# Unit 4: Full Factorial Experiments at Two Levels

Source : Chapter 4 (sections 4.1-4.4, 4.6-4.12, 4.15).
• An Epitaxial Layer Growth Experiment (Section 4.1). • Basic concepts for 2k designs (Section 4.2). • Factorial effects and plots (Section 4.3). • Using Regression to Compute Factorial Effects (Section 4.4). • Fundamental principles (Section 4.6). • Comparisons with ”one-factor-at-a-time” approach (Section 4.7). • Normal and Half-normal plots for detecting effect signiﬁcance (Section 4.8). • Lenth’s Method (Section 4.9). • Nominal-the best problem, quadratic loss function (Section 4.10). • Use of Log Sample Variance for Dispersion Analysis (Section 4.11). • Analysis of Location and Dispersion: Epitaxial Growth Experiment (Section 4.12). • Blocking in 2k design (Section 4.15).
1

Epitaxial Layer Growth Experiment
• An AT&T experiment based on 24 design; four factors each at two levels. There are 6 replicates for each of the 16 (=24 ) level combinations; data given on the next page. Table 1: Factors and Levels, Adapted Epitaxial Layer Growth Experiment Factor A. B. C. D. susceptor-rotation method nozzle position deposition temperature (◦ C) deposition time − Level +

continuous 2 1210 low

oscillating 6 1220 high

• Objective : Reduce variation of y (=layer thickness) around its target 14.5 µm by changing factor level combinations.
2

Data from Epitaxial Layer Growth Experiment
Table 2: Design Matrix and Thickness Data, Adapted Epitaxial Layer Growth Experiment
Factor Run 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 A − − − − − − − − + + + + + + + + B − − − − + + + + − − − − + + + + C − − + + − − + + − − + + − − + + D + − + − + − + − + − + − + − + − 14.506 12.886 13.926 13.758 14.629 14.059 13.800 13.707 15.050 14.249 13.327 13.605 14.274 13.775 13.723 14.031 14.153 12.963 14.052 13.992 13.940 13.989 13.896 13.623 14.361 13.900 13.457 13.190 13.904 14.586 13.914 14.467 Thickness 14.134 13.669 14.392 14.808 14.466 13.666 14.887 14.210 13.916 13.065 14.368 13.695 14.317 14.379 14.913 14.675 14.339 13.869 14.428 13.554 14.538 14.706 14.902 14.042 14.431 13.143 14.405 14.259 14.754 13.775 14.808 14.252 14.953 14.145 13.568 14.283 15.281 13.863 14.461 14.881 14.968 13.708 13.932 14.428 15.188 13.382 14.469 13.658 15.455 14.007 15.074 13.904 15.046 13.357 14.454 14.378 15.294 14.255 13.552 14.223 14.923 13.382 13.973 13.578 y ¯ 14.59 13.59 14.24 14.05 14.65 13.94 14.40 14.14 14.67 13.72 13.84 13.90 14.56 13.88 14.30 14.11 s2 0.270 0.291 0.268 0.197 0.221 0.205 0.222 0.215 0.269 0.272 0.220 0.229 0.227 0.253 0.250 0.192 ln s2 -1.309 -1.234 -1.317 -1.625 -1.510 -1.585 -1.505 -1.537 -1.313 -1.302 -1.514 -1.474 -1.483 -1.374 -1.386 -1.650

3

2k Designs: A General discussion
• 2 × 2 × . . . × 2 = 2k design. • Planning matrix vs model matrix (see Tables 4.3, 4.5). • Run order and restricted randomization (see Table 4.4). • Balance: each factor level appears the same number of times in the design. • Orthogonality : for any pair of factors, each possible level combination appears the same number of times in the design. • Replicated vs unreplicated experiment.

4

4 • • 14. Adapted Epitaxial Layer Growth Experiment 5 .0 • 13. • Advantages of factorial designs (R.8 - A + - B + - C + - D + Figure 1: Main Effects Plot.Main effects and Plots • Main effect of factor A: ME(A) = z ¯(A+) − z ¯(A−). • Informal analysis using the main effects plot can be powerful.2 • • • • • 14. 14.Fisher): reproducibility and wider inductive basis for inference.A.

