The Box-Behnken three-level designs have proved useful for fitting sec
ond-order response surfaces to experimental data. Here, we discuss an
aspect of their value in the detection of lack of model fit. The degre
es of freedom in excess of those needed to fit the model can be used a
dvantageously for contrasts which estimate certain higher-order depart
ures from the model. These can be expressed in various forms. Some for
ms, in which these contrasts are orthogonal, permit the plotting of th
e contrasts in a normal plot for easy detection of model bias.