Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 1 Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 2 Objectives 3 Define blocking effects. Define a randomized complete block design. Generate and analyze a randomized complete block design. Blocking 4 Blocks are groups of experimental units that are formed such that units within blocks are as homogeneous as possible. Blocking is a statistical technique designed to identify and control variation among groups of experimental units. Blocking is a restriction on randomization. Randomized Block Design Model yij i b j ij 5 Drill Tip Experiment 6 Generating a Randomized Complete Block Design This demonstration illustrates the concepts discussed previously. 7 The Design P B G O O B 8 P G O P B G O P B G Is the Blocking Factor Useful? Did the blocking factor help the experiment? 9 Analysis Concerns: Should the Blocking Factor be Deleted from the Model? 10 ... Analysis Concerns: Should the Blocking Factor be Deleted from the Model? NEVER! 11 12 2.01 Multiple Choice Poll Which of the following statements is true? a. Blocking is a restriction on randomization. b. A block effect is not always significant. c. A blocking factor should never be removed from a model. d. All of the above are true. 13 2.01 Multiple Choice Poll – Correct Answer Which of the following statements is true? a. Blocking is a restriction on randomization. b. A block effect is not always significant. c. A blocking factor should never be removed from a model. d. All of the above are true. 14 Analyzing a Randomized Complete Block Design This demonstration illustrates the concepts discussed previously. 15 Prospective versus Retrospective Power Prospective (CRD) 16 Retrospective (CRD) Retrospective (RCBD) Alpha (α) .05 .05 .05 Error Standard Deviation .2 .275 .094 Mean for Orange 9.0 9.450 9.450 Mean for Purple 9.1 9.575 9.575 Mean for Green 9.4 9.600 9.600 Mean for Blue 9.6 9.875 9.875 Sample Size (n) 16 16 16 Power (1-β) .9374 .3357 .9960 17 Exercise This exercise reinforces the concepts discussed previously. 18 19 2.02 Poll In the exercise, the blocking factor Mill had an F statistic of 5.0167. Did the blocking factor help this experiment? Yes No 20 2.02 Poll – Correct Answer In the exercise, the blocking factor Mill had an F statistic of 5.0167. Did the blocking factor help this experiment? Yes No 21 Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 22 Objectives 23 Define an incomplete block. State the requirements for a balanced incomplete block design. Understand the differences between unbalanced and balanced block designs. Generate and analyze an incomplete block design. What Is an Incomplete Block? 24 Properties of a Balanced Incomplete Block Design Each block has the same number of experimental units. Each treatment occurs the same number of times in the experiment. The number of times any two treatments occur together in the same block is the same for all pairs of treatments. If the design is an incomplete block design (that is, each treatment cannot be applied in each block an equal number of times) and any one of these properties is not met, the design is an unbalanced incomplete block design. 25 Balanced Incomplete Block Design BLOCK 26 TREATMENT I 1, 2, 3 II 1, 2, 4 III IV 1, 3, 4 2, 3, 4 Differences between Balanced and Unbalanced Designs 27 A balanced design can require too many runs to be practical. All estimates of treatment means and all comparisons between pairs of treatments have equal precision in a balanced design. A balanced design is more statistically efficient than an unbalanced design. Drill Tip Experiment 28 Generating and Analyzing an Incomplete Block Design This demonstration illustrates the concepts discussed previously. 29 30 Exercise This exercise reinforces the concepts discussed previously. 31 32 2.03 Multiple Choice Poll Which of the following is not a property of the balanced incomplete block design? a. Each block contains the same number of experimental units. b. Each treatment occurs the same number of times in the experiment. c. Each treatment is present in every block. d. Each pair of treatments occurs together in the same block the same number of times. 33 2.03 Multiple Choice Poll – Correct Answer Which of the following is not a property of the balanced incomplete block design? a. Each block contains the same number of experimental units. b. Each treatment occurs the same number of times in the experiment. c. Each treatment is present in every block. d. Each pair of treatments occurs together in the same block the same number of times. 