Interaction Effects • Conditional main effect of B at + level of A: ME(B|A+) = z ¯(B + |A+) − z ¯(B − |A+). • Two-factor interaction effect between A and B: INT (A. The latter is commonly used in standard texts. 2 2 (1) The ﬁrst two deﬁnitions in (1) give more insight on the term ”interaction” than the third one in (1). 6 . B) = = = 1 {ME(B|A+) − ME(B|A−)} 2 1 {ME(A|B+) − ME(A|B−)} 2 1 1 {z ¯(A + |B+) + z ¯(A − |B−)} − {z ¯(A + |B−) + z ¯(A − |B+)}.

2 A+ A-- 14.2 14.0 A+ 14.4 14.6 C-- B+ 14.8 13. Adapted Epitaxial Layer Growth Experiment 7 .6 13.0 13.0 14.4 C+ 14.6 C + D + 13.4 14.6 14.6 B + C + 13.2 14.Interaction Effect Plots AB AC AD 14.4 14.4 B-B+ 14.2 14.0 B-- 13.4 A+ A-- 14.8 13.6 A-- 14.6 14.0 D + Figure 2: Interaction Plots.8 14.6 14.8 14.8 13.8 13.6 13.0 14.6 13.2 D + BC BD CD 14.6 14.6 13.2 14.

0 13. An antagonistic plot suggests a more complex underlying relationship than what the data reveal.6 13.Synergistic and Antagonistic Plots • An A-against-B plot is synergystic if ME(B|A+)ME(B|A−) > 0 and antagonistic if ME(B|A+)ME(B|A−) < 0.4 14.6 D + 13.6 C+ D+ D-- 14.2 14.8 14.0 C + Figure 3: C-against-D and D-against-C Plots. Adapted Epitaxial Layer Growth Experiment 8 . CD DC 14.4 14.6 C-- 14.2 14.8 13.

A2 . Var(θ) = N /2 N /2 N σ2 = variance of an observation. . 2 2 σ 4 2 σ ˆ + = σ . C|B−) = INT(B. . . . C|A−). Ak−1 |Ak +) − INT(A1 . C|B+) 2 2 2 1 1 1 − INT(A. If N is the total number of observations. 9 . 2 2 2 = 1 1 INT(A1 . . B|C−) = INT(A. 2 2 INT(A. C|A+) − INT(B. . A2 . . Ak−1 |Ak −). B|C+) − INT(A. . A2 . where z ¯+ and z ¯− are averages of one half of the observations.More on Factorial Effects 1 1 1 INT(A. Ak ) • A general factorial effect ˆ =z θ ¯+ − z ¯− . . . C) = INT(A1 . B. . .

x6 = x2 x3 .e. Use the regression model zi = β0 + ∑ β j xi j + εi . The regression (i. j=1 7 where i = ith observation.. whose columns are denoted by x1 . The interactions AB.Using Regression Analysis to Compute Factorial Effects Consider the 23 design for factors A. x7 = x1 x2 x3 (see Table 3). x5 = x1 x3 . x2 and x3 (=1 or -1). B and C. ABC are then equal to x4 = x1 x2 . least squares) estimate of β j is ˆj β = 1 (z ¯(xi j = +1) − z ¯(xi j = −1)) 1 − (−1) =1 2 (factorial effect of variable x j ) 10 . AC. BC.

Model Matrix for 23 Design Table 3: Model Matrix for 23 Design 1 A − − − − + + + + 2 B − − + + − − + + 3 C − + − + − + − + 12 AB + + − − − − + + 13 AC + − + − − + − + 23 BC + − − + + − − + 123 ABC − + + − + − − + 11 .

045 -0.030 0.042 0.030 -0.112 0.016 -0.078 0.029 0.020 ln s2 0.085 -0.003 0.010 0.078 0.080 0.032 0.056 0.058 -0. Adapted Epi-Layer Growth Experiment Table 4: Factorial Effects.098 0.490 0.093 -0.008 -0. Adapted Epitaxial Layer Growth Experiment Effect A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD y ¯ -0.050 0.110 0.026 -0.345 0.000 -0.118 -0.103 12 .Factorial Effects.025 -0.173 -0.

McCullagh and Nelder called it the Marginality Principle.Fundamental Principles in Factorial Design • Effect Hierarchy Principle (i) Lower order effects are more likely to be important than higher order effects. For modeling. at least one of its parent factors should be signiﬁcant. Effect hierarchy and sparsity principles are more effective/relevant for screening experiments. This is similar to the Pareto Principle in quality investigation. (ii) Effects of the same order are equally likely to be important. • Effect Sparsity principle (Box-Meyer) The number of relatively important effects in a factorial experiment is small. 13 . • Effect Heredity Principle (Hamada-Wu) In order for an interaction to be signiﬁcant.

