Chapter 6 The 2 k Factorial Design Solutions


 Brendan Clark
 3 years ago
 Views:
Transcription
1 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Chapter 6 The k Factorial Design Solutions 6.. A router is used to cut locating notches on a printed circuit board. The vibration level at the surface of the board as it is cut is considered to be a major source of dimensional variation in the notches. Two factors are thought to influence vibration: bit size (A) and cutting speed (B). Two bit sizes (/6 and /8 inch) and two speeds (40 and 90 rpm) are selected, and four boards are cut at each set of conditions shown below. The response variable is vibration measured as a resultant vector of three accelerometers (x, y, and z) on each test circuit board. Treatment Replicate A B Combination I II III IV   () a b ab (a) Analyze the data from this experiment. Design Expert Output Response Vibration ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ABit Size < BCutting Speed < AB < Pure Error Cor Total The Model Fvalue of 9.0 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. 6
2 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (b) Construct a normal probability plot of the residuals, and plot the residuals versus the predicted vibration level. Interpret these plots. Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted There is nothing unusual about the residual plots. (c) Draw the AB interaction plot. Interpret this plot. What levels of bit size and speed would you recommend for routine operation? To reduce the vibration, use the smaller bit. Once the small bit is specified, either speed will work equally well, because the slope of the curve relating vibration to speed for the small tip is approximately zero. The process is robust to speed changes if the small bit is used. DesignExpert Software Vibration Design Points B B Interaction B: Cutting Speed X = A: Bit Size X = B: Cutting Speed Vibration A: Bit Size 6.. An engineer is interested in the effects of cutting speed (A), tool geometry (B), and cutting angle on the life (in hours) of a machine tool. Two levels of each factor are chosen, and three replicates of a 3 factorial design are run. The results are as follows: Treatment Replicate 6
3 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY A B C Combination I II III () a b ab c ac bc abc (a) Estimate the factor effects. Which effects appear to be large? From the normal probability plot of effects below, factors B, C, and the AC interaction appear to be significant. DESIGNEXPERT Plot Life Normal plot A: Cutting Speed B: Tool Geometry C: Cutting Angle Normal % probability A C B AC Effect (b) Use the analysis of variance to confirm your conclusions for part (a). The analysis of variance confirms the significance of factors B, C, and the AC interaction. Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C AB AC BC ABC Pure Error Cor Total The Model Fvalue of 7.64 implies the model is significant. There is only 63
4 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY a 0.04% chance that a "Model FValue" this large could occur due to noise. The reduced model ANOVA is shown below. Factor A was included to maintain hierarchy. Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B < C AC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.54 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Effects B, C and AC are significant at %. (c) Write down a regression model for predicting tool life (in hours) based on the results of this experiment. y x x x x ijk A B C A x C Design Expert Output Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ACutting Speed BTool Geometry CCutting Angle AC Final Equation in Terms of Coded Factors: Life = * A * B +3.4 * C 4.4 * A * C Final Equation in Terms of Actual Factors: Life = * Cutting Speed * Tool Geometry * Cutting Angle * Cutting Speed * Cutting Angle The equation in part (c) and in the given in the computer output form a hierarchial model, that is, if an interaction is included in the model, then all of the main effects referenced in the interaction are also included in the model. (d) Analyze the residuals. Are there any obvious problems? 64
5 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted.5 99 Normal % probability Residuals Residual Predicted There is nothing unusual about the residual plots. (e) Based on the analysis of main effects and interaction plots, what levels of A, B, and C would you recommend using? Since B has a positive effect, set B at the high level to increase life. The AC interaction plot reveals that life would be maximized with C at the high level and A at the low level. DESIGNEXPERT Plot Life 60 Interaction Graph Cutting Angle DESIGNEXPERT Plot Life 60 One Factor Plot X = A: Cutting Speed Y = C: Cutting Angle 50.5 C C+.000 Actual Factor B: Tool Geometry = Life X = B: Tool Geometry Actual Factors A: Cutting Speed = 0.00 C: Cutting Angle = 0.00 Life Cutting Speed Tool Geometry 6.3. Reconsider part (c) of Problem 6.. Use the regression model to generate response surface and contour plots of the tool life response. Interpret these plots. Do they provide insight regarding the desirable operating conditions for this process? The response surface plot and the contour plot in terms of factors A and C with B at the high level are shown below. They show the curvature due to the AC interaction. These plots make it easy to see the region of greatest tool life. 65
6 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Life C: Cutting Angle Life A: Cutting Speed C: Cutting Angle A: Cutting Speed 6.4. Plot the factor effects from Problem 6. on a graph relative to an appropriately scaled t distribution. Does this graphical display adequately identify the important factors? Compare the conclusions from this MSE plot with the results from the analysis of variance. S 3. 7 n 3 Scaled t Distribution A C C B Factor Effects This method identifies the same factors as the analysis of variance Find the standard error of the factor effects and approximate 95 percent confidence limits for the factor effects in Problem 6.. Do the results of this analysis agree with the conclusions from the analysis of variance? SE( effect ) S k 3 n
7 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Variable Effect A B.333 * AB C * AC * BC ABC .67 The 95% confidence intervals for factors B, C and AC do not contain zero. This agrees with the analysis of variance approach Reconsider the experiment described in Problem 6.. Suppose that the experimenter only performed the eight trials from replicate I. In addition, he ran four center points and obtained the following response values: 36, 40, 43, 45. (a) Estimate the factor effects. Which effects are large? DESIGNEXPERT Plot Life Normal plot A: Cutting Speed B: Tool Geometry C: Cutting Angle Normal % probability AC C B Effect Effects B, C, and AC appear to be large. (b) Perform an analysis of variance, including a check for pure quadratic curvature. What are your conclusions? SS PureQuadratic n F n C n y y F F n C C Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C AB AC
8 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY BC ABC Curvature E not significant Pure Error Cor Total The Model Fvalue of 9.77 implies the model is significant. There is only a 4.39% chance that a "Model FValue" this large could occur due to noise. The "Curvature Fvalue" of 0.00 implies the curvature (as measured by difference between the average of the center points and the average of the factorial points) in the design space is not significant relative to the noise. There is a 96.7% chance that a "Curvature Fvalue" this large could occur due to noise. Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C AC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.9 implies the model is significant. There is only a 0.98% chance that a "Model FValue" this large could occur due to noise. Effects B, C and AC are significant at 5%. There is no effect of curvature. (c) Write down an appropriate model for predicting tool life, based on the results of this experiment. Does this model differ in any substantial way from the model in Problem 6., part (c)? The model shown in the Design Expert output below does not differ substantially from the model in Problem 6., part (c). Design Expert Output Final Equation in Terms of Coded Factors: Life = * A * B * C * A * C (d) Analyze the residuals. 68
9 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted (e) What conclusions would you draw about the appropriate operating conditions for this process? To maximize life run with B at the high level, A at the low level and C at the high level Cube Graph Life B B: Tool Geometry C+ C: Cutting Angle B C A A+ A: Cutting Speed 6.7. A bacteriologist is interested in the effects of two different culture media and two different times on the growth of a particular virus. She performs six replicates of a design, making the runs in random order. Analyze the bacterial growth data that follow and draw appropriate conclusions. Analyze the residuals and comment on the model s adequacy. 69
10 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Time Culture Medium di 5 6 hr hr Design Expert Output Response: Virus growth ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B < AB Residual Lack of Fit Pure Error Cor Total The Model Fvalue of 45. implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, AB are significant model terms. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Growth rate is affected by factor B (Time) and the AB interaction (Culture medium and Time). There is some very slight indication of inequality of variance shown by the small decreasing funnel shape in the plot of residuals versus predicted. 60
11 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Interaction Graph Virus growth 39 Tim e X = A: Culture Medium Y = B: Time Design Points 34.5 B.000 B Virus growth Culture Medium 6.8. An industrial engineer employed by a beverage bottler is interested in the effects of two different types of 3ounce bottles on the time to deliver bottle cases of the product. The two bottle types are glass and plastic. Two workers are used to perform a task consisting of moving 40 cases of the product 50 feet on a standard type of hand truck and stacking the cases in a display. Four replicates of a factorial design are performed, and the times observed are listed in the following table. Analyze the data and draw the appropriate conclusions. Analyze the residuals and comment on the model s adequacy. Worker Bottle Type Glass Plastic Design Expert Output ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant AWorker BBottle Type AB Pure Error Cor Total The Model Fvalue of 3.6 implies the model is significant. There is only a 0.04% chance that a "Model FValue" this large could occur due to noise. There is some indication of nonconstant variance in this experiment. 6
12 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. Worker Residuals vs. Bottle Type Residuals Residuals Worker B:Bottle Type 6.9. In problem 6.8, the engineer was also interested in potential fatigue differences resulting from the two types of bottles. As a measure of the amount of effort required, he measured the elevation of heart rate (pulse) induced by the task. The results follow. Analyze the data and draw conclusions. Analyze the residuals and comment on the model s adequacy. Worker Bottle Type Glass Plastic Design Expert Output Response Fatigue 6
13 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant AWorker < BBottle Type AB Pure Error Cor Total The Model Fvalue of 6.8 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A are significant model terms. Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. Worker Residuals vs. Bottle Type Residuals 0 Residuals Worker B:Bottle Type 63
14 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY There is an indication that one worker exhibits greater variability than the other Calculate approximate 95 percent confidence limits for the factor effects in Problem 6.9. Do the results of this analysis agree with the analysis of variance performed in Problem 6.9? SE( effect ) S k n 4 Variable Effect C.I. A (.96)= 7.3 B (.96)= 7.3 AB (.96)= 7.3 The 95% confidence interval for factor A does not contain zero. This agrees with the analysis of variance approach. 6.. An experiment was performed to improve the yield of a chemical process. Four factors were selected, and two replicates of a completely randomized experiment were run. The results are shown in the following table: Treatment Replicate Replicate Treatment Replicate Replicate Combination I II Combination I II () d a ad 7 76 b 8 85 bd ab abd c cd ac 8 80 acd bc 88 8 bcd abc abcd (a) Estimate the factor effects. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B Error C Error D Error AB Error AC Error AD Error BC Error BD Error CD Error ABC Error ABD Error ACD Error BCD Error ABCD
15 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (b) Prepare an analysis of variance table, and determine which factors are important in explaining yield. Design Expert Output Response: yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B C D AB AC AD BC BD CD ABC < ABD ACD BCD ABCD Residual Lack of Fit Pure Error Cor Total The Model Fvalue of 3.0 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AB, AD, ABC, ABD, ABCD are significant model terms. F 006.,, 853., and F 0056.,, 6. therefore, factors A and D and interactions AB, ABC, and ABD are significant at %. Factor C and interactions AD and ABCD are significant at 5%. (b) Write down a regression model for predicting yield, assuming that all four factors were varied over the range from  to + (in coded units). Model with hierarchy maintained: Design Expert Output Final Equation in Terms of Coded Factors: yield = * A * B .34 * C +.97 * D +.03 * A * B * A * C .09 * A * D 0.8 * B * C * B * D * C * D .59 * A * B * C +.34 * A * B * D * A * C * D * B * C * D +. * A * B * C * D 65
16 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Model without hierarchy terms: Design Expert Output Final Equation in Terms of Coded Factors: yield = * A .34 * C +.97 * D +.03 * A * B .09 * A * D .59 * A * B * C +.34 * A * B * D +. * A * B * C * D Confirmation runs might be run to see if the simpler model without hierarchy is satisfactory. (d) Plot the residuals versus the predicted yield and on a normal probability scale. Does the residual analysis appear satisfactory? There appears to be one large residual both in the normal probability plot and in the plot of residuals versus predicted. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted (e) Two threefactor interactions, ABC and ABD, apparently have large effects. Draw a cube plot in the factors A, B, and C with the average yields shown at each corner. Repeat using the factors A, B, and D. Do these two plots aid in data interpretation? Where would you recommend that the process be run with respect to the four variables? 66
17 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Cube Graph yield Cube Graph yield B B B: B C+ B: B D+ C: C D: D B C A A+ A: A B D A A+ A: A Run the process at A low B low, C low and D high. 6.. A nickeltitanium alloy is used to make components for jet turbine aircraft engines. Cracking is a potentially serious problem in the final part, as it can lead to nonrecoverable failure. A test is run at the parts producer to determine the effects of four factors on cracks. The four factors are pouring temperature (A), titanium content (B), heat treatment method (C), and the amount of grain refiner used (D). Two replicated of a 4 design are run, and the length of crack (in m) induced in a sample coupon subjected to a standard test is measured. The data are shown below: Treatment Replicate Replicate A B C D Combination I II () a b ab c ac bc abc d ad bd abd cd acd bcd abcd (a) Estimate the factor effects. Which factors appear to be large? 67
18 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY From the half normal plot of effects shown below, factors A, B, C, D, AB, AC, and ABC appear to be large. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Model AB Model AC Error AD Error BC Error BD Error CD Model ABC Error ABD Error ACD Error BCD Error ABCD DESIGNEXPERT Plot Crack Length Half Normal plot A: Pour Temp B: Titanium Content C: Heat Treat Method D: Grain Ref iner AC 95 B Half Normal %probability BC D AB C ABC A Effect (b) Conduct an analysis of variance. Do any of the factors affect cracking? Use =0.05. The Design Expert output below identifies factors A, B, C, D, AB, AC, and ABC as significant. Design Expert Output Response: Crack Lengthin mm x 0^ ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B < C < D < AB < AC < AD BC BD CD
19 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY ABC < ABD ACD.96E E BCD ABCD.596E E Residual Lack of Fit Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, D, AB, AC, ABC are significant model terms. (c) Write down a regression model that can be used to predict crack length as a function of the significant main effects and interactions you have identified in part (b). Design Expert Output Final Equation in Terms of Coded Factors: Crack Length= *A +.99 *B .80 *C *D *A*B .00 *A*C +.57 * A * B * C (d) Analyze the residuals from this experiment. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted There is nothing unusual about the residuals. (e) Is there an indication that any of the factors affect the variability in cracking? By calculating the range of the two readings in each cell, we can also evaluate the effects of the factors on variation. The following is the normal probability plot of effects: 69
