higher r2 and lower RFE, only the linear equation by Forsyth & Sinning (BD = 1.10647--0.00162(scapSF)--0.00144(abdSF)--0.00077(triSF++ +) + 0.00071(midaxSF] was a stable predictor of BD during training. These results suggest that many existing equations may not be accurate predictors of changes in BC during training.