3 0.6 S slow fast slow fast slow fast slow fast 14 Percent Burned 11 17 25 29 02 09 37 40 .3 0.6 0.3 0.6 0.3 0.6 0.One-Factor-At-A-Time (ofat) Approach Table 5: Planning Matrix for 23 Design and Response Data For Comparison with One-Factor-At-A-Time Approach Factor P 1200 1200 1200 1200 1400 1400 1400 1400 R 0.

15 .3 1200 P 1400 Figure 4: The Path of a One-Factor-At-A-Time Plan The three steps of ofat as illustrated in the arrows in Figure 4 are detailed in steps 1-3 on page 174 of WH.6             R   x  0.) x     ©       x c c ' x       x   fast S slow x x x 0.One-Factor-At-A-Time (ofat) Approach (Contd.

Conclusions for analysis not as general. For ofat to have the same precision. to be comparable to a 2k design. It requires more runs for the same precision in effect estimation. totaling 16 runs. 2. It can miss optimal settings. It cannot estimate some interactions. For points 2 − 4.Disadvantages of ofat Approach Relative to Factorial Approach 1. totaling (k + 1)2k−1 . see Figure 4. ofat would require 2k−1 runs at each of the k + 1 corners on its path. 16 . In the example. 3. The ratio is (k + 1)2k−1 /2k = (k + 1)/2. the 23 design requires 8 runs. 4. In general. each of the 4 corners on the ofat path needs to have 4 runs.

Needs a new outlook. Easier to think and focus on one factor each time. • No exposure to statistical design of experiments.Why Experimenters Continue to Use ofat? • Most physical laws are taught by varying one factor at a time. • Challenges for DOE researchers: To combine the factorial approach with the good intuition rendered by the the ofat approach. 17 . • Experimenters often have good intuition about the problem when thinking in this mode.

Φ−1 ([i − 0.4 • • • ABC BCD B effects • • • AC • • A AD • • ACD BD • BC • • ABD ABABCD 0. are the factorial effect estimates (example in Table 4). • Suppose θ ˆ ( 1) ≤ · · · ≤ θ ˆ (I ) . I .2 0. i = 1. · · · .Normal Plot of Factorial Effects ˆ i . Adapted Epitaxial Layer Growth Experiment 18 . Normal probability plot (see Unit 2): Order them as θ ˆ i (vertical) vs.5]/I ) (horizontal) θ • D 0.0 C -0.2 • CD -1 0 normal quantiles 1 Figure 5: Normal Plot of Location Effects.

Null hypothesis H0 : all factorial effects = 0 . any effect whose corresponding point falls off the line is declared signiﬁcant (Daniel. Daniel’s idea is to use the normal curve as the reference distribution. 19 . θ • Induction Step. • In Figure 5.Use of Normal Plot to Detect Effect Signiﬁcance • Deduction Step. D. ˆ i ∼ N (0. σ2 ) and the resulting normal plot should follow a straight line. By ﬁtting a straight line to the middle group of points (around 0) in the normal plot. In t test. no estimate of σ2 is required. Method is informal and judgemental. For unreplicated experiments. 1959). Method is especially suitable for unreplicated experiments. • Unlike t or F test. Under H0 . s2 is the reference quantity. CD (and possibly B?) are signiﬁcant.

0 0.5 0.5 • •G D O effects • • • • BE • • • • F J • A I H M -0.5 I •D K • • •J A • • •F • 0.0 1.0 HM B E 0.5 half normal quantiles Figure 6: Comparison of Normal and Half-Normal Plots 20 .5 • L N • -2 K • -1 0 normal quantiles 1 2 • C absolute effects 1.0 2.5 2.5 • • • L• N G• O • 0.Normal and Half Normal Plots • C 1.0 1.5 1.

one may declare C.Visual Misjudgement with Normal Plot Potential misuse of normal plot : In Figure 6 (top). because they “deviate” from the middle straight line. This points to a potential visual misjudgement and misuse with the normal plot. This is wrong because it ignores the obvious fact that K and I are smaller than G and O in magnitude. 21 . by following the procedure for detecting effect signiﬁcance. K and I are signiﬁcant.