20 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Range Normal plot A: Pour Temp B: Titanium Content C: Heat Treat Method D: Grain Refiner Normal % probability AB CD Effect It appears that the AB and CD interactions could be significant. The following is the ANOVA for the range data: Design Expert Output Response: Range ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant AB CD Residual Cor Total The Model Fvalue of.46 implies the model is significant. There is only a 0.4% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case AB, CD are significant model terms. Final Equation in Terms of Coded Factors: Range = * A * B +0.0 * C * D (f) What recommendations would you make regarding process operations? Use interaction and/or main effect plots to assist in drawing conclusions. From the interaction plots, choose A at the high level and B at the low level. In each of these plots, D can be at either level. From the main effects plot of C, choose C at the high level. Based on the range analysis, with C at the high level, D should be set at the low level. From the analysis of the crack length data: 60
21 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Crack Length 9.84 Interaction Graph B: Titanium Content DESIGNEXPERT Plot Crack Length 9.84 Interaction Graph C: Heat Treat Method X = A: Pour Temp Y = B: Titanium Content X = A: Pour Temp Y = C: Heat Treat Method B B+.000 Actual Factors C: Heat Treat Method = D: Grain Refiner = Crack Length C  C Actual Factors B: Titanium Content = 0.00 D: Grain Refiner = Crack Length A: Pour Tem p A: Pour Tem p DESIGNEXPERT Plot Crack Length 9.84 X = D: Grain Refiner One Factor Plot DESIGNEXPERT Plot Crack Length X = A: Pour Temp Y = B: Titanium Content Z = C: Heat Treat Method Cube Graph Crack Length Actual Factors A: Pour Temp = 0.00 B: Titanium Content = 0.00 C: Heat Treat Method = Crack Length Actual Factor D: Grain Refiner = 0.00 B: Titanium Content B C+ C: Heat Treat Metho B C A A+ A: Pour Tem p D: Grain Refiner From the analysis of the ranges: 6
22 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Range 0.66 Interaction Graph B: Titanium Content DESIGNEXPERT Plot Range 0.66 Interaction Graph D: Grain Refiner X = A: Pour Temp Y = B: Titanium Content X = C: Heat Treat Method Y = D: Grain Refiner 0.55 B B+.000 Actual Factors C: Heat Treat Method = 0.00 D: Grain Refiner = Range 0.55 D D+.000 Actual Factors A: Pour Temp = 0.00 B: Titanium Content = Range A: Pour Tem p C: Heat Treat Method 6.3. Continuation of Problem 6.. One of the variables in the experiment described in Problem 6., heat treatment method (c), is a categorical variable. Assume that the remaining factors are continuous. (a) Write two regression models for predicting crack length, one for each level of the heat treatment method variable. What differences, if any, do you notice in these two equations? Design Expert Output Final Equation in Terms of Actual Factors Heat Treat Method  Crack Length = * Pour Temp * Titanium Content * Grain Refiner * Pour Temp * Titanium Content Heat Treat Method Crack Length = * Pour Temp * Titanium Content * Grain Refiner * Pour Temp * Titanium Content (b) Generate appropriate response surface contour plots for the two regression models in part (a). 6
23 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot.00 Crack Length X = A: Pour Temp Y = B: Titanium Content Crack Length 8 DESIGNEXPERT Plot.00 Crack Length X = A: Pour Temp Y = B: Titanium Content Crack Length Actual Factors C: Heat Treat Method = D: Grain Refiner = 0.00 B: Titanium Content Actual Factors C: Heat Treat Method = 0.50 D: Grain Refiner = 0.00 B: Titanium Content A: Pour Tem p A: Pour Tem p (c) What set of conditions would you recommend for the factors A, B and D if you use heat treatment method C=+? High level of A, low level of B, and low level of D. (d) Repeat part (c) assuming that you wish to use heat treatment method C=. Low level of A, low level of B, and low level of D An article in the AT&T Technical Journal (March/April 986, Vol. 65, pp ) describes the application of twolevel factorial designs to integrated circuit manufacturing. A basic processing step is to grow an epitaxial layer on polished silicon wafers. The wafers mounted on a susceptor are positioned inside a bell jar, and chemical vapors are introduced. The susceptor is rotated and heat is applied until the epitaxial layer is thick enough. An experiment was run using two factors: arsenic flow rate (A) and deposition time (B). Four replicates were run, and the epitaxial layer thickness was measured (in mm). The data are shown below: Replicate Factor Levels A B I II III IV Low () High (+) A 55% 59% B Short Long (0 min) (5 min) (a) Estimate the factor effects. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B Error AB Error Lack Of Fit 0 0 Error Pure Error
24 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (b) Conduct an analysis of variance. Which factors are important? From the analysis of variance shown below, no factors appear to be important. Factor B is only marginally interesting with an Fvalue of 4.3. Design Expert Output Response: Thickness ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B AB Residual Lack of Fit Pure Error Cor Total The "Model Fvalue" of.9 implies the model is not significant relative to the noise. There is a 4.5 % chance that a "Model Fvalue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case there are no significant model terms. (c) Write down a regression equation that could be used to predict epitaxial layer thickness over the region of arsenic flow rate and deposition time used in this experiment. Design Expert Output Final Equation in Terms of Coded Factors: Thickness = * A +0.9 * B +0.4 * A * B Final Equation in Terms of Actual Factors: Thickness = * Flow Rate * Dep Time * Flow Rate * Dep Time (d) Analyze the residuals. Are there any residuals that should cause concern? Observation # falls outside the groupings in the normal probability plot and the plot of residual versus predicted. 64
25 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted (e) Discuss how you might deal with the potential outlier found in part (d). One approach would be to replace the observation with the average of the observations from that experimental cell. Another approach would be to identify if there was a recording issue in the original data. The first analysis below replaces the data point with the average of the other three. The second analysis assumes that the reading was incorrectly recorded and should have been The analysis with the run associated with standard order replaced with the average of the remaining three runs in the cell, 3.97, is shown below. Design Expert Output Response: Thickness ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A 7.439E E B < AB.76E E Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B are significant model terms. Final Equation in Terms of Coded Factors: Thickness = * A * B E003 * A * B Final Equation in Terms of Actual Factors: Thickness = * Flow Rate 65
26 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY * Dep Time E004 * Flow Rate * Dep Time Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted A new outlier is present and should be investigated. Analysis with the run associated with standard order replaced with the value 4.65: Design Expert Output Response: Thickness ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B < AB 3.969E E Pure Error Cor Total The Model Fvalue of 45.8 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B are significant model terms. Final Equation in Terms of Coded Factors: Thickness = * A +0.4 * B * A * B Final Equation in Terms of Actual Factors: Thickness = * Flow Rate * Dep Time E003 * Flow Rate * Dep Time 66
27 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Studentized Residuals Studentized Residuals Predicted Another outlier is present and should be investigated Continuation of Problem 6.4. Use the regression model in part (c) of Problem 6.4 to generate a response surface contour plot for epitaxial layer thickness. Suppose it is critically important to obtain layer thickness of 4.5 mm. What settings of arsenic flow rate and deposition time would you recommend? Arsenic flow rate may be set at any of the experimental levels, while the deposition time should be set at.4 minutes. DESIGNEXPERT Plot Thickness X = A: Flow Rate Y = B: Dep Time Design Points 3.75 Thickness DESIGNEXPERT Plot Thickness X = A: Flow Rate Y = B: Dep Time Design Points Interaction Graph Dep Time Dep Time B B Thickness Flow Rate Flow Rate 6.6. Continuation of Problem 6.5. How would your answer to Problem 6.5 change if arsenic flow rate was more difficult to control in the process than the deposition time? Running the process at a high level of Deposition Time there is no change in thickness as flow rate changes. 67
28 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY 6.7. Semiconductor manufacturing processes have long and complex assembly flows, so matrix marks and automated dmatrix readers are used at several process steps throughout factories. Unreadable matrix marks negatively effect factory run rates, because manual entry of part data is required before manufacturing can resume. A 4 factorial experiment was conducted to develop a dmatrix laser mark on a metal cover that protects a substrate mounted die. The design factors are A = laser power (9W, 3W), B = laser pulse frequency (4000 Hz, 000 Hz), C = matrix cell size (0.07 in, 0. in), and D = writing speed (0 in/sec, 0 in/sec), and the response variable is the unused error correction (UEC). This is a measure of the unused portion of the redundant information embedded in the d matrix. A UEC of 0 represents the lowest reading that still results in a decodable matrix while a value of is the highest reading. A DMX Verifier was used to measure UEC. The data from this experiment are shown below. Standard Order Run Order Laser Power Pulse Frequency Cell Size Writing Speed UEC (a) Analyze the data from this experiment. Which factors significantly affect UEC? The normal probability plot of effects identifies A, C, D, and the AC interaction as significant. The Design Expert output including the analysis of variance confirms the significance and identifies the corresponding model. Contour plots identify factors A and C with B held constant at zero and D toggled from  to
29 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot UEC Normal plot A: Laser Power B: Pulse Frequency C: Cell Size D: Writing Speed A Normal % probability C D AC Design Expert Output Response: UEC ANOVA for Selected Factorial Model Analysis of variance table [Terms added sequentially (first to last)] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < C < D AC Residual E003 Cor Total The Model Fvalue of 35.5 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC are significant model terms. Std. Dev RSquared 0.98 Mean 0.7 Adj RSquared C.V Pred RSquared PRESS Adeq Precision Final Equation in Terms of Coded Factors UEC = * A * C * D * A * C Final Equation in Terms of Actual Factors UEC = * Laser Power * Cell Size * Writing Speed * Laser Power * Cell Size Effect 69
30 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot UEC X = A: Laser Power Y = C: Cell Size.00 UEC DESIGNEXPERT Plot UEC X = A: Laser Power Y = C: Cell Size UEC Actual Factors 0.50 B: Pulse Frequency = 0.00 D: Writing Speed = Actual Factors 0.50 B: Pulse Frequency = 0.00 D: Writing Speed = C: Cell Size C: Cell Size A: Laser Power A: Laser Power (b) Analyze the residuals from this experiment. Are there any indications of model inadequacy? The residual plots appear acceptable with the exception of run 8, standard order 6. This value should be verified by the engineer. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 630
31 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Run Residuals vs. Laser Power Residuals Residuals Run Number Laser Power Residuals vs. Pulse Frequency Residuals vs. Cell Size Residuals Residuals Pulse Frequency Cell Size Residuals vs. Writing Speed Residuals Writing Speed 63
32 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY 6.8. Consider a variation of the bottle filling experiment from Example 5.3. Suppose that only two levels of carbonation are used so that the experiment is a 3 factorial design with two replicates. The data are shown below. Coded Factors Fill Height Deviation Run A B C Replicate Replicate Factor Levels Low () High (+) A (%) 0 B (psi) 5 30 C (b/m) (a) Analyze the data from this experiment. Which factors significantly affect fill height deviation? The half normal probability plot of effects shown below identifies the factors A, B, and C as being significant and the AB interaction as being marginally significant. The analysis of variance in the Design Expert output below confirms that factors A, B, and C are significant and the AB interaction is marginally significant. DESIGNEXPERT Plot Fill Deviation Half Normal plot A: Carbonation B: Pressure C: Speed 99 Half Normal % probability AB C B A Effect Design Expert Output Response: Fill Deviation ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant 63
33 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY A < B C AB Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 6.84 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C are significant model terms. Std. Dev. 0.8 RSquared Mean.00 Adj RSquared C.V. 8.8 Pred RSquared PRESS 5.34 Adeq Precision 5.44 (b) Analyze the residuals from this experiment. Are there any indications of model inadequacy? The residual plots below do not identify any violations to the assumptions. Normal plot of residuals Residuals vs. Predicted.5 99 Normal % probability Residuals Residual Predicted 633
34 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Run Residuals vs. Carbonation Residuals 0 Residuals Run Number Carbonation Residuals vs. Pressure Residuals vs. Speed Residuals 0 3 Residuals Pressure Speed (c) Obtain a model for predicting fill height deviation in terms of the important process variables. Use this model to construct contour plots to assist in interpreting the results of the experiment. The model in both coded and actual factors are shown below. Design Expert Output Final Equation in Terms of Coded Factors Fill Deviation = * A +.3 * B * C * A * B Final Equation in Terms of Actual Factors Fill Deviation = * Carbonation * Pressure 634
35 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY * Speed * Carbonation * Pressure The following contour plots identify the fill deviation with respect to carbonation and pressure. The plot on the left sets the speed at 00 b/m while the plot on the right sets the speed at 50 b/m. Assuming a faster bottle speed is better, settings in pressure and carbonation that produce a fill deviation near zero can be found in the lower left hand corner of the contour plot on the right. B: Pressure Fill Deviation Fill Deviation B: Pressure A: Carbonation Speed set at 00 b/m A: Carbonation Speed set at 50 b/m (d) In part (a), you probably noticed that there was an interaction term that was borderline significant. If you did not include the interaction term in your model, include it now and repeat the analysis. What difference did this make? If you elected to include the interaction term in part (a), remove it and repeat the analysis. What difference does this make? The following analysis of variance, residual plots, and contour plots represent the model without the interaction. As in the original analysis, the residual plots do not identify any concerns with the assumptions. The contour plots did not change significantly either. The interaction effect is small relative to the main effects. Design Expert Output Response: Fill Deviation ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 8.84 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C are significant model terms. Std. Dev RSquared
36 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Mean.00 Adj RSquared C.V Pred RSquared PRESS 6.89 Adeq Precision Final Equation in Terms of Coded Factors Fill Deviation = * A +.3 * B * C Final Equation in Terms of Actual Factors Fill Deviation = * Carbonation * Pressure * Speed Normal plot of residuals Residuals vs. Predicted.5 99 Normal % probability Residuals Residual Predicted Residuals vs. Run Residuals vs. Carbonation Residuals 0.5 Residuals Run Number Carbonation 636
37 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Pressure Residuals vs. Speed Residuals 0.5 Residuals Pressure Speed Fill Deviation Fill Deviation B: Pressure B: Pressure A: Carbonation Speed set at 00 b/m A: Carbonation Speed set at 50 b/m 6.9. An experimenter has run a single replicate of a 4 design. The following effect estimates have been calculated: A = AB = 5.3 ABC = .8 B = AC =.69 ABD = C = AD = 9.78 ACD = 0.0 D = BC = 0.78 BCD = BD = 4.74 ABCD = 6.5 CD =.7 (a) Construct a normal probability plot of these effects. The plot from Minitab follows. 637