34 Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 35 Objectives 36 Explain the advantages of multiple factor designs. Define common terms. Generate and analyze a full factorial design. Full Factorial 37 Most experiments involve two or more factors. These factors have two or more levels and can be quantitative or qualitative. A design is said to be a full factorial design if all possible combinations of factor levels are included in the experiment. A large number of runs is required for a full factorial design as the number of factors increase. – For example, for a full factorial design with f factors such that factor i has li levels, the number of runs is given by l1*l2*l3…*lf. Advantages of Factorials 38 Factorials reveal interactions. Factorials are more efficient than experiments that use one factor at a time. Combinations of factor levels provide replication for individual factors, when factors are removed from the design. Full Factorial Designs 39 40 2.04 Multiple Choice Poll How many runs would a full factorial experiment with 4 factors, each at 3 levels, require? a. 12 b. 64 c. 81 41 2.04 Multiple Choice Poll – Correct Answer How many runs would a full factorial experiment with 4 factors, each at 3 levels, require? a. 12 b. 64 c. 81 42 Battery Life Experiment Temperature (categorical) 15, 70, 125 Battery Life Type of Plate Material (categorical) 1, 2, 3 43 Generating a Full Factorial Design This demonstration illustrates the concepts discussed previously. 44 45 Exercise This exercise reinforces the concepts discussed previously. 46 47 2.05 Quiz The output from the exercise on fuel use for the aircraft fleet is below. Based on the output, what factors should the company use in future experiments? 48 2.05 Quiz – Correct Answer The output from the exercise on fuel use for the aircraft fleet is below. Based on the output, what factors should the company use in future experiments? Weight and Speed; the experiment shows that only these two factors are significant so future experiments need only include these factors. 49 Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 50 Objectives 51 Understand the application of full factorial designs such that each factor has only two levels. Understand the reasons for using these designs. Generate and analyze a 2k factorial design. Factorial Designs The 2k factorial design is a special case of the full factorial design, introduced to reduce the number of design points is widely used in industrial experimentation forms basic building blocks for other designs assumes that the response is linear in terms of the factors in the design. 52 Two-Level Full Factorial Design 22 = 4 runs 53 23 = 8 runs Coded Levels in a 2k Factorial Design A 2k design is one such that each factor has exactly two levels. These two levels are usually called low and high and are usually denoted by –1 and +1. 54 The 2k Factorial Design with k Factors There will be k main effects and k two-factor 2 interactions k 3 three-factor interactions .. . 1 k-factor interactions For example, a design with 5 factors, or a 25 design, will have 5 main effects, 10 two-factor interactions, 10 three-factor interactions, 5 four-factor interactions, and 1 five-factor interaction. 55 Two-Factor Factorial Experiment 22 The main effect Catalyst = (55+22)/2 – (45+10)/2 = 11 Pressure 1 High 45 1 Low 10 1 Low 56 55 1 High Catalyst Interactions Yes Rate Rate No -1 Catalyst -1 +1 Low Pressure High Pressure 57 Catalyst +1 Completely Randomized Design for Two-Level Full Factorial Design Catalyst Pressure High (-1-1)1 Low High (-1+1)1 (+1-1)1 (+1+1)1 (-1+1)2 (+1-1)2 (+1+1)2 (+1+1)1 (+1+1)2 Low (-1-1)1 (-1+1)1 (+1-1)1 (-1-1)2 (-1+1)2 (-1-1)2 (+1-1)2 In a completely randomized design, the order of the design points is selected at random. 58 The Model 59 2k Factorial with No Replicates 60 These are designs with one observation at each corner of the cube. One unusual value of the response could cause an experimenter to draw incorrect conclusions. The experimenter runs the risk of modeling only noise and not the effects of the factors. These designs are widely used. 2k Factorial with Replication 61 Replication allows for an estimation of pure error and for a test of lack of fit. Center points can be used for replication, if the factors are quantitative. Yield Experiment 62 Temperature (continuous) Pressure (continuous) 225 & 250 14 & 18 Concentration (continuous) Time (continuous) 60 & 80 2.5 & 3 63 2.06 Quiz Match the Yield experiment component on the left with the correct value on the right. 1. Number of factors A. 4 B. 16 2. Number of levels (for each factor) C. 2 3. Number of treatments 64 2.06 Quiz – Correct Answer Match the Yield experiment component on the left with the correct value on the right. 1. Number of factors A. 4 B. 16 2. Number of levels (for each factor) C. 2 3. Number of treatments 1-A, 2-C, 3-B 65 2k Factorial Experiment with No Replication This demonstration illustrates the concepts discussed previously. 66 67 Exercise This exercise reinforces the concepts discussed previously. 