half-normal plots will be used for detecting effect signiﬁcance (2) 22 . This would avoid the potential misjudgement between the positive and negative values. . 2k − 1. .5[i − 0. for i = 1. only C is declared signiﬁcant.Half-Normal Plot ˆ (i) values as |θ ˆ |( 1) ≤ · · · |θ ˆ |(I ) and plot them on the • Idea: Order the absolute θ positive axis of the normal distribution (thus the term “half-normal”). (Φ−1 (0.5 + 0. |θ (i) • In Figure 6 (bottom). Notice that K and I no longer stand out in terms of the absolute values.5]/I ). . • For the rest of the book. • The half-normal probability plot consists of the points ˆ | ). .

ˆ i | with |θ ˆ i | < 2.5 · median|θ (Justiﬁcation : If θi = 0 and error is normal.) 23 . • Lenth’s Method 1. Use of median gives “robustness” to outlying the standard deviation of θ values.5s0 } |θi |. s0 = 1.5 · median{|θ ˆ i |<2. Compute the pseudo standard error ˆ PSE = 1.A Formal Test of Effect Signiﬁcance : Lenth’s Method • Sometimes it is desirable to have a formal test that can assign p values to the effects.5s0 and where the median is computed among the |θ ˆ i |. The following method is also available in packages like SAS. s0 is a consistent estimate of ˆ i .

) 2. θ is declared signiﬁcant. PSE ˆi If |tPSE . for each i.A Formal Test of Effect Signiﬁcance (Contd.i = . 24 . • Two versions of the critical values are considered next.i | exceeds the critical value given in Appendix H (or from software). Compute ˆi θ tPSE .

for i = 1.i has the same distribution under H0 for all i. I .i | > EERα |H0 ) = α. (Why?) 25 . . . (Note : Because θi = 0. normal error. i = 1.i | > EERα for at least one i. In screening experiments. I |H0 ) = Prob( max |tPSE . 1≤i≤I • EER accounts for the number of tests done in the experiment but often gives conservative results (less powerful). The EER critical values can be inﬂated by considering many θi values. .Two Versions of Lenth’s Method • Individual Error Rate (IER) H0 : all θi ’s = 0. IER is more powerful and preferable because many of the θi ’s are negligible (recall the effect sparsity principle). IERα at level α is determined by Prob(|tPSE . · · · . .) • Experiment-wise Error Rate (EER) Prob(|tPSE . tPSE .i | > IERα |H0 ) = α.

5s0 = 2.117.292. 1. • p values of effects can be obtained from packages or by interpolating the critical values in the tables in appendix H.117 = 0.8.45 (for I = 15) will not detect any effect signiﬁcance.63 for I = 15. 26 . By comparing with the |tPSE | values.058.01.078. ˆ Then median{|θ ˆ i |<2. s0 = 1. Use of EER0. (See page 182 for illustration).Illustration with Adapted Epi-Layer Growth Experiment ˆ i | = 0.5 × 0. For α = 0.01 level.345 (CD).087.058 = 0. which eliminates 0. IER0. median|θ Trimming constant 2. In Table 4.490 (D) and 0.5 × 0.5s0 } |θi | = 0.01 = 3.078 = 0. PSE = 1.10 of section 4.01 = 6. thus conﬁrming the half-normal plot analysis in Figure 4. Analysis of the |tPSE | values for ln s2 (Table 6) detects no signiﬁcant effect (details on page 182 of WH). The corresponding |tPSE | values appear in Table 6. 2.5 × 0. D and CD are signiﬁcant at 0.

87 1.63 0.26 0.78 0.34 3.23 ln s2 0.07 0.63 27 .71 0.05 1.46 1.25 1.89 0.00 0.90 1. Adapted Epitaxial Layer Growth Experiment Effect A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD y ¯ 0.35 0.67 0.27 0.97 1.16 1.90 5.57 0.41 0.|tPSE | Values for Adapted Epi-Layer Growth Experiment Table 6: |tPSE | Values.51 0.34 1.09 1.13 0.29 0.67 0.99 0.