38 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Probability Plot A Probability B AB.00 Average: .57 StDev: N: Value AndersonDarling Normality Test ASquared: 0.78 PValue: (b) Identify a tentative model, based on the plot of the effects in part (a). ŷ Intercept x x 5. 66x A B A x B 6.0. I am always interested in improving my golf scores. Since a typical golfer uses the putter for about 3545% of his or her strokes, it seems logical that in improving one s putting score is a logical and perhaps simple way to improve a golf score ( The man who can putt is a match for any man. Willie Parks, , twotime winner of the British Open). An experiment was conducted to study the effects of four factors on putting accuracy. The design factors are length of putt, type of putter, breaking putt vs. straight putt, and level versus downhill putt. The response variable is distance from the ball to the center of the cup after the ball comes to rest. One golfer performs the experiment, a 4 factorial design with seven replicates was used, and all putts were made in random order. The results are as follows. Design Factors Distance from cup (replicates) Length of putt (ft) Type of putter Break of putt Slope of putt Mallet Straight Level Mallet Straight Level Cavityback Straight Level Cavityback Straight Level Mallet Breaking Level Mallet Breaking Level Cavityback Breaking Level Cavityback Breaking Level Mallet Straight Downhill Mallet Straight Downhill Cavityback Straight Downhill Cavityback Straight Downhill Mallet Breaking Downhill Mallet Breaking Downhill Cavityback Breaking Downhill Cavityback Breaking Downhill
39 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Analyze the data from this experiment. Which factors significantly affect putting performance? The half normal probability plot of effects identifies only factors A and B, length of putt and type of putter, as having a potentially significant affect on putting performance. The analysis of variance with only these significant factors is presented as well and confirms significance. DESIGNEXPERT Plot Distance f rom cup Half Normal plot A: Length of putt B: Ty pe of putter C: Break of putt D: Slope of putt 99 A Half Normal %probability B Effect Design Expert Output with Only Factors A and B Response: Distance from cup ANOVA for Selected Factorial Model Analysis of variance table [Terms added sequentially (first to last)] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.69 implies the model is significant. There is only a 0.08% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, are significant model terms. Std. Dev. 9. RSquared 0.37 Mean.30 Adj RSquared C.V Pred RSquared PRESS Adeq Precision 6.66 (b) Analyze the residuals from this experiment. Are there any indications of model inadequacy? The residual plots for the model containing only the significant factors A and B are shown below. The normality assumption appears to be violated. Also, as a golfer might expect, there is a slight inequality of variance with regards to the length of putt. A square root transformation is applied which corrects the violations. The analysis of variance and corrected residual plots are also presented. Finally, an effects plot identifies a 0 foot putt and the cavityback putter reduce the mean distance from the cup. 639
40 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. Length of putt Residuals vs. Type of putter Residuals Residuals Length of putt Type of putter Residuals vs. Break of putt Residuals vs. Slope of putt Residuals Residuals Break of putt Slope of putt 640
41 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Design Expert Output with Only Factors A and B and a Square Rot Transformation Response: Distance from cuptransform:square root Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Terms added sequentially (first to last)] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.85 implies the model is significant. There is only a 0.07% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, are significant model terms. Std. Dev..54 RSquared 0.59 Mean 3. Adj RSquared C.V Pred RSquared PRESS Adeq Precision Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted 64
42 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Length of putt Residuals vs. Type of putter Residuals Residuals Length of putt Type of putter Residuals vs. Break of putt Residuals vs. Slope of putt Residuals Residuals Break of putt Slope of putt DESIGNEXPERT Plot Sqrt(Distance f rom cup) 0 X = A: Length of putt Y = B: Ty pe of putter 6 B Mallet B Cavityback Actual Factors C: Break of putt = Straight D: Slope of putt = Lev el Distance from cup 8 Interaction Graph B: Type of putter A: Length of putt 64
43 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY 6.. A company markets its products by direct mail. An experiment was conducted to study the effects of three factors on the customer response rate for a particular product. The three factors are A = type of mail used (3 rd class, st class), B = type of descriptive brochure (color, blackandwhite), and C = offer price ($9.95, $4.95). The mailings are made to two groups of 8,000 randomly selected customers, with,000 customers in each group receiving each treatment combination. Each group of customers is considered as a replicate. The response variable is the number of orders placed. The experimental data is shown below. Coded Factors Number of Orders Run A B C Replicate Replicate Factor Levels Low () High (+) A (class) 3rd st B (type) BW Color C ($) $9.95 $4.95 (a) Analyze the data from this experiment. Which factors significantly affect the customer response rate? The half normal probability plot of effects identifies the two factor interactions, AB, AC, BC, and factors A and C as significant. Factor B is not significant; however, remains in the model to satisfy the hierarchal principle. The analysis of variance confirms the significance of two factor interactions and factor C. Factor A is marginally significant. HalfNormal Plot 99 HalfNormal % Probability A ABC AB C BC AC Standardized Effect 643
44 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response Number of Orders ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant AClass BType C$ AB AC BC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.0 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case C, AB, AC, BC are significant model terms. (b) Analyze the residuals from this experiment. Are there any indications of model inadequacy? The residual plots below do not identify model inadequacy. Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted 644
45 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Class Residuals vs. Type Residuals 0.5 Residuals Class B:Type Residuals vs. $ Residuals C:$ (c) What would you recommend to the company? Based on the interaction plots below, we recommend st class mail, color brochures, and an offered price of $4.95 would achieve the greatest number of orders. 645
46 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Number of Orders Design Points B BW B Color 56 5 Interaction B: Type X = A: Class X = B: Type Actual Factor C: $ = $9.95 Number of Orders rd st A: Class DesignExpert Software Number of Orders Design Points B BW B Color Interaction B: Type X = A: Class X = B: Type Actual Factor C: $ = $4.95 Number of Orders rd st A: Class 6.. Reconsider the experiment in Problem 6.7. Suppose that four center points are available, and the UEC response at these four runs is 0.98, 0.95, 0.93 and 0.96, respectively. Reanalyze the experiment incorporating a test for curvature into the analysis. What conclusions can you draw? What recommendations would you make to the experimenters? As with the results of Problem 6.7, factors A, C, D, and the AC interaction remain significant. However, the CD interaction and curvature are significant as well. The curvature is the strongest effect; unfortunately, we are not able to determine which factor(s) have a quadratic term. We recommend that the engineer augment the experiment with additional experimental runs, such as axial points and a couple of 646
47 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY extra center points for blocking purposes. These extra runs will determine the pure quadratic effects and allow us to fit a second order model. Design Expert Output Response: UEC ANOVA for Selected Factorial Model Analysis of variance table [Terms added sequentially (first to last)] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < C < D < AC CD 5.65E E Curvature < significant Residual E003 Lack of Fit E not significant Pure Error.300E E004 Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC, CD are significant model terms. Std. Dev RSquared Mean 0.76 Adj RSquared C.V. 4.8 Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept E ALaser Power E CCell Size E DWriting Speed E AC E E CD E E Center Point An experiment was run in a semiconductor fabrication plant in an effort to increase yield. Five factors, each at two levels, were studied. The factors (and levels) were A = aperture setting (small, large), B = exposure time (0% below nominal, 0% above nominal), C = development time (30 s, 45 s), D = mask dimension (small, large), and E = etch time (4.5 min, 5.5 min). The unreplicated 5 design shown below was run. () = 7 d = 8 e = 8 de = 6 a = 9 ad = 0 ae = ade = 0 b = 34 bd = 3 be = 35 bde = 30 ab = 55 abd = 50 abe = 5 abde = 53 c = 6 cd = 8 ce = 5 cde = 5 ac = 0 acd = ace = acde = 0 bc = 40 bcd = 44 bce = 45 bcde = 4 abc = 60 abcd = 6 abce = 65 abcde = 63 (a) Construct a normal probability plot of the effect estimates. Which effects appear to be large? 647
48 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY From the normal probability plot of effects shown below, effects A, B, C, and the AB interaction appear to be large. DESIGNEXPERT Plot Yield Normal plot A: Aperture B: Exposure Time C: Develop Time D: Mask Dimension E: Etch Time Normal % probability C AB A B Effect (b) Conduct an analysis of variance to confirm your findings for part (a). Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B < C < AB < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. (c) Write down the regression model relating yield to the significant process variables. Design Expert Output Final Equation in Terms of Actual Factors: Aperture small Yield = * Exposure Time * Develop Time Aperture large Yield = * Exposure Time * Develop Time (d) Plot the residuals on normal probability paper. Is the plot satisfactory? 648
49 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Normal % probability E Residual There is nothing unusual about this plot. (e) Plot the residuals versus the predicted yields and versus each of the five factors. Comment on the plots. Residuals vs. Aperture Residuals vs. Exposure Time Residuals 3.557E05 Residuals 3.557E Aperture Exposure Time 649
50 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Residuals vs. Develop Time Residuals vs. Mask Dimension Residuals 3.557E05 Residuals 3.557E Develop Time Mask Dimension Residuals vs. Etch Time Residuals 3.557E Etch Time The plot of residual versus exposure time shows some very slight inequality of variance. There is no strong evidence of a potential problem. (f) Interpret any significant interactions. 650
51 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Yield 65 Interaction Graph Aperture X = B: Exposure Time Y = A: Aperture 50.5 A small A large Actual Factors C: Develop Time = D: Mask Dimension = Small 35.5 E: Etch Time = 5.00 Yield Exposure Time Factor A does not have as large an effect when B is at its low level as it does when B is at its high level. (g) What are your recommendations regarding process operating conditions? To achieve the highest yield, run B at the high level, A at the high level, and C at the high level. (h) Project the 5 design in this problem into a k design in the important factors. Sketch the design and show the average and range of yields at each run. Does this sketch aid in interpreting the results of this experiment? DESIGNEASE Analysis Actual Yield R= R=5 E x p o s u r e T i m e B R= R= R=5 B C A R= R=3 A+ Aperture R= D e v e l o p C+ T i m e This cube plot aids in interpretation. The strong AB interaction and the large positive effect of C are clearly evident Continuation of Problem 6.3. Suppose that the experimenter had run four runs at the center points in addition to the 3 trials in the original experiment. The yields obtained at the center point runs were 68, 74, 76, and
52 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Reanalyze the experiment, including a test for pure quadratic curvature. Because aperture and mask dimension are not continuous variables, the four center points were split amongst these two factors as follows. Aperture Mask Dimension Yield Small Small 68 Large Small 74 Small Large 76 Large Large 70 The sum of squares for the curvature can be estimated with the following equation and is confirmed with the analysis of variance shown in the Design Expert output. SS PureQuadratic n F n C n y y F F n C C Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B < C < AB < Curvature < significant Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. Discuss what your next step would be. Add axial points for factors B and C along with four more center points to fit a secondorder model and satisfy blocking concerns Consider the single replicate of the 4 design in Example 6.. Suppose we had arbitrarily decided to analyze the data assuming that all three and fourfactor interactions were negligible. Conduct this analysis and compare your results with those obtained in the example. Do you think that it is a good idea to arbitrarily assume interactions to be negligible even if they are relatively highorder ones? Design Expert Output Response: Filtration Ratein A/min ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F 65
53 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Model significant A B C D AB E AC AD BC BD CD Residual Cor Total The Model Fvalue of.9 implies the model is significant. There is only a 0.6% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC, AD are significant model terms. This analysis of variance identifies the same effects as the normal probability plot of effects approach used in Example 6.. In general, it is not a good idea to arbitrarily pool interactions. Use the normal probability plot of effect estimates as a guide in the choice of which effects to tentatively include in the model An experiment was conducted on a chemical process that produces a polymer. The four factors studied were temperature (A), catalyst concentration (B), time (C), and pressure (D). Two responses, molecular weight and viscosity, were observed. The design matrix and response data are shown below: Actual Run Run Molecular Factor Levels Number Order A B C D Weight Viscosity Low () High (+) A (ºC) B (%) C (min) D (psi)
54 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Consider only the molecular weight response. Plot the effect estimates on a normal probability scale. What effects appear important? DESIGNEXPERT Plot Molecular Wt Half Normal plot A: Temperature B: Catalyst Con. C: Time D: P re ssu re Half Normal % probability AB A C A, C and the AB interaction appear to be important. (b) Use an analysis of variance to confirm the results from part (a). Is there an indication of curvature? A,C and the AB interaction are significant. While the main effect of B is not significant, it could be included to preserve hierarchy in the model. There is no indication of quadratic curvature. Design Expert Output Response: Molecular Wt ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model.809E < significant A < C.60E E < AB < Curvature not significant Residual Lack of Fit not significant Pure Error Cor Total 3.037E The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, AB are significant model terms. (c) Write down a regression model to predict molecular weight as a function of the important variables. Design Expert Output Final Equation in Terms of Coded Factors: Molecular Wt = * A * C * A * B Effect 654
55 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (d) Analyze the residuals and comment on model adequacy. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted There are two residuals that appear to be large and should be investigated. (e) Repeat parts (a)  (d) using the viscosity response. DESIGNEXPERT Plot Viscosity Normal plot A: Temperature B: Catalyst Con. C: Time D: P re ssu re Normal % probability B A Effect Design Expert Output Response: Viscosity ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B < Curvature not significant Residual Lack of Fit not significant 655
56 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. Final Equation in Terms of Coded Factors: Viscosity = * A * B Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted There is one large residual that should be investigated. 656
