Slide set 8 Stat402 (Spring 2016) Last update: February 29, 2016 Stat 402 (Spring 2016): Slide Set 8 The Two-way Factorial in a Completely Randomized Design • Example 5.1: Measure maximum output voltage of storage batteries. All combinations of 3 plate materials and 3 temperatures are used in a completely randomized design with 4 replications per each combination. Nine treatment combinations allocated to 36 runs completely at random. The treatment arrangement is called a 3 × 3 factorial. 9 treatments combinations: 3 materials × 3 temperatures 1 Stat 402 (Spring 2016): Slide Set 8 • Analysis The analysis is best considered in two stages. The first part of the analysis ignores the treatment structure, i.e., that it is a factorial. Analyze the data as one-way classification with 9 treatments (i.e., 1 factor with 9 levels) and equal sample sizes. Data will look like: Treatment 1 Treatment 2 Treatment 3 . . . y111 y121 ... ... y112 y122 ... ... y113 y123 ... ... y114 y124 ... ... • Treatment 9 y331 y332 y333 y334 Model: yijk = µij + ijk , i = 1, . . . , 3 (Material); j = 1, . . . , 3 (Temperature) k = 1, . . . , 4 (Replications) i.i.d and ijk ∼ N (0, σ 2) 2 Stat 402 (Spring 2016): Slide Set 8 The Two-way Factorial in a CRD (Cont’d) • ANOVA SV d.f. Treatment 8 Error 27 Total 35 Now, consider the modeling of the treatment structure. We consider the non-additive model µij = µ + τi + βj + (τ β)ij for the treatment mean µij where τi = incremental effect due to plate material i βj = incremental effect due to plate temperature j (τ β)ij =incremental effect due to combination of material i and temp. j 3 Stat 402 (Spring 2016): Slide Set 8 This partitioning partitions the d.f. due to treatment above as follows: SV Trt Material Temp Material × Temp Error Total Look at the ANOVA table on d.f. 8 2 2 4 27 35 p.194 of Montgomery 4 Stat 402 (Spring 2016): Slide Set 8 Two-way Factorial in a Completely Randomized Design Data: Factor A 1 2 .. a 1 y111, . . . , y11n y211, . . . , y21n .. ya11, . . . , ya1n ȳ.1. Factor B 2 ... y121, . . . , y12n . . . ... .. ... ȳ.2. ... b y1ba, . . . , y1bn yab1, . . . , yabn ȳ.b. ȳ1 ȳ1.. ȳ2.. .. ȳa.. ȳ... where Pb Pn ȳi.. = ( j=1 k=1 yijk )/bn (Factor A Means) Pa Pn ȳ.j. = ( i=1 k=1 yijk )/an (Factor B Means) Pn ȳij. = k=1 yijk /n (Cell Means) P P P ȳ... = ( i j k yijk )/abn (Grand Mean) 5 Stat 402 (Spring 2016): Slide Set 8 Model: yijk = τi = βj = (τ β)ij = i = 1, . . . , a µ + τi + βj + (τ β)ij + ijk j = 1, . . . , b k = 1, . . . , n effect of i-th level of A a=# levels of Factor A effect of j-th level of B b=# levels of Factor B interaction effect of i-th level n=# of replications of each of A and j-th level of B treatment combination Estimates µ̂ τ̂i β̂j (τˆβ)ij = = = = ȳ... ȳi.. − ȳ... or µ̂ij = ȳij. ȳ.j. − ȳ... ȳij. − ȳi.. − ȳ.j. + ȳ... 6 Stat 402 (Spring 2016): Slide Set 8 ANOVA Table SV Treatment A B AB Error Total d.f. ab − 1 a−1 b−1 (a − 1)(b − 1) ab(n − 1) abn − 1 SS SST rt SSA SSB SSAB SSE SST MS M ST rt M SA M SB M SAB M SE (= s2E ) F M ST rt /M SE 1 M SA /M SE 2 M SB /M SE 3 M SAB /M SE Hypotheses Tested F statistic What is tested? 