68 69 2.07 Poll In the exercise, Time*Pressure was found to be significant but Pressure was not found to be significant. Should Pressure remain in the model? Yes No 70 2.07 Poll – Correct Answer In the exercise, Time*Pressure was found to be significant but Pressure was not found to be significant. Should Pressure remain in the model? Yes No 71 Chapter 2: Multiple Factor Designs and Blocking 2.1 Randomized Complete Block Design 2.2 Randomized Incomplete Block Design 2.3 Full Factorial Designs 2.4 The 2k Factorial Design 2.5 Split-Plot Designs 72 Objectives 73 Define a split-plot design. Define random and fixed effects. Generate and analyze a split-plot design. Carbon Anode Experiment 180 (°C) 2 1 180 (°C) 3 1 190 (°C) 3 1 2 2 1 74 1 2 3 3 2 1 2 2 1 3 2 1 2 3 1 3 3 2 3 1 3 2 1 200 (°C) 2 3 1 Weight 2 210 (°C) 210 (°C) 3 3 190 (°C) 200 (°C) 210 (°C) 2 1 180 (°C) 190 (°C) 200 (°C) 210 (°C) 3 2 190 (°C) 200 (°C) 1 3 180 (°C) 1 3 2 1 Randomized Experiment 75 The runs in every experiment so far have been randomized. – Treatment and sample are randomly selected. – Conditions are reset before every run. Randomization helps meet model assumptions and reduces the chance that changes in an uncontrolled variable will bias the model estimates. Limited Randomization Complete randomization might be difficult or impossible; factor (such as Furnace Temperature) is hard to change. Restricted randomization should be part of the design to maintain an optimal study. Restricted randomization should be part of the analysis so that the model accounts for multiple components of variance. 76 Split-Plot Designs 77 Split-plot designs are used when factors are impractical, inconvenient, or costly to change. This methodology originated in agricultural studies. – The hard-to-change factor was at the same level for an entire plot of land (whole plot factor). – The easy-to-change factor varied levels within a plot of land (sub-plot factor). Randomization is restricted in split-plot designs. These designs have two components of variation: between plots and within a plot. Split-Plot Designs Whole Plots Subplot 78 Subplot Subplot Subplot Whole Plot Effect Restricted randomization is represented by a new variable, Whole Plots. The levels of this factor indicate when hard-to-change factors are randomized. This factor appears as a new term in the regression model. 79 The Model Yijk i j ij k (i ) ijk k i ~ N (0, ) 2 ijk ~ N (0, 2 ) 80 Hard-to-Change Oven Temperature This demonstration illustrates the concepts discussed previously. 81 82 Fixed Effects The levels of interest result from deliberate choice, not sampling a distribution. For example, the levels of Furnace Temperature and Material Type are deliberately chosen and controlled. 83 Inferences are to be made only to those levels included in the study. In this case, you are interested in the mean response at a given factor level. Random Effects 84 Levels represent a sample, for example, different runs in the furnace Inferences apply to the population of levels, not just the subset of levels included in the study. In this example, you are interested in the variation across oven runs in general, not just the particular set of runs in the study. Mixed Models Models in which some factors are treated as fixed effects and other factors are treated as random effects are called mixed effects models, or simply mixed models. – Whole plot variation is the random effect. – Furnace Temperature and Material Type are the fixed effects. 85 The Whole Plots term in the JMP model specification includes the Random attribute. Restricted Maximum Likelihood (REML) is the recommended method to estimate the variance components and parameters. 86 2.08 Quiz Match the term on the left with the appropriate description on the right. A. receives the easy-to-change factor 1. Fixed effect 2. Random effect B. extends inferences to all population levels 3. Whole plot C. receives the hard-to-change factor 4. Subplot D. extends inferences only to the levels in the experiment 87 2.08 Quiz – Correct Answer Match the term on the left with the appropriate description on the right. A. receives the easy-to-change factor 1. Fixed effect 2. Random effect B. extends inferences to all population levels 3. Whole plot C. receives the hard-to-change factor 4. Subplot D. extends inferences only to the levels in the experiment 1-D, 2-B, 3-C, 4-A 88 Split-Plot Analysis This demonstration illustrates the concepts discussed previously. 89 90 Exercise This exercise reinforces the concepts discussed previously. 91 92 2.09 Quiz The output from the split-plot exercise is below. Based on the output, which is the largest component of variance? 93 2.09 Quiz – Correct Answer The output from the split-plot exercise is below. Based on the output, which is the largest component of variance? The larger component of variance is the Whole Plot component. Therefore more of the error was accounted for by the differences between the whole plots rather than from random noise that was not accounted for by the whole plots. 94