57 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY 6.7. Continuation of Problem 6.6. Use the regression models for molecular weight and viscosity to answer the following questions. (a) Construct a response surface contour plot for molecular weight. In what direction would you adjust the process variables to increase molecular weight? Increase temperature, catalyst and time. DESIGNEXPERT Plot Molecular Wt X = A: Temperature Y = B: Catalyst Con. Design Points Actual Factors C: Time = 5.00 D: Pressure = B: Catalyst Con Molecular Wt A: Temperature (b) Construct a response surface contour plot for viscosity. In what direction would you adjust the process variables to decrease viscosity? DESIGNEXPERT Plot Viscosity X = A: Temperature Y = B: Catalyst Con Viscosity 575 Design Points Actual Factors C: Time = 5.00 D: Pressure = B: Catalys t Con A: Temperature Decrease temperature and catalyst. (c) What operating conditions would you recommend if it was necessary to produce a product with a molecular weight between 400 and 500, and the lowest possible viscosity? 657
58 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Overlay Plot X = A: Temperature Y = B: Catalyst Con. Actual Factors C: Time = 4.50 D: Pressure = Overlay Plot B: Catalys t Con Molecular Wt: 500 Viscosity: A: Temperature Set the temperature between 00 and 05, the catalyst between 4 and 5%, and the time at 4.5 minutes. The pressure was not significant and can be set at conditions that may improve other results of the process such as cost In a process development study on yield, four factors were studied, each at two levels: time (A), concentration (B), pressure (C), and temperature (D). A single replicate of a 4 design was run, and the resulting data are shown in the following table: Actual Run Run Yield Factor Levels Number Order A B C D (lbs) Low () High (+) A (h) B (%) C (psi) D (ºC)
59 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Construct a normal probability plot of the effect estimates. Which factors appear to have large effects? DESIGNEXPERT Plot Yield Normal plot A: Time B: Concentration C: P re ssu re D: Temperature Normal % probability AC C AD D A Effect A, C, D and the AC and AD interactions appear to have large effects. (b) Conduct an analysis of variance using the normal probability plot in part (a) for guidance in forming an error term. What are your conclusions? Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < C D AC < AD < Residual Cor Total The Model Fvalue of 33.9 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC, AD are significant model terms. (c) Write down a regression model relating yield to the important process variables. Design Expert Output Final Equation in Terms of Coded Factors: Yield = *A +.00 *C +.63 *D .3 *A*C +.00 *A*D 659
60 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Yield = * Time * Pressure * Temperature * Time * Pressure * Time * Temperature (d) Analyze the residuals from this experiment. Does your analysis indicate any potential problems? Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Residuals vs. Run Residuals Run Number There is nothing unusual about the residual plots. (e) Can this design be collapsed into a 3 design with two replicates? If so, sketch the design with the average and range of yield shown at each point in the cube. Interpret the results. 660
61 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEASE Analysis Actual yield 8.0 R=.0 R= t e m p e r a t u r e D+.5 R=3 8.5 R=3 4.5 R= D C A R= R= A+ time 5.0 C+ R=0 p r e s s u r e 6.9. Continuation of Problem 6.8. Use the regression model in part (c) of Problem 6.8 to generate a response surface contour plot of yield. Discuss the practical purpose of this response surface plot. The response surface contour plot shows the adjustments in the process variables that lead to an increasing or decreasing response. It also displays the curvature of the response in the design region, possibly indicating where robust operating conditions can be found. Two response surface contour plots for this process are shown below. These were formed from the model written in terms of the original design variables. DESIGNEXPERT Plot Yield X = A: Time Y = C: Pressure Yield DESIGNEXPERT Plot Yield X = A: Time Y = D: Temperature Yield Actual Factors B: Concentration = D: Temperature = Actual Factors B: Concentration = C: Pressure = Pressure Temperature Tim e Tim e 66
62 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY The scrumptious brownie experiment. The author is an engineer by training and a firm believer in learning by doing. I have taught experimental design for many years to a wide variety of audiences and have always assigned the planning, conduct, and analysis of an actual experiment to the class participants. The participants seem to enjoy this practical experience and always learn a great deal from it. This problem uses the results of an experiment performed by Gretchen Krueger at Arizona State University. There are many different ways to bake brownies. The purpose of this experiment was to determine how the pan material, the brand of brownie mix, and the stirring method affect the scrumptiousness of brownies. The factor levels were Factor Low () High (+) A = pan material Glass Aluminum B = stirring method Spoon Mixer C = brand of mix Expensive Cheap The response variable was scrumptiousness, a subjective measure derived from a questionnaire given to the subjects who sampled each batch of brownies. (The questionnaire dealt with such issues as taste, appearance, consistency, aroma, and so forth.) An eightperson test panel sampled each batch and filled out the questionnaire. The design matrix and the response data are shown below: Brownie Test Panel Results Batch A B C (a) Analyze the data from this experiment as if there were eight replicates of a 3 design. Comment on the results. Design Expert Output ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C AB AC BC ABC Residual Lack of Fit Pure Error Cor Total The Model Fvalue of.0 implies the model is significant. There is only a 4.75% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A are significant model terms. 66
63 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY In this analysis, A, the pan material and B, the stirring method, appear to be significant. There are 56 degrees of freedom for the error, yet only eight batches of brownies were cooked, one for each recipe. (b) Is the analysis in part (a) the correct approach? There are only eight batches; do we really have eight replicates of a 3 factorial design? The different rankings by the tastetest panel are not replicates, but repeat observations by different testers on the same batch of brownies. It is not a good idea to use the analysis in part (a) because the estimate of error may not reflect the batchtobatch variation. (c) Analyze the average and standard deviation of the scrumptiousness ratings. Comment on the results. Is this analysis more appropriate than the one in part (a)? Why or why not? DESIGNEXPERT Plot Average Normal plot DESIGNEXPERT Plot Stdev Normal plot A: Pan Material B: Stirring Method C: Mix Brand A A: Pan Material B: Stirring Method C: Mix Brand C Normal % probability B Normal % probability AC A Effect Effect Design Expert Output Response: Average ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B Residual Cor Total.65 7 The Model Fvalue of 76.3 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. Design Expert Output Response: Stdev ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A C AC Residual Cor Total
64 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY The Model Fvalue of 9.77 implies the model is significant. There is only a.59% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case AC are significant model terms. Variables A and B affect the mean rank of the brownies. Note that the AC interaction affects the standard deviation of the ranks. This is an indication that both factors A and C have some effect on the variability in the ranks. It may also indicate that there is some inconsistency in the taste test panel members. For the analysis of both the average of the ranks and the standard deviation of the ranks, the mean square error is now determined by pooling apparently unimportant effects. This is a more accurate estimate of error than obtained assuming that all observations were replicates Consider the single replicate of the 4 design in Example 6.. Suppose that we ran five points at the center (0, 0, 0, 0) and observed the following responses: 9, 95, 9, 89, and 96. Test for curvature in this experiment. Interpret the results. Design Expert Output Response Filtration Rate ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ATemperature < CFormaldehyde DStirring Rate < AC < AD < Curvature < significant Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC, AD are significant model terms. The "Curvature Fvalue" of 8.65 implies there is significant curvature (as measured by difference between the average of the center points and the average of the factorial points) in the design space. There is only a 0.0% chance that a "Curvature Fvalue" this large could occur due to noise. The curvature test indicates that the model has significant pure quadratic curvature A missing value in a k factorial. It is not unusual to find that one of the observations in a k design is missing due to faulty measuring equipment, a spoiled test, or some other reason. If the design is replicated n times (n>) some of the techniques discussed in Chapter 5 can be employed. However, for an unreplicated factorial (n) some other method must be used. One logical approach is to estimate the missing value with a number that makes the highestorder interaction contrast zero. Apply this technique 664
65 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY to the experiment in Example 6. assuming that run ab is missing. Compare the results with the results of Example 6.. Treatment Combination Response * ABCD ABCD A B C D Response () a b ab missing missing *   c ac bc abc d ad bd abd cd acd bcd abcd Contrast (ABCD) = Missing 54 = 0 Missing = 54 Substitute the value 54 for the missing run at ab. From the effects list and half normal plot shown below, factors A, C, D, AC, and AD appear to be large; the same result as found in Example 6.. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Model AB Model AC Model AD Model BC Model BD Model CD Model ABC Model ABD Model ACD Model BCD Model ABCD Lenth's ME.5676 Lenth's SME
66 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Filtration Rate Normal plot A: Temperature B: Pressure C: Concentration D: Stirring Rate Normal % probability AC C A AD D Effect Resistivity on a silicon wafer is influenced by several factors. The results of a 4 factorial experiment performed during a critical process step is shown below. Run A B C D Resistivity
67 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Estimate the factor effects. Plot the effect estimates on a normal probability plot and select a tentative model. DesignExpert Software Resistivity Normal Plot ShapiroWilk test Wvalue = 0.98 pvalue = 0.99 A: A B: B C: C D: D Positive Effects Negative Effects Normal % Probability B AB A Standardized Effect From the normal plot, factors A, B and the AB interaction appear to be significant. (b) Fit the model identified in part (a) and analyze the residuals. Is there any indication of model inadequacy? Design Expert Output Response Resistivity ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA < BB < AB < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, AB are significant model terms. 667
68 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. A Residuals vs. B Residuals Residuals A B:B The normal probability plot of residuals is not satisfactory. The plots of residual versus predicted, residual versus factor A, and the residual versus factor B are funnel shaped indicating nonconstant variance. 668
69 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (c) Repeat the analysis from parts (a) and (b) using ln(y) as the response variable. Is there any indication that the transformation has been useful? Normal Plot A Normal % Probability B Standardized Effect Design Expert Output Response Resistivity Transform: Natural log Constant: 0 ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA < BB < Residual Cor Total.0 5 The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. The transformed data no longer indicates that the AB interaction is significant. A simpler model has resulted from the log transformation. 669
70 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. A Residuals vs. B Residuals Residuals A B:B The residual plots are much improved. (d) Fit a model in terms of the coded variables that can be used to predict the resistivity. Design Expert Output Final Equation in Terms of Coded Factors: Ln(Resistivity) = * A 0.3 * B 670
71 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Continuation of Problem Suppose that the experiment had also run four center points along with the 6 runs in Problem The resistivity measurements at the center points are: 8.5, 7.63, 8.95, and Analyze the experiment again incorporating the center points. What conclusions can you draw now? DesignExpert Software Resistivity Normal Plot Error from replicates ShapiroWilk test Wvalue = 0.98 pvalue = 0.99 A: A B: B C: C D: D Positive Effects Negative Effects Normal % Probability B AB A Standardized Effect Design Expert Output Response Resistivity ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA < BB < AB < Curvature < significant Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 9.5 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, AB are significant model terms. 67
72 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Because of the funnel shaped residual versus predicted plot, the analysis was repeated with the natural log transformation. DesignExpert Software Ln(Resistivity) Normal Plot Error from replicates ShapiroWilk test Wvalue = 0.96 pvalue = 0.79 A: A B: B C: C D: D Positive Effects Negative Effects Normal % Probability A 5 B Standardized Effect Design Expert Output Transform: Natural log Constant: 0 ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA < BB < Curvature < significant Residual Lack of Fit not significant Pure Error Cor Total
73 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. The "Curvature Fvalue" of 88.9 implies there is significant curvature (as measured by difference between the average of the center points and the average of the factorial points) in the design space. There is only a 0.0% chance that a "Curvature Fvalue" this large could occur due to noise. The curvature test indicates that the model has significant pure quadratic curvature An engineer has performed an experiment to study the effect of four factors on the surface roughness of a machined part. The factors (and their levels) are A = tool angle ( degrees, 5 degrees), B = cutting fluid viscosity (300, 400), C = feed rate (0 in/min, 5 in/min), and D = cutting fluid cooler used (no, yes). The data from this experiment (with the factors coded to the usual , + levels) are shown below. Run A B C D Surface Roughness
74 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (a) Estimate the factor effects. Plot the effect estimates on a normal probability plot and select a tentative model. DESIGNEXPERT Plot Surface Roughness Normal plot A: Tool Angle B: Viscosity C: Feed Rate D: Cutting Fluid Normal % probability B AB A C Effect (b) Fit the model identified in part (a) and analyze the residuals. Is there any indication of model inadequacy? Design Expert Output Response: Surface Roughness ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model 6.406E E < significant A 8.556E E < B 3.080E E006. < C.030E E < AB.440E E < Residual.53E E008 Cor Total 6.559E The Model Fvalue of 4.97 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. 674
75 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability E005 Residuals 5E E E005 5E E Residual Predicted Residuals vs. Tool Angle E005 Residuals 5E E Tool Angle The plot of residuals versus predicted shows a slight ushaped appearance in the residuals, and the plot of residuals versus tool angle shows an outwardopening funnel. (c) Repeat the analysis from parts (a) and (b) using /y as the response variable. Is there and indication that the transformation has been useful? The plots of the residuals are more representative of a model that does not violate the constant variance assumption. 675
76 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot.0/(Surface Roughness) Normal plot A: Tool Angle B: Viscosity C: Feed Rate D: Cutting Fluid Normal % probability C AB A B Effect Design Expert Output Response: Surface RoughnessTransform:Inverse ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model.059E < significant A < B < C < AB < Residual Cor Total.063E The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. DESIGNEXPERT Plot.0/(Surface Roughness) Residuals vs. Predicted DESIGNEXPERT Plot.0/(Surface Roughness) Residuals vs. Tool Angle Residuals Residuals Predicted Tool Angle 676