1 H0 : τ1 = · · · = τa = 0 vs Ha : at least one τi 6= 0 ≡ [H0 : µ̄1. = · · · = µ̄a. vs Ha : at least one ineq.] 2 H0 : β1 = · · · = βb = 0 vs Ha : at least one βj 6= 0 ≡ [H0 : µ̄.1 = · · · = µ̄.b vs Ha : at least one ineq.] 3 H0 : (τ β)ij = 0 for all i, j vs Ha : at least one 6= 0 ≡ [H0 : no interaction vs Ha : interaction] 7 Stat 402 (Spring 2016): Slide Set 8 Hypotheses Tested (Cont’d) Predicted Values: ŷijk = µ̂ + τ̂i + β̂j + (τˆβ)ij = ȳ... + (ȳi.. − ȳ...) + (ȳ.j. − ȳ...) + (ȳij. − ȳi.. − ȳ.j. + ȳ...) Residuals rijk = yijk − ŷijk = yijk − (ȳ... + (ȳi.. − ȳ...) + (ȳ.j. − ȳ...) + (ȳij. − ȳi.. − ȳ.j. + ȳ...)) = (yijk − ȳij.) SSRes: a X b X n X i=1 j=1 k=1 rijk = a X b X n X (yijk − ȳij.)2 i=1 j=1 k=1 8 Stat 402 (Spring 2016): Slide Set 8 Algebraic Decomposition of Total Variation Total SS = SST = SST rt = n Pa i=1 SSA SSB SS AB Pa i=1 Pb Pb j=1 (yij j=1 Pn 2 (y − ȳ ) ijk ... k=1 − ȳij.)2 = Pa = bn( i=1(ȳi.. − ȳ...)2) Pb = an( j=1(ȳ.j. − ȳ...)2) Pa Pb = n( i=1 j=1(ȳij. − ȳi.. − ȳ.j. + ȳ...)2) SSE = Pa i=1 Pb j=1 Pn k=1 (yijk − ȳij.)2 SST = SST rt + SSE = SSA + SSB + SSAB + SSE 9 Stat 402 (Spring 2016): Slide Set 8 Suggested procedure • • • • • Test the No Interaction Hypothesis first. If no significant interaction between A and B then: Go ahead and test for main effects. These tests compare factor means i.e. means of each factor averaged over the levels of the other factor If A or/and B are significant, then use methods such as LSD, Tukey, Prelanned comparisons, Orthogonal Polynominals to compare factor means. A 100(1 − α)% C.I. for τ1 − τ2 (or µ̄1. − µ̄2.) q 2 (ȳ1.. − ȳ2..) ∓ t α2 ,ν · sE bn A 100(1 − α) C.I. for β1 − β2 (or µ̄.1 − µ̄.2) q 2 (ȳ.1. − ȳ.2.) ± t α2 ,ν · sE an where ν = ab(n − 1) s2E = M SE 10 Stat 402 (Spring 2016): Slide Set 8 Interpretation of Results: Battery Life Data ANOVA table Interaction Plot 11 Stat 402 (Spring 2016): Slide Set 8 Comparison of Means: Battery Life Data • • • • Interaction is significant This implies that rather than comparing battery means or material means, we need to make comparisons among the 9 cell means (observe the interaction plot) The interaction plot suggests that it makes sense to compare the average battery life for the material types at each of the temperature. For example, from the JMP output, we have, that all pairs of material type means at each of temperatures 15 and 125 are not different from each other as evidenced by the interaction plot. From JMP, 95% CI’s for differences in Material means at Temperature 70 are: µ1 − µ2 : (−100.2, −24.8)∗ µ1 − µ3 : (−126.2, −50.8)∗ µ2 − µ3 : (−63.7, 11.7005) 12