Slide set 10 Stat402 (Spring 2016) Last update: March 22, 2016 Stat 402 (Spring 2016): Slide set 10 The 23 Factorial The Treatment combinations are: (1), a, b, ab, c, ac, bc, and abc. The effects A, B, C, AB, BC, AC, ABC are difined as in the 22 case. Example Simple Effects of A are: a-(1), ab-b, ac-c, abc-bc Main Effects of Factor A is the average of these: A = 1/4[(a − (1)) + (ab − b) + (ac − c) + (abc − bc)] = 1/4[abc + ab + ac + a − b − c − bc − (1)] 1 Stat 402 (Spring 2016): Slide set 10 Algebraic Identities for the 23 case • Contrasts for main effects and interactions: A B C AB • : : : : (a-1)(b+1)(c+1) (a+1)(b-1)(c+1) (a+1)(b+1)(c-1) (a-1)(b-1)(c+1) = = = = abc+ab+ac+a-bc-b-c-(1) abc+ab+bc+b-ac-a-c-(1) abc+bc+ac+c-ab-a-b-(1) abc+ab+c+(1)-ac-bc-a-b The interaction effect AB is the average of the two interaction effects at each level of C Effect AB at level 0 of C: 1/2[(ab-b)-(a-(1))] Effect AB at level 1 of C: 1/2[(abc-bc)-(ac-c)] AB = 1/2{1/2[ab − b − a + (1)] + 1/2[abc − bc − ac + c]} = 1/4[abc + ab + c + (1) − ac − bc − a − b] 2 Stat 402 (Spring 2016): Slide set 10 Algebraic Identities for the 23 case (Cont’d) The same procedure is used to define the AC and BC effects. These effects are called 2-factor or 1st order interactions (usually shortened to 2-fi’s). • Three-factor interactions (3-fi’s): It is defined as the average difference between the interaction effects of AB at each level of C: ABC = 1/2{1/2[abc − bc − ac + c] − 1/2[ab − b − a + (1)]} = 1/4[abc + a + b + c − ab − ac − bc − (1)] The signs for the above contrast can be obtained from the identity (a-1)(b-1)(c-1). 3 Stat 402 (Spring 2016): Slide set 10 Contrasts for all effects in the 23 factorial In the 23 factorial, the set of defining contrasts are: Treatment Combinations (1) a b ab c ac bc abc I + + + + + + + + A + + + + B + + + + Factorial Effect AB C AC BC + + + + + + + + + + + + + + + + ABC + + + + 4 Stat 402 (Spring 2016): Slide set 10 Notes 1. Each effect is a contrast of the observations. 2. These are orthogonal contrasts implying that the corresponding effects are independent. 3. In a 23 experiment, there are 8 different treatment combinations and hence there are 7 d.f. for treatment. What we have done so far is to partition these 7 d.f. to 7 orthogonal single d.f. contrasts. 4. In general, in a 2k experiment we can find (2k -1) orthogonal contrasts to partition the treatment sum of squares. 5. See Example 6.1 in Section 6.3 of Montgomery. 6. The general 2k factorial is discussed in Section 6.4 7. The unreplicated 2k factorial is discussed in Section 6.5. 5 Stat 402 (Spring 2016): Slide set 10 An Example of a 23 Factorial: Plasma Etch Experiment A 23 factorial design was used to develop a nitride etch process on a single-wafer plasma etching tool. There are three quantitative factors-gap between electrodes, gas flow and the RF power applied to the cathode. The response is the etch rate for silicon nitride. Each treatment combination was replicated twice and the 16 experimental runs were done in completely random order. 6 Stat 402 (Spring 2016): Slide set 10 Example 6.1 (Cont’d) • • • Since this is a completely randomized design with 2 replications per treatment combination, the standard analysis is carried out as follows: The model is yijkl = µijk + ijkl, i = 1, 2; j = 1, 2; k = 1, 2; l = 1, 2 iid where ijkl ∼ N (0, σ 2). The expected response from the treatment combination ijk is µijk which may be expressed using factorial effects as µijk = µ + τi + βj + (τ β)ij + γk + (τ γ)ik + (βγ)jk + (τ βγ)ijk The ANOVA table (from JMP) is SV Trt Error Total DF 7 8 15 SS 513,400.4375 18,020.5000 531,420.9347 MS 73342.9196 2 sE =2252.5625 F 32.56 7 Stat 402 (Spring 2016): Slide set 10 Example 6.1 (Cont’d) Now the 7 d.f. for Treatment