In industrial experiments, even when an experiment is run in a completely random order, it is unusual to independently reset the level of hard-to-change factors which have the same level in successive runs. This causes the experiment to have more than one error term. The author studies experiments with exactly two error terms, one associated with hard-to-change factors and the other associated with easy-to-change factors. This assumption differs from the classical split-plot designs, which also consider two error terms.

The disadvantage of running an experiment in a completely randomized fashion, when there are inherently two error terms, is quantified in this thesis. Various ways of restricting the randomization, with corresponding consequences, are described. The author recommends some specific blocking procedures. L$\sp{\rm k}$ factorial and response surface designs are discussed. More details are given about situations with one hard-to-change factor or one easy-to-change factor. The derivation can be extended to cases with multiple error terms.