77 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY (d) Fit a model in terms of the coded variables that can be used to predict the surface roughness. Convert this prediction equation into a model in the natural variables. Design Expert Output Final Equation in Terms of Coded Factors:.0/(Surface Roughness) = * A * B * C * A * B Suppose that you want to run a 3 factorial design. The variance of an individual observation is expected to be about 4. Suppose that you want the length of a 95% confidence interval on any effect to be less than or equal to.5. How many replicates of the design do you need to run? With the equations for the se(effect) and 00() percent confidence interval on the effects shown below, we can iteratively estimate the number of replicates. From the table of iterations, 0 replicates are required. se(effect) = S n k Effect ± t (/,Np) se(effect) n se(effect) t(0.05,np) 95% CI Length Often the fitted regression model from a k factorial design is used to make predictions at points of interest in the design space. (a) Find the variance of the predicted response ŷ at the point x, x,, x k in the design space. Hint: Remember that the x s are coded variables, and assume a k design with an equal number of replicates n at each design point so that the variance of a regression coefficient ˆ is covariance between any pair of regression coefficients is zero. Let s assume that the model can be written as follows:?? yˆ( x )= 0 xx... pxp k n and that the where x [ x, x,..., x k ] are the values of the original variables in the design at the point of interest where a prediction is required, and the variables in the model x, x,..., xp potentially include interaction terms among the original k variables. Now the variance of the predicted response is 677
78 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY V[ yˆ ( x)] V(?? 0 xx... pxp) V(? ) V( x ) V(? x )... V( x ) n 0 p k xi i This result follows because the design is orthogonal and all model parameter estimates have the same variance. Remember that some of the x s involved in this equation are potentially interaction terms. (b) Use the result of part (a) to find an equation for a 00()% confidence interval on the true mean response at the point x, x,, x k in the design space. The confidence interval is /, dfe /, dfe y? ( x) t V[ y( x)] y( x) y? ( x) t V[ y( x )] p p where df E is the number of degrees of freedom used to estimate and the estimate of has been used in computing the variance of the predicted value of the response at the point of interest Hierarchical Models. Several times we have utilized the hierarchy principal in selecting a model; that is, we have included nonsignificant terms in a model because they were factors involved in significant higherorder terms. Hierarchy is certainly not an absolute principle that must be followed in all cases. To illustrate, consider the model resulting from Problem 6., which required that a nonsignificant main effect be included to achieve hierarchy. Using the data from Problem 6.: (a) Fit both the hierarchical model and the nonhierarchical model. Design Expert Output for Hierarchial Model Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B < C AC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.54 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, C, AC are significant model terms. Std. Dev RSquared Mean Adj RSquared C.V Pred RSquared PRESS 98.5 Adeq Precision The "Pred RSquared" of is in reasonable agreement with the "Adj RSquared" of A difference greater than 0.0 between the "Pred RSquared" and the "Adj RSquared" 678
79 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY indicates a possible problem with your model and/or data. Design Expert Output for NonHierarchical Model Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant B < C AC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.57 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, C, AC are significant model terms. The "Lack of Fit Fvalue" of 0.78 implies the Lack of Fit is not significant relative to the pure error. There is a 55.66% chance that a "Lack of Fit Fvalue" this large could occur due to noise. Nonsignificant lack of fit is good  we want the model to fit. Std. Dev RSquared Mean Adj RSquared C.V. 3.5 Pred RSquared PRESS 89.9 Adeq Precision.30 The "Pred RSquared" of is in reasonable agreement with the "Adj RSquared" of A difference greater than 0.0 between the "Pred RSquared" and the "Adj RSquared" indicates a possible problem with your model and/or data. (b) Calculate the PRESS statistic, the adjusted R and the mean square error for both models. The PRESS and R are in the Design Expert Output above. The PRESS is smaller for the nonhierarchical model than the hierarchical model suggesting that the nonhierarchical model is a better predictor. (c) Find a 95 percent confidence interval on the estimate of the mean response at a cube corner ( x = x = x 3 = ). Hint: Use the result of Problem Design Expert Output Prediction SE Mean 95% CI low 95% CI high SE Pred 95% PI low 95% PI high Life Life Life Life Life Life Life Life (d) Based on the analyses you have conducted, which model would you prefer? Notice that PRESS is smaller and the adjusted R is larger for the nonhierarchical model. This is an indication that strict adherence to the hierarchy principle isn t always necessary. Note also that the confidence interval is shorter for the nonhierarchical model. 679
80 Solutions from Montgomery, D. C. (004) Design and Analysis of Experiments, Wiley, NY 680
81 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter 7 Blocking and Confounding in the k Factorial Design Solutions 7.. Consider the experiment described in Problem 6.. Analyze this experiment assuming that each one of the four replicates represents a block. Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 Bit Size (A) * Cutting Speed (B) * AB * Blocks Error Total These results agree with those from Problem 6.. Bit size, cutting speed and their interaction are significant at the % level. Design Expert Output Response : Vibration ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Block Model < significant ABit Size < BCutting Speed < AB < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, AB are significant model terms. Values greater than indicate the model terms are not significant. 7. Consider the experiment described in Problem 6.. Analyze this experiment assuming that each replicate represents a block of a single production shift. Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 Cutting Speed (A) < Tool Geometry (B) * Cutting Angle (C) * AB < AC * 7
82 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY BC ABC < Blocks Error Total Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model significant A B C AC Residual Cor Total The Model Fvalue of.3 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, C, AC are significant model terms. These results agree with the results from Problem 6.. Tool geometry, cutting angle and the interaction between cutting speed and cutting angle are significant at the 5% level. The Design Expert program also includes factor A, cutting speed, in the model to preserve hierarchy Consider the data from the first replicate of Problem 6.. Suppose that these observations could not all be run using the same bar stock. Set up a design to run these observations in two blocks of four observations each with ABC confounded. Analyze the data. Block Block () a ab b ac c bc abc From the normal probability plot of effects, B, C, and the AC interaction are significant. Factor A was included in the analysis of variance to preserve hierarchy. 7
83 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Life Normal plot A: Cutting Speed B: Tool Geometry C: Cutting Angle B Normal % probability AC C Effect Design Expert Output Response: Life in hours ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model not significant A B C AC Residual Cor Total The "Model Fvalue" of 7.3 implies the model is not significant relative to the noise. There is a.38 % chance that a "Model Fvalue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case there are no significant model terms. This design identifies the same significant factors as Problem Consider the data from the first replicate of Problem 6.. Construct a design with two blocks of eight observations each with ABCD confounded. Analyze the data. Block Block () a ab b ac c bc d ad abc bd abd cd acd abcd bcd 73
84 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The significant effects are identified in the normal probability plot of effects below: DESIGNEXPERT Plot yield Normal plot A: A B: B C: C D: D Normal % probability A AD ABC D ABD AB Effect AC, BC, and BD were included in the model to preserve hierarchy. Design Expert Output Response: yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model significant A B C D AB AC AD BC BD ABC ABD Residual Cor Total The Model Fvalue of 9.64 implies the model is significant. There is only a 4.38% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, D, ABC, ABD are significant model terms Repeat Problem 7.4 assuming that four blocks are required. Confound ABD and ABC (and consequently CD) with blocks. The block assignments are shown in the table below. The normal probability plot of effects identifies factors A and D, and the interactions AB, AD, and the ABCD as strong candidates for the model. For hierarchal purposes, factor B was included in the model; however, hierarchy is not preserved for the ABCD interaction allowing an estimate for error. 74
85 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Block Block Block 3 Block 4 () ac c a ab bc abc b acd d ad cd bcd abd bd abcd DESIGNEXPERT Plot yield Normal plot A: A B: B C: C D: D D Normal %probabili ty B AB ABCD 0 0 AD 5 A Effect Design Expert Output Response: yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model significant A B D AB AD ABCD Residual Cor Total The Model Fvalue of 9.6 implies the model is significant. There is only a 0.% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, D, AB, AD, ABCD are significant model terms Consider the alloy cracking experiment described in Problem 6.. Suppose that only 6 runs could be made on a single day, so each replicate was treated as a block. Analyze the experiment and draw conclusions. The analysis of variance for the full model is as follows: Design Expert Output Response: Crack Lengthin mm x 0^ ANOVA for Selected Factorial Model 75
86 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B < C < D < AB < AC < AD BC BD CD ABC < ABD ACD.96E E BCD ABCD.596E E Residual Cor Total The Model Fvalue of 445. implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, D, AB, AC, ABC are significant model terms. The analysis of variance for the reduced model based on the significant factors is shown below. The BC interaction was included to preserve hierarchy. Design Expert Output Response: Crack Lengthin mm x 0^ ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B < C < D < AB < AC < BC ABC < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, D, AB, AC, ABC are significant model terms. Blocking does not change the results of Problem Consider the fill height deviation experiment in Problem 6.8. Suppose that each replicate was run on a separate day. Analyze the data assuming that the days are blocks. Design Expert Output Response: Fill Deviation 76
87 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B C AB Residual Cor Total The Model Fvalue of 8.30 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C are significant model terms. The analysis is very similar to the original analysis in chapter 6. The same effects are significant Consider the fill height deviation experiment in Problem 6.8. Suppose that only four runs could be made on each shift. Set up a design with ABC confounded in replicate and AC confounded in replicate. Analyze the data and comment on your findings. Design Expert Output Response: Fill Deviation ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model significant A B C AB Residual Cor Total The Model Fvalue of 4.6 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C are significant model terms. The analysis is very similar to the original analysis of Problem 6.8 and that of Problem 7.7. The AB interaction is less significant in this scenario Using the data from the 4 design in Problem 6.7, construct and analyze a design in two blocks with ABCD confounded with blocks. Design Expert Output Response: UEC ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block.500E E005 Model < significant A <
88 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY C < D AC Residual E003 Cor Total The Model Fvalue of 3.33 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC are significant model terms. The analysis is similar to that of Problem 6.7. The significant effects are A, C, D and AC Using the data from the 5 design in Problem 6.3, construct and analyze a design in two blocks with ABCDE confounded with blocks. Block Block Block Block () ae a e ab be b abe ac ce c ace bc abce abc bce ad de d ade bd abde abd bde cd acde acd cde abcd bcde bcd abcde The normal probability plot of effects identifies factors A, B, C, and the AB interaction as being significant. This is confirmed with the analysis of variance. DESIGNEXPERT Plot Yield Normal plot A: Aperture B: Exposure Time C: Develop Time D: Mask Dimension E: Etch Time Normal % probability C AB A B Effect Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F 78
89 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Block Model < significant A < B < C < AB < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. 7.. Repeat Problem 7.0 assuming that four blocks are necessary. Suggest a reasonable confounding scheme. Use ABC and CDE, and consequently ABDE. The four blocks follow. Block Block Block 3 Block 4 () a ac c ab b bc abc acd cd d ad bcd abcd abd bd ace ce e ae bce abce abe be de ade acde cde abde bde bcde abcde The normal probability plot of effects identifies the same significant effects as in Problem 7.0. DESIGNEXPERT Plot Yield Normal plot A: Aperture B: Exposure Time C: Develop Time D: Mask Dimension E: Etch Time Normal % probability C AB A B Effect Design Expert Output Response: Yield ANOVA for Selected Factorial Model 79
90 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B < C < AB < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms. 7.. Consider the data from the 5 design in Problem 6.3. Suppose that it was necessary to run this design in four blocks with ACDE and BCD (and consequently ABE) confounded. Analyze the data from this design. Block Block Block 3 Block 4 () a b c ae e abe ace cd acd bcd d abc bc ac ab acde cde abcde ade bce abce ce be abd bd ad abcd bde abde de bcde Even with four blocks, the same effects are identified as significant per the normal probability plot and analysis of variance below: DESIGNEXPERT Plot Yield Normal plot A: Aperture B: Exposure Time C: Develop Time D: Mask Dimension E: Etch Time Normal % probability C AB A B Effect 70
91 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B < C < AB < Residual Cor Total The Model Fvalue of 9.6 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, C, AB are significant model terms Consider the putting experiment in Problem 6.0. Analyze the data considering each replicate as a block. The analysis is similar to that of Problem 6.0. Blocking has not changed the significant factors, however, the residual plots show that the normality assumption has been violated. The transformed data also has similar analysis to the transformed data of Problem 6.0. The ANOVA shown is for the transformed data. Design Expert Output Response: Distance from cuptransform:square root Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model significant A B Residual Cor Total The Model Fvalue of 7.83 implies the model is significant. There is only a 0.07% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms Design an experiment for confounding a 6 factorial in four blocks. Suggest an appropriate confounding scheme, different from the one shown in Table 7.8. We choose ABCE and ABDF, which also confounds CDEF. Block Block Block 3 Block 4 a c ac () b abc bc ab cd ad d acd 7
92 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY abcd bd abd bcd ace e ae ce bce abe be abce de acde cde ade abde bcde abcde bde cf af f acf abcf bf abf bcf adf cdf acdf df bdf abcdf bcdf abdf ef acef cef aef abef bcef abcef bef acdef def adef cdef bcdef abdef bdef abcdef 7.5. Consider the direct mail experiment in Problem 6.. Suppose that each group of customers is in different parts of the country. Support an appropriate analysis for the experiment. Set up each Group (replicate) as a geographic region. The analysis is similar to that of Problem 6.. Factors A and B are included to achieve a hierarchical model. Design Expert Output Response Number of Orders ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Block Model significant AClass BType C$ AB AC BC Residual Cor Total The Model Fvalue of 9.97 implies the model is significant. There is only a 0.4% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case C, AB, AC, BC are significant model terms Consider the design in two blocks with AB confounded. Prove algebraically that SS AB = SS Blocks. If AB is confounded, the two blocks are: Block Block () a ab b () + ab a + b 7
