The purpose of this work was the study of the efficiency of a 3³ factorial design, confounded in blocks of 9 units, in comparison with the completely randomized design. Approximately 300 cotton experiments were conducted during the years of 57/58, 58/59, 59/60 on different kinds of soils of the State of São Paulo. From the results of the experiments in which confounding was employed, we estimated what the experimental error would have been if this technique had not been used; a comparison was mode between the two experimental errors and the efficiency was calculated. Confounding in blocks of 9 units has presented a gain in precision of about 20% when compared with the completely randomized design.