The reproducibility of bone mineral density (BMD) measurements by dual-energy X-ray absorptiometry (DXA), based on 12 successive monthly determinations, was assessed in a group of 24 subjects (23 postmenopausal women, 1 man) using six trained operators. The variability (S2A) was calculated from both duplicate operator measurements and the standard error of estimate from nonparametric regression of the individual subject series. Robust estimates of SA from the 90th percentile of the sampling distribution of variances were calculated for the spine (25 mg/cm2), femur neck (20 mg/cm2) and total femur (15.5 mg/cm2) using the bootstrap technique. The critical difference for a significant decrease (p = 0.05) at the spine, femoral neck and total femur was estimated at 57, 46 and 36 mg/cm2 respectively. Estimation of S2A allowed calculation of the probability that the true BMD, for an observed BMD near the osteoporosis diagnostic threshold (T-score < -2.5), is not misclassified. Analysis of covariance established a significant operator-subject interaction at all sites, but only the total femur was associated with a significant difference between operators. The percentage of body fat was a significant covariate for the spine and total femur regions. ANOVA showed that the greater proportion of variance was instrument-related. The limitations of DXA as an analytical method are discussed.