93 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY SS SS SS Blocks Blocks Blocks ab ab abab 4 ab ab a b ab ab a b ab a b a ab b ab ab 4 ab a b ab ab a b aabbab SSBlocks ab a b SS 4 AB Consider the 6 design in eight blocks of eight runs each with ABCD, ACE, and ABEF as the independent effects chosen to be confounded with blocks. Generate the design. Find the other effects confound with blocks. Block Block Block 3 Block 4 Block 5 Block 6 Block 7 Block 8 b abc a c ac () bc ab acd d bcd abd bd abcd ad cd ce ae abce be abe bce e ace abde bcde de acde cde ade abcde bde abcf bf cf af f acf abf bcf df acdf abdf bcdf abcdf bdf cdf adf aef cef bef abcef bcef abef acef ef bcdef abdef acdef def adef cdef bdef abcdef The factors that are confounded with blocks are ABCD, ABEF, ACE, BDE, CDEF, BCF, and ADF Repeat the analysis of Problem 6. assuming that ABC was confounded with blocks in each replicate. Replicate I, II, and III (ABC Confounded) Block> () a ab b ac c bc abc Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 A < B C AB < 73
94 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY AC BC Blocks (or ABC) Replicates/Lack of Fit 0.58 Error Total Suppose that the data in Problem 6. we had confounded ABC in replicate I, AB in replicate II, and BC in replicate III. Construct the analysis of variance table. Replicate I Replicate II Replicate III (ABC Confounded) (AB Confounded) (BC Confounded) Block> () a () a () b ab b ab b bc c ac c abc ac abc ab bc abc c bc a ac Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 A < B C AB (reps and III) < AC BC (reps I and II) < ABC (reps II and III) < Blocks within replicates Replicates 0.58 Error Total Consider the data in Example 7.. Suppose that all the observations in block are increased by 0. Analyze the data that would result. Estimate the block effect. Can you explain its magnitude? Do blocks now appear to be an important factor? Are any other effect estimates impacted by the change you made in the data? Block Effect yblock yblock This is the block effect estimated in Example 7. plus the additional 0 units that were added to each observation in block. All other effects are the same. 74
95 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 A C D AC AD Blocks Error Total Design Expert Output Response: Filtration in gal/hr ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < C D AC < AD < Residual Cor Total The Model Fvalue of 53.3 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, C, D, AC, AD are significant model terms. 7.. Construct a 3 design with ABC confounded in the first two replicates and BC confounded in the third. Outline the analysis of variance and comment on the information obtained. Replicate I Replicate II Replicate III (ABC Confounded) (ABC Confounded) (BC Confounded) Block> () a () a () b ab b ab b bc c ac c ac c abc ab bc abc bc abc a ac 75
96 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Source of Degrees of Variation Freedom A B C AB AC BC ABC Replicates Blocks 3 Error Total 3 This design provides twothirds information on BC and onethird information on ABC. 7.. Suppose that in Problem 6. ABCD was confounded in replicate I and ABC was confounded in replicate II. Perform the statistical analysis of variance. Source of Sum of Degrees of Mean Variation Squares Freedom Square F0 A B C D AB AC < AD BC < BD < CD ABC ABD ACD < BCD < ABCD Replicates.8.8 Blocks Error Total
97 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter 8 TwoLevel Fractional Factorial Designs Solutions 8.. Consider the plasma etch experiment described in Example 6.. Suppose that only a onehalf fraction of the design could be run. Set up the design and analyze the data. Because Example 6. is a replicated 3 factorial experiment, a half fraction of this design is a 3 with four runs. The experiment is replicates to assure an adequate estimate of the MS E. Etch Rate Factor Levels A B C=AB (A/min) Low () High (+) A (Gap, cm) B (C F 6 flow, SCCM) C (Power, W) The analysis shown below identifies all three main effects as significant. Because this is a resolution III design, the main effects are aliased with two factor interactions. The original analysis from Example 6 identifies factors A, C, and the AC interaction as significant. In our replicated half fraction experiment, factor B is aliased with the AC interaction. This problem points out the concerns of running small resolution III designs. Design Expert Output Response: Etch Rate ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model.5E significant A B C.58E E Pure Error Cor Total.39E The Model Fvalue of 3.6 implies the model is significant. There is only a 0.30% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared 0.99 C.V. 6. Pred RSquared PRESS Adeq Precision.48 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AGap BCF6 Flow CPower Final Equation in Terms of Coded Factors: 8
98 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Etch Rate = * A * B * C Final Equation in Terms of Actual Factors: Etch Rate = * Gap * CF6 Flow * Power 8.. Suppose that in the chemical process development experiment in Problem 6., it was only possible to run a onehalf fraction of the 4 design. Construct the design and perform the statistical analysis, using the data from replicate. The required design is a 4 with I=ABCD. A B C D=ABC () ad bd ab cd ac bc abcd 80 Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Model AB Model AC Model AD Error BC Aliased Error BD Aliased Error CD Aliased Error ABC Aliased Error ABD Aliased Error ACD Aliased Error BCD Aliased Error ABCD Aliased Lenth's ME.5856 Lenth's SME
99 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Half Normal plot 99 Half Normal % probability D B AD AB A Effect The largest effect is A. The next largest effects are the AB and AD interactions. A plausible tentative model would be A, AB and AD, along with B and D to preserve hierarchy. Design Expert Output Response: yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B D AB AD Residual Cor Total The "Model Fvalue" of 4.87 implies the model is not significant relative to the noise. There is a 7.9 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev. 4. RSquared 0.94 Mean Adj RSquared C.V Pred RSquared PRESS Adeq Precision 6.44 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AA BB DD AB AD Final Equation in Terms of Coded Factors: yield = * A * B * D * A * B .50 * A * D 83
100 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: yield = * A * B * D * A * B * A * D The DesignExpert output indicates that we really only need the main effect of factor A. The updated analysis is shown below: Design Expert Output Response: yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A Residual Cor Total The Model Fvalue of 0.80 implies the model is significant. There is only a.67% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 5.6 RSquared Mean Adj RSquared C.V Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AA Final Equation in Terms of Coded Factors: yield = * A Final Equation in Terms of Actual Factors: yield = * A 8.3. Suppose that in Problem 6., only a onehalf fraction of the 4 design could be run. Construct the design and perform the analysis, using the data from replicate I. The required design is a 4 with I=ABCD. A B C D=ABC () ad bd ab cd ac bc
101 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY abcd 0.85 Design Expert Output Term Effect Effects SumSqr % Contribtn Require Intercept Model A Model B Model C Model D Model AB Model AC Model AD Lenth's ME Lenth's SME B, D, and AC + BD are the largest three effects. Now because the main effects of B and D are large, the large effect estimate for the AC + BD alias chain probably indicates that the BD interaction is important. It is also possible that the AB interaction is actually the CD interaction. This is not an easy decision. Additional experimental runs may be required to dealias these two interactions. DESIGNEXPERT Plot Crack Length Half Normal plot A: Pour Temp B: Titanium Content C: Heat Treat Method D: Grain Ref iner Half Normal %probability AC B D 40 A C 0 AB Effect Design Expert Output Response: Crack Lengthin mm x 0^ ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B C D AB BD Residual Cor Total The Model Fvalue of 87.9 implies there is a 8.8% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean.04 Adj RSquared C.V Pred RSquared
102 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY PRESS 9.3 Adeq Precision 5.0 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept APour Temp BTitanium Content CHeat Treat Method DGrain Refiner AB BD Final Equation in Terms of Coded Factors: Crack Length = * A +.0 * B .73 * C +.53 * D +0.9 * A * B .80 * B * D Final Equation in Terms of Actual Factors: Crack Length = * Pour Temp * Titanium Content * Heat Treat Method * Grain Refiner * Pour Temp * Titanium Content * Titanium Content * Grain Refiner 8.4. An article in Industrial and Engineering Chemistry ( More on Planning Experiments to Increase Research Efficiency, 970, pp ) uses a 5 design to investigate the effect of A = condensation, B = amount of material, C = solvent volume, D = condensation time, and E = amount of material on yield. The results obtained are as follows: e = 3. ad = 6.9 cd = 3.8 bde = 6.8 ab = 5.5 bc = 6. ace = 3.4 abcde = 8. (a) Verify that the design generators used were I = ACE and I = BDE. A B C D=BE E=AC e ad bde ab cd ace bc abcde (b) Write down the complete defining relation and the aliases for this design. I=BDE=ACE=ABCD. A (BDE) =ABDE A (ACE) =CE A (ABCD) =BCD A=ABDE=CE=BCD 86
103 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY B (BDE) =DE B (ACE) =ABCE B (ABCD) =ACD B=DE=ABCE=ACD C (BDE) =BCDE C (ACE) =AE C (ABCD) =ABD C=BCDE=AE=ABD D (BDE) =BE D (ACE) =ACDE D (ABCD) =ABC D=BE=ACDE=ABC E (BDE) =BD E (ACE) =AC E (ABCD) =ABCDE E=BD=AC=ABCDE AB (BDE) =ADE AB (ACE) =BCE AB (ABCD) =CD AB=ADE=BCE=CD AD (BDE) =ABE AD (ACE) =CDE AD (ABCD) =BC AD=ABE=CDE=BC (c) Estimate the main effects. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Model E (d) Prepare an analysis of variance table. Verify that the AB and AD interactions are available to use as error. The analysis of variance table is shown below. Part (b) shows that AB and AD are aliased with other factors. If all twofactor and three factor interactions are negligible, then AB and AD could be pooled as an estimate of error. Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B C D E Residual Cor Total The "Model Fvalue" of 3. implies the model is not significant relative to the noise. There is a 5.37 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev..3 RSquared Mean 9.4 Adj RSquared C.V..57 Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ACondensation BMaterial CSolvent DTime EMaterial Final Equation in Terms of Coded Factors: Yield = * A .59 * B +.4 * C * D 87
104 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY +.4 * E Final Equation in Terms of Actual Factors: Yield = * Condensation * Material * Solvent * Time * Material (e) Plot the residuals versus the fitted values. Also construct a normal probability plot of the residuals. Comment on the results. The residual plots are satisfactory. Residuals vs. Predicted Normal plot of residuals Residuals E Normal % probability E Predicted Residual 8.5. An article by J.J. Pignatiello, Jr. And J.S. Ramberg in the Journal of Quality Technology, (Vol. 7, 985, pp ) describes the use of a replicated fractional factorial to investigate the effects of five factors on the free height of leaf springs used in an automotive application. The factors are A = furnace temperature, B = heating time, C = transfer time, D = hold down time, and E = quench oil temperature. The data are shown in Table P8.. Table P8. A B C D E Free Height
105 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) Write out the alias structure for this design. What is the resolution of this design? I=ABCD, Resolution IV A (ABCD)= BCD B (ABCD)= ACD C (ABCD)= ABD D (ABCD)= ABC E (ABCD)= ABCDE AB (ABCD)= CD AC (ABCD)= BD AD (ABCD)= BC AE (ABCD)= BCDE BE (ABCD)= ACDE CE (ABCD)= ABDE DE (ABCD)= ABCE (b) Analyze the data. What factors influence the mean free height? Design Expert Output Term Effect SumSqr % Contribtn Require Intercept Model AFurnace Temp Model BHeating Time Error CTransfer Time Error DHold Down Time Model EQuench Oil Temp Error AB Error AC E Error AD Error AE Aliased BC Aliased Aliased BD Aliased Model BE Aliased CD Aliased Error CE Error DE Error Pure Error Lenth's ME Lenth's SME
106 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Free Height Normal Plot Error from replicates ShapiroWilk test Wvalue = pvalue = A: Furnace Temp B: Heating Time C: Transfer Time D: Hold Down Time E: Quench Oil Temp Positive Effects Negative Effects Normal % Probability BE A 5 E B Standardized Effect Design Expert Output Response Free Height ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AFurnace Temp < BHeating Time EQuench Oil Temp < BE Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 3.9 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0.4 RSquared Mean 7.63 Adj RSquared C.V. %.88 Pred RSquared PRESS.0 Adeq Precision Factor Estimate df Error Low High VIF Intercept AFurnace Temp BHeating Time EQuench Oil Temp BE Final Equation in Terms of Coded Factors: Free Height = * A * B 0. * E * B * E 80
107 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Free Height = * Furnace Temp * Heating Time * Quench Oil Temp * Heating Time * Quench Oil Temp (c) Calculate the range and standard deviation of the free height for each run. Is there any indication that any of these factors affects variability in the free height? Design Expert Output (Range) Term Effect SumSqr % Contribtn Require Intercept Model AFurnace Temp Model BHeating Time Error CTransfer Time Error DHold Down Time Error EQuench Oil Temp Error AB Error AC Error AD Error AE E Aliased BC Aliased Aliased BD Aliased Error BE Aliased CD Aliased Model CE Error DE Aliased ABC Aliased Aliased ABD Aliased Error ABE Aliased ACD Aliased Error ACE Model ADE Aliased BCD Aliased Aliased BCE Aliased Aliased BDE Aliased Aliased CDE Aliased Lenth's ME Lenth's SME
108 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Range HalfNormal Plot ShapiroWilk test Wvalue = 0.90 pvalue = 0.96 A: Furnace Temp B: Heating Time C: Transfer Time D: Hold Down Time E: Quench Oil Temp Positive Effects Negative Effects HalfNormal % Probability A B BCE CE C E BE Standardized Effect Design Expert Output (Range) ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant AFurnace Temp BHeating Time CTransfer Time.756E E EQuench Oil Temp 7.563E E BC 5.56E E BE.056E E CE BCE Residual E003 Cor Total The Model Fvalue of 5.70 implies the model is significant. There is only a.67% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 0. Adj RSquared C.V. % 35.5 Pred RSquared PRESS 0. Adeq Precision 7.66 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AFurnace Temp BHeating Time CTransfer Time EQuench Oil Temp E BC BE8.5E CE BCE
109 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Range = * A * B * C E003 * E * B * C +8.5E003 * B * E * C * E * B * C * E Final Equation in Terms of Actual Factors: Range = * Furnace Temp * Heating Time * Transfer Time E003 * Quench Oil Temp * Heating Time * Transfer Time E003 * Heating Time * Quench Oil Temp * Transfer Time * Quench Oil Temp * Heating Time * Transfer Time * Quench Oil Temp Design Expert Output (StDev) Term Effect SumSqr % Contribtn Require Intercept Model AFurnace Temp Model BHeating Time Model CTransfer Time Error DHold Down Time Model EQuench Oil Temp Error AB Error AC Aliased AD Aliased Error AE E Model BC Aliased BD Aliased Model BE Aliased CD Aliased Model CE Error DE E Aliased ABC Aliased Aliased ABD Aliased Error ABE Aliased ACD Aliased Error ACE Aliased ADE Aliased Aliased BCD Aliased Model BCE Aliased BDE Aliased Aliased CDE Aliased Lenth's ME Lenth's SME
110 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Std. Dev. HalfNormal Plot ShapiroWilk test Wvalue = pvalue = 0.89 A: Furnace Temp B: Heating Time C: Transfer Time D: Hold Down Time E: Quench Oil Temp Positive Effects Negative Effects HalfNormal % Probability BC A B CE BCE C BE E Standardized Effect Design Expert Output (StDev) Response Std. Dev. ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant AFurnace Temp BHeating Time CTransfer Time 4.08E E EQuench Oil Temp.8E004.8E BC.37E E BE.704E E CE BCE Residual E003 Cor Total The Model Fvalue of 6.70 implies the model is significant. There is only a.06% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 0. Adj RSquared C.V. % Pred RSquared PRESS Adeq Precision 7.75 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept E AFurnace Temp E E BHeating Time E CTransfer Time 5.0E E EQuench Oil Temp E E BC 9.58E E BE 4.E E CE E BCE E
111 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Std. Dev. = * A * B +5.0E003 * C E003 * E +9.58E003 * B * C +4.E003 * B * E * C * E * B * C * E Final Equation in Terms of Actual Factors: Std. Dev. = * Furnace Temp * Heating Time E003 * Transfer Time E003 * Quench Oil Temp E003 * Heating Time * Transfer Time E003 * Heating Time * Quench Oil Temp * Transfer Time * Quench Oil Temp * Heating Time * Transfer Time * Quench Oil Temp (d) Analyze the residuals from this experiment, and comment on your findings. The residual plot follows. All plots are satisfactory. Normal Plot of Residuals Residuals vs. Predicted 99 Normal % Probability Residuals Residual Predicted 85
