Milk analysis proficiency testing has been carried out monthly by a group 10 laboratories for about 10 years. Patterns of differences in total solids (TS) results among labs have been observed, but the causes of some of these differences (i.e., systematic increase or decrease in difference from the all-laboratory mean value as function of milk total solids level) have not been identified. Our objective was to identify specific sources that cause these variations in results among laboratories. The atmospheric force air oven method number 990.20 of the Association of Official Analytical Chemists was used. A set of 14 milks with a range of TS content from about 8.4 to 15.2% were tested in duplicate by 10 different labs and replicated in 4 mo. In mo 1 and 3, the 14 milks were tested on each of 3 different days by the same analyst and in mo 2 and 4, the 14 milks were tested on each of 3 different days by the different analyst within each lab. The differences from the all-lab mean with statistical outliers removed were analyzed using Proc GLM of SAS to test for effects of lab, day, sample, and their interactions. The ANOVA terms for lab and lab by sample interaction were significant (P < 0.05) and explained most of the variation. The lab by sample interaction manifested itself as some labs having upward slopes, while other laboratories had downward slopes of the residual plot of 28 differences from the 14 sample all lab means as TS increased. The direction of the slope, when it occurred, was a function of laboratory or technician within laboratory. Follow-up work was conducted to identify causes of this behavior of TS results. Oven temperatures and oven temperature recovery rates were within method specifications for all labs. Attention was focused on the interaction of the analyst and balance. At the various weighing steps, the weights were recorded 3 times: immediately, after the balance indicated the weight was stable, and 7 s after the indication of stability. Based on these observations, a series of sensitivity analyses were done. The accuracy of TS results was most sensitive to variation (both random and systematic) in control of the balance zero during the steps of weighing the empty pan and the pan plus dry solids.