Procedure is effective only if an adjustment factor can be found. This is often done on engineering ground. t )) = cVar(y) + c(E (y) − t )2 . E (L(y. Deﬁne a quantitative loss due to deviation of y from t . location and spacing of markings on the dial of a weighing scale). t ) = c(y − t )2 . mold size in tile fabrication.Nominal-the-Best Problem • There is a nominal or target value t based on engineering design requirements. Quadratic loss : L(y. A factor in step (ii) is an ad justment factor if it has a signiﬁcant effect on E (y) but not on Var(y). (Examples of adjustment factors : deposition time in surface ﬁlm deposition process. (ii)Select the level of a factor not in (i) to move E (y) closer to t . • Two-step procedure for nominal-the-best problem: (i) Select levels of some factors to minimize Var(y). 28 .

i. • Variance stabilizing property: next page. • Better for variance prediction. z ˆ = predicted value of ln σ2 . always nonnegative. Suppose z = ln s2 . ˆ then eZ = predicted variance of σ2 .∞) to ln s2 over (−∞. over (−∞. Log converts multiplicity into additivity.Why Take ln s2 ? • It maps s2 over (0. ∞). Regression and ANOVA assume the responses are nearly normal. 29 . ∞). • Most physical laws have a multiplicative component.e.

h a smooth function. 2(ni − 1)−1 . (3) • X a random variable. is nearly constant for ni − 1 ≥ 9. 2 2 ln s2 i = ln σi + ln(χni −1 /(ni − 1)). Then E (X ) = 1 and Var(X ) = ν . ν ν ′ 22 2 −1 2 In (2). σ2 ) . ν = ni − 1. 30 . Var(h(X )) ≈ [h′ (E (X ))]2 Var(X ) • Take X = χ2 ν ν 2 and h = ln. Then ( n − 1 ) s ∑ j=1 i j ¯i )2 ∼ σ2 i i i i χni −1 . ln s2 i ∼ N (ln σi . 2(ni − 1) ).ln s2 as a Variance Stabilizing Transformation 2 = n i (y − y 2 • Assume yi j ∼ N (0. The variance of ln si . χ2 ν ν (4) • Applying (3) to X = leads to 2 Var(ln(X )) ≈ [h (1)] = .

060 15.108 15.311 -5. Table 7: Design Matrix and Thickness Data.868 14.870 14.918 13.886 14.718 13.958 14.003 0.430 -1.736 13.704 14.771 -5.850 13.938 14.016 0.936 14.402 14.326 15.037 14.384 13.614 14.544 14.000 14.592 y ¯ 14. Original Epitaxial Layer Growth Experiment Design A − − − − − − − − + + + + + + + + B − − − − + + + + − − − − + + + + C − − + + − − + + − − + + − − + + D + − + − + − + − + − + − + − + − 14.932 14.722 13.656 14.896 14.921 14.130 Thickness 14.432 15.350 14.082 14.165 14.876 14.415 13.860 13.778 14.126 15.772 13.182 14.876 14.870 14.907 14.588 14.003 0.256 14.640 14.878 14.988 14.579 -1.001 0.738 14.202 14. Tsui and Wu (1991).237 -1.034 13.810 14.914 14.984 -5.485 -4.774 13.932 13.324 13.821 13.328 14.932 14.258 13.005 0.003 0.154 14.118 -2.896 14.470 13.880 14.962 14.917 -5.648 13.054 14.226 0.147 0.215 0.812 13.121 0.970 15.448 14.914 s2 0.004 0.914 13.536 14.846 14.184 13.682 13.843 13.036 13.070 ln s2 -5.972 14.272 14.970 14.653 31 .794 13.058 14.228 15.738 15.264 15.327 0.206 0.704 -6.170 13.860 14.032 14.926 13.866 14.768 14.758 13.198 13.002 0.682 14.107 -6.538 -2.044 13.487 -1.088 0.172 14.996 15.784 13.778 13.964 14.028 14.490 14.274 15.774 13.504 14.757 13.916 -2.860 14.044 14.116 -1.424 14.888 14.Epi-layer Growth Experiment Revisited Original data from Shoemaker.136 14.932 14.092 14.

060 0.305 0.002 0.067 -0.109 32 .836 -0. Original Epitaxial Layer Growth Experiment Effect A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD y ¯ -0.036 ln s2 3.055 0.098 0.142 -0.331 0.037 0.314 0.582 -0.214 0.086 -0.025 0.047 0.428 0.077 0.632 -0.074 -0.056 0.834 0.109 0.010 -0.032 -0.078 0.494 0.Epi-layer Growth Experiment: Effect Estimates Table 8: Factorial Effects.335 0.

