Total body water (TBW) was measured by deuterium oxide (D2O) dilution and predicted from bioelectrical impedance (Z) in nineteen anorexic and twenty-seven control women. The equation of Kushner et al. (1992) based on the impedance index (ZI = height2/Z) gave biases of 0.9 (sd 2.5) and 0.8 (sd 2.5) litres in controls and patients respectively (NS, ANOVA). The ZI-based equation of Deurenberg et al. (1993) gave biases of 1.5 (sd 2.4) litres (NS) and 3.0 (sd 2.1) litres (P <0.001) in controls and patients respectively. Despite the fact that weight was the most powerful predictor of TBW on the study sample (n 46, r2 0.90, P < 0.0001, se of the estimate 1.6 litres, CV 5.7%), the formulas of Segal et al. (1991) and Kushner et al. (1992) based on the association of weight and ZI gave an inaccurate prediction of TBW in both control and anorexic subjects, with a bias ranging from -3.2 (sd 2.4) to 2.9 (sd 2.1) litres (P ≤0.001). Population-specific formulas based on ZI (n 46) gave a more accurate prediction of TBW by bioelectrical impedance analysis on the study subjects, with biases of -0.1 (sd 1.8) and 0.5 (sd 1.7) litres in controls and patients respectively (NS). However, the individual bias was sometimes high. It is concluded that bioelectrical impedance analysis can be used to predict TBW in anorexic women at a population level, but the predictions are less good than those based on body weight alone.