112 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Furnace Temp Residuals vs. Heating Time Residuals Residuals Furnace Temp B:Heating Time Residuals vs. Transfer Time Residuals vs. Hold Down Time Residuals Residuals C:Transfer Time D:Hold Down Time Residuals vs. Quench Oil Temp Residuals E:Quench Oil Temp 86
113 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (e) Is this the best possible design for five factors in 6 runs? Specifically, can you find a fractional design for five factors in 6 runs with a higher resolution than this one? This was not the best design. A resolution V design is possible by setting the generator equal to the highest order interaction, ABCDE Consider the leaf spring experiment in Problem 8.5. Suppose that factor E (quench oil temperature) is very difficult to control during manufacturing. Where would you set factors A, B, C and D to reduce variability in the free height as much as possible regardless of the quench oil temperature used? DesignExpert Software.00 Std. Dev. Std. Dev C: Transfer Time X = B: Heating Time X = C: Transfer Time 0.00 Actual Factors A: Furnace Temp = D: Hold Down Time = 0.00 E: Quench Oil Temp = B: Heating Time To minimize the variability in the free height, run the process with A at the low level, B at the high level, C at the low level and D at either level (the low level of D may give a faster process) R.D. Snee ( Experimenting with a Large Number of Variables, in Experiments in Industry: Design, Analysis and Interpretation of Results, by R.D. Snee, L.B. Hare, and J.B. Trout, Editors, ASQC, 985) describes an experiment in which a 5 design with I=ABCDE was used to investigate the effects of five factors on the color of a chemical product. The factors are A = solvent/reactant, B = catalyst/reactant, C = temperature, D = reactant purity, and E = reactant ph. The results obtained were as follows: e = d = 6.79 a =.5 ade = 5.47 b = .68 bde = 3.45 abe =.66 abd = 5.68 c =.06 cde = 5. ace =. acd = 4.38 bce = .09 bcd = 4.30 abc =.93 abcde = 4.05 (a) Prepare a normal probability plot of the effects. Which effects seem active? Factors A, B, D, and the AB, AD interactions appear to be active. 87
114 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Color Normal plot A: Solvent/Reactant B: Catalyst/Reactant C: Temperature D: Reactant Purity E: Reactant ph Normal % probability B AD A AB D Effect Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Error C Model D Error E Model AB Error AC Model AD Error AE Error BC Error BD Error BE Error CD Error CE Error DE Lenth's ME Lenth's SME Design Expert Output Response: Color ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B D < AB AD Residual Cor Total The Model Fvalue of 4.53 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean.7 Adj RSquared C.V Pred RSquared PRESS.4 Adeq Precision
115 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ASolvent/Reactant BCatalyst/Reactant DReactant Purity AB AD Final Equation in Terms of Coded Factors: Color = * A * B +. * D * A * B * A * D Final Equation in Terms of Actual Factors: Color = * Solvent/Reactant * Catalyst/Reactant * Reactant Purity * Solvent/Reactant * Catalyst/Reactant * Solvent/Reactant * Reactant Purity (b) Calculate the residuals. Construct a normal probability plot of the residuals and plot the residuals versus the fitted values. Comment on the plots. Design Expert Output Diagnostics Case Statistics Standard Actual Predicted Student Cook's Outlier Run Order Value Value Residual Leverage Residual Distance t Order
116 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Norm al % probability Residuals Residual Predicted The residual plots are satisfactory. (c) If any factors are negligible, collapse the 5 design into a full factorial in the active factors. Comment on the resulting design, and interpret the results. The design becomes two replicates of a 3 in the factors A, B and D. When reanalyzing the data in three factors, D becomes labeled as C. Design Expert Output Response: Color ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C < AB AC BC ABC Residual Lack of Fit Pure Error Cor Total The Model Fvalue of 4.89 implies the model is significant. There is only a 0.05% chance that a "Model FValue" this large could occur due to noise. Std. Dev..0 RSquared Mean.7 Adj RSquared C.V Pred RSquared PRESS 3.7 Adeq Precision.736 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ASolvent/Reactant BCatalyst/Reactant CReactant Purity AB
117 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY AC BC ABC Final Equation in Terms of Coded Factors: Color = * A * B +. * C * A * B * A * C +0. * B * C 0. * A * B * C Final Equation in Terms of Actual Factors: Color = * Solvent/Reactant * Catalyst/Reactant * Reactant Purity * Solvent/Reactant * Catalyst/Reactant * Solvent/Reactant * Reactant Purity * Catalyst/Reactant * Reactant Purity * Solvent/Reactant * Catalyst/Reactant * Reactant Purity 8.8. Problem 6.3 describes a process improvement study in the manufacturing process of an integrated circuit. Suppose that only eight runs could be made in this process. Set up an appropriate 5 design and find the alias structure. Use the appropriate observations from Problem 6.3 as the observations in this design and estimate the factor effects. What conclusions can you draw? I = ABD = ACE = BCDE A (ABD) =BD A (ACE) =CE A (BCDE) =ABCDE A=BD=CE=ABCDE B (ABD) =AD B (ACE) =ABCE B (BCDE) =CDE B=AD=ABCE=CDE C (ABD) =ABCD C (ACE) =AE C (BCDE) =BDE C=ABCD=AE=BDE D (ABD) =AB D (ACE) =ACDE D (BCDE) =BCE D=AB=ACDE=BCE E (ABD) =ABDE E (ACE) =AC E (BCDE) =BCD E=ABDE=AC=BCD BC (ABD) =ACD BC (ACE) =ABE BC (BCDE) =DE BC=ACD=ABE=DE BE (ABD) =ADE BE (ACE) =ABC BE (BCDE) =CD BE=ADE=ABC=CD A B C D=AB E=AC de a be abd cd ace bc abcde 63 Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Error E Error BC Error BE
118 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Lenth's ME 8.3 Lenth's SME Half Normal plot 99 Half Normal % probability D C A B Effect The main A, B, C, and D are large. However, recall that we are really estimating A+BD+CE, B+AD, C+DE and D+AD. There are other possible interpretations of the experiment because of the aliasing. Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C D Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.7% chance that a "Model FValue" this large could occur due to noise. Std. Dev..73 RSquared 0.99 Mean Adj RSquared C.V Pred RSquared PRESS 59. Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AAperture BExposure Time CDevelop Time DMask Dimension Final Equation in Terms of Coded Factors: Yield = * A * B * C 8
119 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY * D Final Equation in Terms of Actual Factors: Aperture small Mask Dimension Small Yield = * Exposure Time * Develop Time Aperture large Mask Dimension Small Yield = * Exposure Time * Develop Time Aperture small Mask Dimension Large Yield = * Exposure Time * Develop Time Aperture large Mask Dimension Large Yield = * Exposure Time * Develop Time 8.9. Continuation of Problem 8.8. Suppose you have made the eight runs in the 5 design in Problem 8.8. What additional runs would be required to identify the factor effects that are of interest? What are the alias relationships in the combined design? We could fold over the original design by changing the signs on the generators D = AB and E = AC to produce the following new experiment. A B C D=AB E=AC () ade bd abe ce acd bcde abc 60 A (ABD) =BD A (ACE) =CE A (BCDE) =ABCDE A=BD=CE=ABCDE B (ABD) =AD B (ACE) =ABCE B (BCDE) =CDE B=AD=ABCE=CDE C (ABD) =ABCD C (ACE) =AE C (BCDE) =BDE C=ABCD=AE=BDE D (ABD) =AB D (ACE) =ACDE D (BCDE) =BCE D=AB=ACDE=BCE E (ABD) =ABDE E (ACE) =AC E (BCDE) =BCD E=ABDE=AC=BCD BC (ABD) =ACD BC (ACE) =ABE BC (BCDE) =DE BC=ACD=ABE=DE BE (ABD) =ADE BE (ACE) =ABC BE (BCDE) =CD BE=ADE=ABC=CD Assuming all three factor and higher interactions to be negligible, all main effects can be separated from their twofactor interaction aliases in the combined design. 83
120 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 8.0. Analyze the data in Problem 6.8 as if it came from a IV design with I = ABCD. Project the design into a full factorial in the subset of the original four factors that appear to be significant. 4 Run Yield Factor Levels Number A B C D=ABC (lbs) Low () High (+) () A (h) ad 5 B (%) bd 3 C (psi) ab 6 D (ºC) cd ac bc abcd 3 Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Error B Model C Model D Error AB Model AC Model AD Lenth's ME.74 Lenth's SME Half Normal plot 99 Half Normal % probability C A AD D AC Effect Design Expert Output Response: Yield in lbs ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A C D AC AD
121 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residual Cor Total The Model Fvalue of 48.5 implies the model is significant. There is only a.03% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 7.88 Adj RSquared C.V. 4.4 Pred RSquared PRESS 0.00 Adeq Precision 7.89 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ATime CPressure DTemperature AC AD Final Equation in Terms of Coded Factors: Yield = * A +.37 * C +.3 * D .3 * A * C +.3 * A * D Final Equation in Terms of Actual Factors: Yield = * Time * Pressure * Temperature * Time * Pressure * Time * Temperature 8.. Repeat Problem 8.0 using I = ABCD. Does use of the alternate fraction change your interpretation of the data? Run Yield Factor Levels Number A B C D=ABC (lbs) Low () High (+) d 0 A (h) a 8 B (%) b 3 C (psi) abd 4 D (ºC) c acd bcd abc 5 Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Error B Model C Model D Error AB Model AC Model AD Lenth's ME
122 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Lenth's SME Half Normal plot 99 Half Normal % probability C D AD AC A Effect Design Expert Output Response: Yield in lbs ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A C D AC AD Residual.5. Cor Total The Model Fvalue of 3.58 implies the model is significant. There is only a 4.% chance that a "Model FValue" this large could occur due to noise. Std. Dev..06 RSquared Mean 6.88 Adj RSquared C.V. 6.9 Pred RSquared PRESS Adeq Precision 4.45 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ATime CPressure DTemperature AC AD Final Equation in Terms of Coded Factors: Yield = * A * C +.3 * D .3 * A * C +.88 * A * D 86
123 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Yield = * Time * Pressure * Temperature * Time * Pressure * Time * Temperature 8.. Consider the 5 design in Problem 6.3. Suppose that only a onehalf fraction could be run. Furthermore, two days were required to take the 6 observations, and it was necessary to confound the 5 design in two blocks. Construct the design and analyze the data. A B C D E=ABCD Data Blocks = AB Block e a b abe c ace bce abc d ade bde abd cde acd bcd abcde 63 + Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Error D Error E Error AB Aliased Error AC Error AD Error AE Error BC Error BD Error BE Error CD Error CE Error DE Lenth's ME.4663 Lenth's SME The AB interaction in the above table is aliased with the threefactor interaction BCD, and is also confounded with blocks. 87
124 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Half Normal plot 99 Half Normal % probability A C B Effect Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Block Model < significant A < B < C < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..70 RSquared Mean Adj RSquared C.V Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept Block 3.56 Block AAperture BExposure Time CDevelop Time Final Equation in Terms of Coded Factors: Yield = * A +6.8 * B +5.3 * C Final Equation in Terms of Actual Factors: Aperture small Yield =
125 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY * Exposure Time * Develop Time Aperture large Yield = * Exposure Time * Develop Time 8.3. Construct a 7 design by choosing two fourfactor interactions as the independent generators. Write down the complete alias structure for this design. Outline the analysis of variance table. What is the resolution of this design? I=CDEF=ABCG=ABDEFG, Resolution IV A B C D E F=CDE G=ABC () ag bg ab cfg acf bcf abcfg df adfg bdfg abdf cdg acd bcd abcdg ef aefg befg abef ceg ace bce abceg de adeg bdeg abde cdefg acdef bcdef abcdefg Alias Structure A(CDEF)= ACDEF A(ABCG)= BCG A(ABDEFG)= BDEFG A=ACDEF=BCG=BDEFG B(CDEF)= BCDEF B(ABCG)= ACG B(ABDEFG)= ADEFG B=BCDEF=ACG=ADEFG C(CDEF)= DEF C(ABCG)= ABG C(ABDEFG)= ABCDEFG C=DEF=ABG=ABCDEFG D(CDEF)= CEF D(ABCG)= ABCDG D(ABDEFG)= ABEFG D=CEF=ABCDG=ABEFG E(CDEF)= CDF E(ABCG)= ABCEG E(ABDEFG)= ABDFG E=CDF=ABCEG=ABDFG F(CDEF)= CDE F(ABCG)= ABCFG F(ABDEFG)= ABDEG F=CDE=ABCFG=ABDEG G(CDEF)= CDEFG G(ABCG)= ABC G(ABDEFG)= ABDEF G=CDEFG=ABC=ABDEF AB(CDEF)= ABCDEF AB(ABCG)= CG AB(ABDEFG)= DEFG AB=ABCDEF=CG=DEFG AC(CDEF)= ADEF AC(ABCG)= BG AC(ABDEFG)= BCDEFG AC=ADEF=BG=BCDEFG AD(CDEF)= ACEF AD(ABCG)= BCDG AD(ABDEFG)= BEFG AD=ACEF=BCDG=BEFG AE(CDEF)= ACDF AE(ABCG)= BCEG AE(ABDEFG)= BDFG AE=ACDF=BCEG=BDFG AF(CDEF)= ACDE AF(ABCG)= BCFG AF(ABDEFG)= BDEG AF=ACDE=BCFG=BDEG AG(CDEF)= ACDEFG AG(ABCG)= BC AG(ABDEFG)= BDEF AG=ACDEFG=BC=BDEF BD(CDEF)= BCEF BD(ABCG)= ACDG BD(ABDEFG)= AEFG BD=BCEF=ACDG=AEFG 89
126 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY BE(CDEF)= BCDF BE(ABCG)= ACEG BE(ABDEFG)= ADFG BE=BCDF=ACEG=ADFG BF(CDEF)= BCDE BF(ABCG)= ACFG BF(ABDEFG)= ADEG BF=BCDE=ACFG=ADEG CD(CDEF)= EF CD(ABCG)= ABDG CD(ABDEFG)= ABCEFG CD=EF=ABDG=ABCEFG CE(CDEF)= DF CE(ABCG)= ABEG CE(ABDEFG)= ABCDFG CE=DF=ABEG=ABCDFG CF(CDEF)= DE CF(ABCG)= ABFG CF(ABDEFG)= ABCDEG CF=DE=ABFG=ABCDEG DG(CDEF)= CEFG DG(ABCG)= ABCD DG(ABDEFG)= ABEF DG=CEFG=ABCD=ABEF EG(CDEF)= CDFG EG(ABCG)= ABCE EG(ABDEFG)= ABDF EG=CDFG=ABCE=ABDF FG(CDEF)= CDEG FG(ABCG)= ABCF FG(ABDEFG)= ABDE FG=CDEG=ABCF=ABDE Analysis of Variance Table Source Degrees of Freedom A B C D E F G AB=CG AC=BG AD AE AF AG=BC BD BE BF CD=EF CE=DF CF=DE DG EG FG Error 9 Total Construct a design is performed. The 5 III design is shown below. 5 III design. Determine the effects that may be estimated if a full fold over of this The design with the fold over included is as follows. A B C D=AB E=AC de a be abd cd ace bc abcde Block A B C D=AB E=AC de a be abd cd 830
127 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The effects are shown in the table below ace bc abcde abc bcde acd ce abe bd ade () Principal Fraction A=A+BD+CE l B= l B + AD C l = C + AE D l = D + AB E= l E + AC BC= l BC + DE BE l = BE + CD Second Fraction l* A=ABDCE l* B= B  AD l* C = C  AE l* D = D  AB l* E= E  AC l* BC= BC + DE l* BE = BE + CD By combining the two fractions we can estimate the following: ( l i +l* i)/ ( i l l* i)/ A BD+CE B AD C AE D AB E AC BC+DE BE+CD These estimates are confirmed with the Design Expert output shown below. Design Expert Output Design Matrix Evaluation for Factorial Reduced 3FI Model Factorial Effects Aliases [Est. Terms] Aliased Terms [Intercept] = Intercept [Block ] = Block + ABD + ACE [Block ] = Block  ABD  ACE [A] = A [B] = B + CDE [C] = C + BDE [D] = D + BCE [E] = E + BCD [AB] = AB [AC] = AC [AD] = AD [AE] = AE [BC] = BC + DE [BD] = BD + CE [BE] = BE + CD [ABC] = ABC + ADE [ABE] = ABE + ACD 83
128 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 8.5. Construct a 63 III design. Determine the effects that may be estimated if a second fraction of this design is run with all signs reversed. The 6 3 III design is shown below. The effects are shown in the table below. A B C D=AB E=AC F=BC def af be abd cd ace bcf abcdef Principal Fraction A=A+BD+CE l B l =B+AD+CF C l =C+AE+BF D=D+AB+EF l E=E+AC+DF l F l =F+BC+DE BE l =BE+CD+AF Second Fraction l* A=ABDCE l* B =BADCF l* C =CAEBF l* D=DABEF l* E=EACDF l* F =FBCDE l* BE =BE+CD+AF By combining the two fractions we can estimate the following: ( l i +l* i )/ ( i l l* i )/ A BD+CE B AD+CF C AE+BF D AB+EF E AC+DF F BC+DE BE+CD+AF 8.6. Project the 4 IV design in Example 8. into two replicates of a design in the factors A and B. Analyze the data and draw conclusions. The Design Expert output below does not identify a significant effect. Design Expert Output Response: Filtration Rate ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B E
129 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY AB E Residual Lack of Fit Pure Error Cor Total The "Model Fvalue" of 0.4 implies the model is not significant relative to the noise. There is a 75.3 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev. 4.0 RSquared 0.37 Mean Adj RSquared C.V. 34. Pred RSquared PRESS Adeq Precision.98 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ATemperature BPressure AB Final Equation in Terms of Coded Factors: Filtration Rate = * A * B * A * B Final Equation in Terms of Actual Factors: Filtration Rate = * Temperature * Pressure * Temperature * Pressure 8.7. Fold over a 5 III design to produce a sixfactor design. Verify that the resulting design is a design. Compare this 6 IV design to the in Table IV F A B C D=AB E=BC Original Design Second Set of Runs w/ all Signs Switched If we relabel the factors from left to right as A, B, C, D, E, F, then this design becomes IV with generators I=ABDF and I=BCEF. It is not a minimal design, since k=(6)= runs, and the design contains 6 runs. 833