0 • • • • • • • CD • D 0.0 0.5 2.5 0 • • • • 0.0 2.4 0.5 1.0 • 0.5 half-normal quantiles half-normal quantiles Figure 7 : Location effects 33 Figure 8 : Dispersion effects .0 1.Epi-layer Growth Experiment: Half-Normal Plots • D • A 0.6 absolute effects 0.5 2.0 1.8 absolute effects 0.2 1 2 3 • • • • • • • • • • • • • C B • • 0.5 1.0 2.

33 seconds.418xD .389 + 0. ˆ +β α ˆ0 + γ ˆ A xA = −3.389 + 0.917xA .266 to satisfy 14.418xD If xD = 30 and 40 seconds for D = − and +. γ • Two-step procedure: (i) Choose A at − level (continuous rotation).772 + 1.Epi-layer Growth Experiment: Analysis and Optimization • From the two plots. xD = 0. (ii)Choose xD = 0.5 = 14.266(5) = 36. D is an adjustment factor.266 would correspond to 35+0. D is signiﬁcant for y ¯ and A is signiﬁcant for z = ln s2 . 34 . Fitted models : y ˆ = z ˆ = ˆ D xD = 14.

Yet in practice.058)2 = 0. any recommended setting should be veriﬁed by conﬁrmation experiments.Epi-layer Growth Experiment: Analysis and Optimization (Cont’d) • Predicted variance ˆ 2 = exp(−3. among the observed values of s2 . 35 . three or four are comparable.917(−1)) = (0.0034. σ This may not be too optimistic because.772 + 1.

• Why sacriﬁcing 123? ans: Effect hierarchy principle. The block effect B and the interaction 123 are called confounded. B = 123. 36 . Use the 123 column in Table 9 to deﬁne the blocking scheme: block I if 123 = − and block II if 123 = +. Notationally. Therefore the block effect estimate y ¯(II ) − y ¯(I ) is identical to the estimate of the 123 interaction y ¯(123 = +) − y ¯(123 = −).2k Designs in 2q Blocks • Example: Arranging a 23 design in 2 blocks (of size 4). this blocking scheme increases the precision in the estimates of main effects and 2ﬁ’s by arranging 8 runs in two homogeneous blocks. • By giving up the ability to estimate 123.

Arrangement of 23 Design in 2 Blocks Table 9: Arranging a 23 Design in Two Blocks of Size Four Run 1 2 3 4 5 6 7 8 1 − − − − + + + + 2 − − + + − − + + 3 − + − + − + − + 12 + + − − − − + + 13 + − + − − + − + 23 + − − + + − − + 123 − + + − + − − + Block I II II I II I I II 37 .

Confounding relationships: B1 = 12. The 4 blocks I. B2 = 13. II. Thus 12. 38 .13 and 23 are confounded with block effects and thus sacriﬁced.A 23 Design in 4 Blocks • Similarly we can use B1 = 12 and B2 = 13 to deﬁne two independent blocking variables. III and IV are deﬁned by B1 = ± and B2 = ± : B1 B2 − + − I II + III IV • A 23 design in 4 blocks is given in Table 9. B1 B2 = 12 × 13 = 23.

Arranging a 23 Design in 4 Blocks Table 10: Arranging a 23 Design in Four Blocks of Size Two Run 1 2 3 4 5 6 7 8 1 − − − − + + + + 2 − − + + − − + + 3 − + − + − + − + B1 (= 12) + + − − − − + + B2 (= 13) + − + − − + − + 23 + − − + + − − + 123 − + + − + − − + block IV III II I I II III IV 39 .

b1 is said to have less aberration than scheme b2 . let r = smallest i such that gi (b1 ) = gi (b2 ).12.1 of WH. 40 . A blocking scheme has minimum aberration if no other blocking schemes have less aberration. Must require g1 (b) = 0 (because no main effect should be confounded with block effects). If gr (b1 ) < gr (b2 ). let gi (b) = number of i-factor interactions that are confounded with block effects. {I. • Minimum aberration blocking schemes are given in Table 4A. (This is justiﬁed by the effect hierarchy principle). For any two blocking schemes b1 and b2 .13. • Theory is developed under the assumption of no block × treatment interactions. • For any blocking scheme b. see page 196 of WH.23} forms the block deﬁning contrast subgroup for the 23 design in 4 blocks. For a more complicated example (25 design in 8 blocks).Minimum Aberration Blocking Scheme • On page 38.