Estimation of weight loss plays a key role in the nutritional assessment of patients. The loss is usually calculated by subtracting the patient's observed weight (O) either from his recalled weight (R) or from his predicted weight (P) taken from standard tables or equations. We have compared these two assessments of weight loss (R-O, P-O) in a cross-sectional study of patients in the surgical wards of a teaching hospital. There are large differences (up to 15 kg) between average predicted weights taken from the various published tables. We have obtained predicted weights using equations derived from a healthy local population. We have devised a general technique with which the measured, R and P weights in a group can be used to give random errors of R, P, and true weight loss. In our patients there were 3.6, 10.7, and 6.0 kg respectively. As the random error of R was smaller than that of P in our patients, R-O gave better estimates of the mean and SD of the weight losses of the group of patients than did P-O, and R-O was a more precise estimate of the true weight loss of an individual. Nevertheless, R-O is only an estimate of the true weight loss and had a large random error (up to 7.2 kg). This error which can be calculated for any group of patients, must be borne in mind when using R-O to estimate the weight loss in an individual, and when selecting patients with a given weight loss. In 100 patients such as ours, 24 would have R-O greater than 10 kg, but only 18 of these would have lost 10 kg, and nine other patients who had lost 10 kg would be missed.