This paper describes a comparative study of the three most commonly used field density tests: sand-cone, nuclear, and drive-cylinder. In order to evaluate the range of variability of these tests, a large-scale soil compaction apparatus was constructed so that soil could be compacted to a known water content and dry density under close laboratory control conditions before running field density tests. The equipment consisted of a 4-ft (122-cm) mold, and a hydraulic system for compacting the soil in ten 4-in. (10-cm) thick layers inside the mold. A cohesive soil with gravel up to 3/4-in. (19 mm) in size was used. Five series of tests were performed on this soil compacted to five different water contents and dry densities. In each series, side-by-side sand-cone, nuclear, and 3-in. (7.6-cm) drive-cylinder tests were made. It was found that sand-cone test results were closer to the placement values than the nuclear test results. This was partly due to inaccuracies in water content readings by the nuclear device. When the water content data measured by the nuclear device were ignored and the placement water contents measured by oven-drying were used instead, the results of the nuclear tests became more accurate, but the range of variability in the measured data did not decrease significantly. The drive-cylinder test had a bias toward underestimating the actual field density, primarily due to sampling disturbance. Despite careful control of test conditions, data from all three test methods had a wide range of variability.