• Stat 402 (Spring 2016): slide set 13 Rep 2 Block III Block IV abc ab a ac b bc c (1) Rep 3 Block V Block VI abc ab a ac b bc c (1) 2 If we carry out a similar analysis as before (same for each replicate), we see that we can estimate A, B, C, AB, AC and BC effects free of block effects and that ABC interaction effect is confounded in each of the reps. Rep 1 Block I Block II abc ab a ac b bc c (1) A 2 Factorial in Blocks of Size 4 (Cont’d 1) 3 Last update: April 22, 2016 Stat402 (Spring 2016) Slide set 13 • • • • • • Stat 402 (Spring 2016): slide set 13 = = ABC 1 4 [μabc + μa + μb + μc + 14 [abc + a + c + b 3 − μab − μac − μbc − μ(1)] + 12 (β1 − β) − ab − ac − bc − (1)] + β1 + abc + μa + β1 + a + μb + β1 + b + μc + β1 μab − β2 − ab − μac − β2 − ac − μbc − β2 − bc −μ(1) − β2 − (1)] 1 4 [μabc +c − 1 4 [contrast = − μb − μc − μbc − μ(1)] − b − c − bc − (1)] of means] + 14 [contrast of errors] 1 4 [μabc + μab + μac + μa + 14 [abc + ab + ac + a −μ(1) − β2 − (1)] β2 − bc Stat 402 (Spring 2016): slide set 13 1 4 [μabc + β1 + abc + μab + β2 + ab + μac + β2 + ac +μa + β1 + a − μb − β1 − b − μc − β1 − c − μbc − = Â = For example using the appropriate model, 1 1/2[abc + a + b + c − ab − ac − bc − (1) Put all the treatment combinations with + sign in one block and the treatment combinations with - sign in the other. Now consider a 23 factorial experiment, where we only have blocks of size 4 available (i.e., need 2 blocks to accomodate the 8 treatment combination.) Suppose we are going to take 3 replications, i.e., each treatment combination will occur in the design 3 times. We also decide to confound the ABC interaction in each of the replications. To do this, first look at the defining contrast of the ABC interaction effect. Since ABC:(a − 1)(b − 1)(c − 1), ABC effect is defined by A 2 Factorial in Blocks of Size 4 3 DF 6 1 1 1 1 1 1 5 12 23 Block abc a b c Rep 3 Block V Block VI abc ab a ac bc b (1) c 6 Here we will confound different effects with blocks in each rep: − in Rep 1 confound ABC with blocks, − in Rep 2 confound AB with blocks, − in Rep 3 confound BC with blocks. • Rep 2 Block III Block IV abc a c b ab ac (1) bc Instead of confounding the same effect in each replicate we may decide to confound a different effect in each of the replicates. • Rep 1 I Block II (1) ab ac bc Consider the same 23 factorial used in the example used before, where we have blocks of size 4 and replicate the experiment 3 times. We say that the effects ABC, AB,and BC are partially confounded with blocks. • 7 We have not lost 100% of the information of any one effect as before but have lost part of the information about AB, BC, ABC. We can estimate BC from Rep 1 and Rep 2 or use 66.7% information. • • We can estimate AB effect from Rep 1 and Rep 3 or use 66.7% of the information. We can estimate ABC effect from Rep 2 and Rep 3, thus using 66.7% of the information. • • We can estimate A, B, C, and AC from all 3 reps, thus using 100% of the available information. Partial Confounding (contd.) Stat 402 (Spring 2016): slide set 13 Stat 402 (Spring 2016): slide set 13 • For example, in Rep 1 confound ABC with blocks, in Rep 2 confound AB with blocks, in Rep 3 confound BC with blocks. • 5 A 2 Factorial in Blocks of Size 4 3 Consider the same 23 factorial used in the example used before, where we have blocks of size 4 and replicate the experiment 3 times. Instead of confounding the same effect in each replicate we may decide to confound a different effect in each of the replicates. 1. We have been able to put 12 of the 8 treatment combinations in 1 block and the other half in another and come up with a design which gives all the information provided by a complete block design except that one effect is confounded with blocks. 2. See Example 7-2 for a computational example of a factorial in an incomplete block experiment. Remarks: Stat 402 (Spring 2016): slide set 13 • • Partial Confounding 4 Notes: No information about ABC is available from the above analysis. SV Treatments A B C AB BC AC Blocks Error Total We say in this case that the ABC interaction is completely confounded with blocks, meaning that it cannot be estimated from any of the reps (replications). A 2 Factorial in Blocks of Size 4 (Cont’d 2) Stat 402 (Spring 2016): slide set 13 • • • 3 DF 7 1 1 1 1 1 1 1 5 11 23 100% 100% 66.6% 100% 100% 66.7% 66.7% (Relative Information) the the the the the AB effect ABC effect BC effect ABC effect AC effect 10 Remarks − If balanced confounding is used the estimate of all effects of the same order have the same variance. − See Example 7.3. confound confound confound confound confound • 1: 2: 3: 4: 5: Stat 402 (Spring 2016): slide set 13 We use 100% information to estimate A, B, and C effects, 80% of information to estimate AB, AC and BC effects and 60% of information to estimate the ABC effects. We are using balanced confounding. Rep Rep Rep Rep Rep Another Example: A 23 factorial with blocks of size 4 and use 5 reps. 8 Remarks Estimate of partially confounded effects have a larger variance than the unconfounded effects which imply that we have more information about unconfounded effects as shown above. SV Treatments A B AB C AC BC ABC Blocks Error Total • • • ANOVA Table Stat 402 (Spring 2016): slide set 13 Partial Confounding (Cont’d) Stat 402 (Spring 2016): slide set 13 • 9 We see that all 2nd order interactions are confounded an equal number of times, i.e., once in each rep. Rep 1: confound the AB effect Rep 2: confound the BC effect Rep 3: confound the AC effect. We use a 23 factorial experiment with blocks of size 4 as before and use 3 replicates. Suppose that we use the following design. Definition An incomplete block design for a factorial experiment involves Balanced Confounding if all effects of the same order are confounded with blocks an equal number of times. Remarks and Definition