Sum of Squares is partitioned into 7 single degree of d.f. sums of squares corresponding to the 7 factorial effects: Treatment Combination (1) a b ab c ac bc abc Contrast Divisor for Estimate Estimate of Effect Divisor for SS SSA = SSC = Observed Total 1154 1319 1234 1277 2089 1617 2138 1589 I + + + + + + + + 12,417 16 776.0625 16 (−813)2 = 41, 310.5625 16 (2449)2 SSBC = 16 A + + + + -813 8 -101.625 16 SSB = (59)2 16 = 217.5625 (−1229)2 = 374, 850.0625 SSAC = (−17)2 = 18.0625 16 SSABC = B + + + + 59 8 7.375 16 Factorial Effect AB C + + + + + + + + -199 2449 8 8 -24.875 306.125 16 16 16 SSAB = AC + + + + -1229 8 -153.625 16 BC + + + + -17 8 -2.125 16 ABC + + + + 45 8 5.625 16 (−199)2 = 2475.0625 16 = 94, 402.5625 (45)2 16 = 126.5625 8 Stat 402 (Spring 2016): Slide set 10 Example 6.1 (Cont’d) This results in the summary and ANOVA tables: 9 Stat 402 (Spring 2016): Slide set 10 Discussion 1. Since the above experiment was a replicated experiment it was possible to construct the above analysis of variance table for testing hypotheses of interest. 2. Many factorial experiments in industry are not replicated experiments, since even for a moderate number of factors the total number of treatment combinations in a 2k factorial is large. Thus limited availability of resources or high costs may inhibit experimenters from replicating factorial experiments. 3. The result of not replicating an experiment is that there will be no d.f. available for estimating the error variance on; hence the anova table cannot be used for testing hypotheses concerning main effects and interactions. 4. Other methods have to be devised to determine the effects that are significant from such experiments. See Section 6.5 for a discussion of the analysis of a single replicate of a 2k factorial experiment. 5. When an estimate of the error variance is available, as part of the analysis the estimates of effects are usually reported (instead of an analysis of variance table) as follows: 10 Stat 402 (Spring 2016): Slide set 10 Estimating Effects Estimating of Effects from the Plasma-Etch Example √ 2×sE s.e.(E)= √ n×2k = 2 √2252.56 = 23.73; t0.025,8 = 2.31 & 2.31 × 23.73 = 54.82 3 2(2 ) Thus approximate 95% CI’s are: Effects Main Effects Gap(A) C2F6 Flow (B) Power(C) Two-factor Interactions AB AC BC Three-factor Interactions ABC Estimate(E) ± t-value × s.e.(E) -101.625±54.82 7.375±54.82 306.125±54.82 -24.875 ±54.82 -153.625±54.82 -2.125±54.82 5.625 ±54.82 11 Stat 402 (Spring 2016): Slide set 10 Discussion • • • • Comparison of the estimates with their standard errors suggests that the bold items A, C and the two-factor interaction AC require interpretation, while the remaining effects could be due to noise or error. The main effect of a factor should be individually interpreted only if there is no evidence that the variable interacts with other variables. When there is evidence of one or more such interaction effects, the interacting variables should be considered jointly. In the above table there are large Gap (A) and Power (C ) effects, but since a large AC effect indicates interaction between the two, we make no statement of Gap (A) and Power (C ) effects alone. The effects of Gap (A) and Power (C) can best be considered using the two-way table of means shown below and the corresponding graph shown next page: Power (C) 275 325 Gap (A) 0.8 1.2 597.00 649.00 1059.75 801.50 12 Stat 402 (Spring 2016): Slide set 10 Discussion (contd.) Power 6 375 275 1056.75 801.5 597.0 649.0 - 0.8 • Gap 1.2 The AC interaction evidently arises due to a difference in effect of the change in power at the two different gaps. With Gap=0.8 the effect of temperature is to increase mean etch rate by almost 460 but with Gap=1.2, the mean increases by only 152. 13