82 7 Balance and non-orthogonality7.1 Confounding and efficiency factorsIn the split-plot design it is the main effect of one of the treatment factors that isestimated in the higher stratum. Statistically, we would say that this main effect isconfounded with whole plots within blocks. For the factor Variety in Section 5.1, thisis completely acceptable; the main interest in the trial was to look at the Nitrogen factorand the interaction between Nitrogen and Variety. However, on other occasions, wemay want all the main effects to be estimated with the extra precision that should beavailable in the bottom stratum, and so we may want the interactions to be estimated inthe higher strata instead.n 0 0 k n 0 0 k 0 0 0 0 n k n kThe plan above shows a design in which the interaction between the factors N and K isconfounded with blocks. The definition of the N × K interaction is that it is the differencebetween the effect of N estimated at the different levels of K. Here we have factors at twolevels 0 and n for N, and 0 and k for K. For the 0 level of K, the effect of adding N is givenby the mean of the plots with the combination (n, 0) minus the mean of the plots with(0, 0); while for K at level k, it is given by the mean of the plots with (n, k) minus themean of the plots with (0, k). So the difference between the two estimates (which givesthe interaction contrast) is{ mean of plots with (n, 0) + mean of plots with (0, k) } { mean of plots with (0, 0) + mean of plots (n, k) }The left-hand block above contains only combinations (n, 0) and (0, k), while the righthandblock contains only combinations (0, 0) and (n, k). Consequently the differencebetween the means of the plots in the two blocks also estimates the interaction: that is,the N × K interaction is confounded with blocks.Usually, in a situation like this, youwould have more than two blocks. Infact, the two blocks above are part of adesign with eight blocks, each with fourplots, that was used to study factors N, Kand D (see Yates, 1937, Design andAnalysis of Factorial Experiments, page21; also John, 1972, Statistical Designand Analysis of Experiments, page 135).The left-hand block in the plan is block3 of the design, and the right-hand blockis block 4. If we analyse just those twoblocks with treatment model N*K, theanalysis of variance table belowconfirms that the interaction isestimated in the Blocks stratum (and,Figure 7.1as we have analysed only these two blocks, there are no degrees of freedom left over forthe residual).
7.1 Confounding and efficiency factors 83Analysis of varianceVariate: Yield of potatoes in tons/acreSource of variation d.f. s.s. m.s. v.r.Blocks stratumN.K 1 0.56 0.56Blocks.*Units* stratumN 1 0.48 0.48 0.04K 1 29.86 29.86 2.25Residual 4 53.17 13.29Total 7 84.060 0 0 n k 0 n 0 d 0 k d n 0 0 0 k 0 0 0 d n k dn 0 0 0 k 0 n 0 d 0 k d 0 0 0 0 0 d n k 0 n k dn 0 0 0 0 d n k 0 0 k d 0 0 0 0 k 0 n 0 d n k d0 k 0 0 0 d n k 0 n 0 d 0 0 0 n 0 0 0 k d n k dThe plan for the whole design, above, illustrates some further sophistication. It is set upso that N.K.D is confounded in blocks 1 and 2, N.K in blocks 3 and 4, N.D in blocks 5and 6, and K.D in blocks 7 and 8. Thus, for example, N.K is estimated between blocks3 and 4, and within blocks 1, 2, 5, 6, 7 and 8. So 6/8 of the information about N.K is inthe Blocks.Plots stratum, and 2/8 is in the Blocks.Plots stratum. The main effectsof N, K and D can be estimated in every block: they are orthogonal to blocks and all theirinformation is in the Blocks.Plots stratum.The amount of information available about a term in a particular stratum is known asits efficiency factor. The efficiency factors of non-orthogonal terms (i.e. those whoseefficiency is less than one) are listed in the Information Summary, which can be obtainedby checking the Information box in the ANOVA Options menu.The whole design can be analysed using the general Analysis of Variance menu, with theDesign drop-down list box set to General Analysis of Variance, the Block Structure set toBlocks/Plots, and the Treatment Structure set to N*K*D. The analysis is shown below.