130 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 8.8. Consider the 63 III design in Problem 8.5. Determine the effects that may be estimated if a single factor fold over of this design is run with the signs for factor A reversed. Principal Fraction A l =A+BD+CE B=B+AD+CF l C=C+AE+BF l D l =D+AB+EF E l =E+AC+DF F=F+BC+DE l BE=BE+CD+AF l Second Fraction l* A =A+BD+CE l* B=BAD+CF l* C=CAE+BF l* D =DAB+EF l* E =EAC+DF l* F=F+BC+DE l* BE=BE+CDAF By combining the two fractions we can estimate the following: ( l i l* I)/ ( l i +l* I)/ A BD+CE AD B+CF AE C+BF AB D+EF AC E+DF F+BC+DE AF 8.9. Fold over the 74 III design in Table 8.9 to produce a eightfactor design. Verify that the resulting design is a 8 4 IV design. Is this a minimal design? H A B C D=AB E=AC F=BC G=ABC Original Design Second Set of Runs w/ all Signs Switched After folding the original design over, we add a new factor H, and we have a design with generators 8 4 D=ABH, E=ACH, F=BCH, and G=ABC. This is a IV design. It is a minimal design, since it contains k=(8)=6 runs An industrial engineer is conducting an experiment using a Monte Carlo simulation model of an inventory system. The independent variables in her model are the order quantity (A), the reorder point (B), the setup cost (C), the backorder cost (D), and the carrying cost rate (E). The response variable is average 834
131 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY annual cost. To conserve computer time, she decides to investigate these factors using a III design with I = ABD and I = BCE. The results she obtains are de = 95, ae = 34, b = 58, abd = 90, cd = 9, ac = 87, bce = 55, and abcde = 85. (a) Verify that the treatment combinations given are correct. Estimate the effects, assuming threefactor and higher interactions are negligible. The treatment combinations given are correct. The effects are shown in the Design Expert output and the normal and half normal probability plot of effects identify factors A and B as important. A B C D=AB E=BC de ae b abd cd ac bce abcde 5 Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Error C Error D Error E Error AC Error AE Lenth's ME Lenth's SME DESIGNEXPERT Plot Avg Annual Cost Normal plot Half Normal plot A: Order Quantity B: Reorder Point C: Setup Cost D: Backorder Cost E: Carrying Cost Normal % probability B A Half Normal % probability B A Effect Effect (b) Suppose that a second fraction is added to the first, for example ade = 36, e = 93, ab = 87, bd = 53, acd = 39, c = 99, abce = 9, and bcde = 50. How was this second fraction obtained? Add this data to the original fraction, and estimate the effects. 835
132 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY This second fraction is formed by reversing the signs of factor A. A B C D=AB E=BC ade e ab bd acd c abce bcde Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Error C Error D Error E Error AB Error AC Error AD Error AE Error BD Error CD Error DE Error ACD Error ADE Lenth's ME 5.88 Lenth's SME DESIGNEXPERT Plot Avg Annual Cost Normal plot Half Normal plot A: Order Quantity B: Reorder Point C: Setup Cost D: Backorder Cost E: Carrying Cost Normal % probability A B Half Normal % probability A B Effect Effect (c) Suppose that the fraction abc = 89, ce = 96, bcd = 54, acde = 35, abe = 93, bde = 5, ad = 37, and () = 98 was run. How was this fraction obtained? Add this data to the original fraction and estimate the effects. This second fraction is formed by reversing the signs of all factors. 836
133 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A B C D=AB E=BC abc bcd acde ce abe bde ad () Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Error C Error D Error E Error AB Error AC Error AD Error AE Error BC Error BD Error BE Error ABC Error ABE Lenth's ME Lenth's SME DESIGNEXPERT Plot Avg Annual Cost Normal plot Half Normal plot A: Order Quantity B: Reorder Point C: Setup Cost D: Backorder Cost E: Carrying Cost Normal % probability A B Half Normal % probability A B Effect Effect 8.. Construct a 5 design. Show how the design may be run in two blocks of eight observations each. Are any main effects or twofactor interactions confounded with blocks? 837
134 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A B C D E=ABCD Blocks = AB Block e a b abe c ace bce abc d ade bde abd cde acd bcd abcde + Blocks are confounded with AB and CDE. 8.. Construct a 7 design. Show how the design may be run in four blocks of eight observations each. Are any main effects or twofactor interactions confounded with blocks? A B C D E F=CDE G=ABC Block=ACE Block=BFG Block assignment () ag bg ab cfg acf bcf abcfg df adfg bdfg abdf cdg acd bcd abcdg ef aefg befg abef ceg ace bce abceg de adeg bdeg abde cdefg acdef bcdef abcdefg Blocks are confounded with ACE, BFG, and ABCEFG. 838
135 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 8.3. Irregular fractions of the k [John (97)]. Consider a 4 design. We must estimate the four main effects and the six twofactor interactions, but the full 4 factorial cannot be run. The largest possible block contains runs. These runs can be obtained from the four onequarter fractions defined by I = AB = ACD = BCD by omitting the principal fraction. Show how the remaining three 4 fractions can be combined to estimate the required effects, assuming that threefactor and higher interactions are negligible. This design could be thought of as a threequarter fraction. The four 4 fractions are as follows: () I=+AB=+ACD=+BCD Runs: c,d,ab,abcd () I=+AB=ACD=BCD Runs: (), cd, abc, abd (3) I=AB=+ACD=BCD Runs: a, bc, bd, acd (4) I=AB=ACD=+BCD Runs: b, ac, ad, bcd If we do not run the principal fraction (), then we can combine the remaining 3 fractions to 3 onehalf fractions of the 4 as follows: Fraction : () + (3) implies I=BCD. This fraction estimates: A, AB, AC, and AD Fraction : () + (4) implies I=ACD. This fraction estimates: B, BC, BD, and AB Fraction 3: (3) + (4) implies I=AB. This fraction estimates: C, D, and CD In estimating these effects we assume that all threefactor and higher interactions are negligible. Note that AB is estimated in two of the onehalf fractions: and. We would average these quantities and obtain a single estimate of AB. John (97, pp. 663) discusses this design and shows that the estimates obtained above are also the least squares estimates. John also derives the variances and covariances of these estimators Harry and Judy PetersonNedry (two friends of the author) own a vineyard and winery in Newberg, Oregon. They grow several varieties of grapes and manufacture wine. Harry and Judy have used factorial designs for process and product development in the winemaking segment of their business. This problem describes the experiment conducted for their 985 Pinot Noir. Eight variables, shown in Table P8., were originally studied in this experiment: Table P8. Variable Low Level High Level A Pinot Noir Clone Pommard Wadenswil B Oak Type Allier Troncais C Age of Barrel Old New D Yeast/Skin Contact Champagne Montrachet E Stems None All F Barrel Toast Light Medium G Whole Cluster None 0% H Fermentation Temperature Low (75 F Max) High (9 F Max) 839
136 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Harry and Judy decided to use a 84 IV design with 6 runs. The wine was tastetested by a panel of experts on 8 March 986. Each expert ranked the 6 samples of wine tasted, with rank being the best. The design and tastetest panel results are shown in Table P8.3. Table P8.3 Run A B C D E F G H HPN JPN CAL DCM RGB ybar s (a) What are the alias relationships in the design selected by Harry and Judy? E = BCD, F = ACD, G = ABC, H = ABD Defining Contrast : I = BCDE = ACDF = ABEF = ABCG = ADEG = BDFG = CEFG = ABDH = ACEH = BCFH = DEFH = CDGH = BEGH = AFGH = ABCDEFGH Aliases: A = BCG = BDH = BEF = CDF = CEH = DEG = FGH B = ACG = ADH = AEF = CDE = CFH = DFG = EGH C = ABG = ADF = AEH = BDE = BFH = DGH = EFG D = ABH = ACF = AEG = BCE = BFG = CGH = EFH E = ABF = ACH = ADG = BCD = BGH = CFG = DFH F = ABE = ACD = AGH = BCH = BDG = CEG = DEH G = ABC = ADE = AFH = BDF = BEH = CDH = CEF H = ABD = ACE = AFG = BCF = BEG = CDG = DEF AB = CG = DH = EF AC = BG = DF = EH AD = BH = CF = EG AE = BF = CH = DG AF = BE = CD = GH AG = BC = DE = FH AH = BD = CE = FG (b) Use the average ranks ( y ) as a response variable. Analyze the data and draw conclusions. You will find it helpful to examine a normal probability plot of effect estimates. The effects list and normal probability plot of effects are shown below. Factors D, E, F, and G appear to be significant. Also note that the DF and FG interactions were chosen instead of AC and AH based on the alias structure shown above. Design Expert Output Term Effect SumSqr % Contribtn Require Intercept Error A
137 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Error B Error C Model D Model E Model F Model G Error H Error AB Ignore AC Aliased Error AD Error AE Error AF Error AG Ignore AH Aliased Model DF Model FG Lenth's ME Lenth's SME DESIGNEXPERT Plot Taste Av g Normal plot A: A B: B C: C D: D E: E F: F G: G H: H Normal %probabi li ty G DF FG E D F Effect Design Expert Output Response: Taste Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant D E F G DF FG Residual Cor Total The Model Fvalue of 5.55 implies the model is significant. There is only a.5% chance that a "Model FValue" this large could occur due to noise. Std. Dev..5 RSquared Mean 8.50 Adj RSquared C.V Pred RSquared PRESS Adeq Precision 7.83 Coefficient Standard 95% CI 95% CI 84
138 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Factor Estimate DF Error Low High VIF Intercept DD EE FF GG DF FG Final Equation in Terms of Coded Factors: Taste Avg = * D +.0 * E .00 * F +.57 * G +.30 * D * F +.3 * F * G Factors D and G and are important. Factor E and the DF and FG interactions are moderately important and were included in the model because the PRESS statistic showed improvement with their inclusion. Factor F is added to the model to preserve hierarchy. As stated earlier, the interactions are aliased with other twofactor interactions that could also be important. So the interpretation of the twofactor interaction is somewhat uncertain. Normally, we would add runs to the design to isolate the significant interactions, but that will not work very well here because each experiment requires a full growing season. In other words, it would require a very long time to add runs to dealias the alias chains of interest. DESIGNEXPERT Plot Taste Av g 5. Interaction Graph F: F DESIGNEXPERT Plot Taste Av g 5. Interaction Graph G: G X = D: D Y = F: F X = F: F Y = G: G Design Points.3847 Design Points.8847 F F F F Actual Factors A: A = A B: B = B C: C = C E: E = E G: G = G H: H = H Taste Avg G G G G Actual Factors A: A = A B: B = B C: C = C D: D = D E: E = E H: H = H Taste Avg D D F F D: D F: F 84
139 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY One Factor Plot Taste Avg E E E: E (c) Use the standard deviation of the ranks (or some appropriate transformation such as log s) as a response variable. What conclusions can you draw about the effects of the eight variables on variability in wine quality? DESIGNEXPERT Plot Ln(Taste StDev) Normal plot A: A B: B C: C D: D E: E F: F G: G H: H Normal %probabi li ty Warning! No terms are selected Effect There do not appear to be any significant factors. (d) After looking at the results, Harry and Judy decide that one of the panel members (DCM) knows more about beer than he does about wine, so they decide to delete his ranking. What affect would this have on the results and on conclusions from parts (b) and (c)? Design Expert Output Term Effect SumSqr % Contribtn Require Intercept Error A Error B Error C Model D
140 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Model E Model F Model G Error H Error AB Ignore AC Aliased Error AD Error AE Error AF Error AG Ignore AH Aliased Model DF Model FG Lenth's ME Lenth's SME.705 DESIGNEXPERT Plot Taste Av g Half Normal plot A: A B: B C: C D: D E: E F: F G: G H: H Half Normal %probability E DF F FG G D Effect Design Expert Output Response: Taste Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant D E F G DF FG Residual Cor Total The Model Fvalue of 5.48 implies the model is significant. There is only a.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..50 RSquared Mean 8.50 Adj RSquared C.V Pred RSquared PRESS 78.0 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept
141 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DD EE FF GG DF FG Final Equation in Terms of Coded Factors: Taste Avg = * D +. * E .9 * F +.47 * G +.9 * D * F +.6 * F * G The results are very similar for average taste without DCM as they were with DCM. DESIGNEXPERT Plot Ln(Taste StDev) Normal plot A: A B: B C: C D: D E: E F: F G: G H: H Normal %probabi li ty Warning! No terms are selected Effect The standard deviation response is much the same with or without DCM s responses. Again, there are no significant factors. (e) Suppose that just before the start of the experiment, Harry and Judy discovered that the eight new barrels they ordered from France for use in the experiment would not arrive in time, and all 6 runs would have to be made with old barrels. If Harry and Judy just drop column C from their design, what does this do to the alias relationships? Do they need to start over and construct a new design? The resulting design is a 7 3 IV BEGH = AFGH. with defining relations: I = ABEF = ADEG = BDFG = ABDH = DEFH = (f) Harry and Judy know from experience that some treatment combinations are unlikely to produce good results. For example, the run with all eight variables at the high level generally results in a poorly rated wine. This was confirmed in the 8 March 986 taste test. They want to set up a new design for their 986 Pinot Noir using these same eight variables, but they do not want to make the run with all eight factors at the high level. What design would you suggest? 845
142 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY By changing the sign of any of the design generators, a design that does not include the principal fraction will be generated. This will give a design without an experimental run combination with all of the variables at the high level A 6run fractional factorial experiment in 0 factors on sandcasting of engine manifolds was conducted by engineers at the Essex Aluminum Plant of the Ford Motor Company and described in the article Evaporative Cast Process 3.0 Liter Intake Manifold Poor Sandfill Study, by D. Becknell (Fourth Symposium on Taguchi Methods, American Supplier Institute, Dearborn, MI, 986, pp. 030). The purpose was to determine which of 0 factors has an effect on the proportion of defective castings. The design and the resulting proportion of nondefective castings p observed on each run are shown in Table P8.4. This is a resolution III fraction with generators E=CD, F=BD, G=BC, H=AC, J=AB, and K=ABC. Assume that the number of castings made at each run in the design is 000. Table P8.4 F&T s Run A B C D E F G H J K p arcsin Modification (a) Find the defining relation and the alias relationships in this design. I=CDE=BDF=BCG=ACH=ABJ=ABCK=BCEF=BDEG=ADEH=ABCDEJ=ABDEK=CDFG=ABCDFH = ADFJ=ACDFK=ABGH=ACGJ=AGK=BCHJ=BHK=CjK (b) Estimate the factor effects and use a normal probability plot to tentatively identify the important factors. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B Error C Error D Error E Model F Error G Error H Error J Model K Error AB Aliased Error AC Aliased Error AD E Error AE Error AF Error BE Error DK Lenth's ME
143 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Lenth's SME DESIGNEXPERT Plot p Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J K: K Normal % probability K F Effect (c) Fit an appropriate model using the factors identified in part (b) above. Design Expert Output Response: p ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant F K Residual E003 Cor Total The Model Fvalue of 0.7 implies the model is significant. There is only a 0.8% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared 0.65 Mean 0.89 Adj RSquared C.V Pred RSquared 0.48 PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept FF KK Final Equation in Terms of Coded Factors: p = * F * K Final Equation in Terms of Actual Factors: p = * F * K 847
144 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (d) Plot the residuals from this model versus the predicted proportion of nondefective castings. Also prepare a normal probability plot of the residuals. Comment on the adequacy of these plots. The residual versus predicted plot identifies an inequality of variances. This is likely caused by the response variable being a proportion. A transformation could be used to correct this. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted (e) In part (d) you should have noticed an indication that the variance of the response is not constant (considering that the response is a proportion, you should have expected this). The previous table also shows a transformation on ˆp, the arcsin square root, that is a widely used variance stabilizing transformation for proportion data (refer to the discussion of variance stabilizing transformations is Chapter 3). Repeat parts (a) through (d) above using the transformed response and comment on your results. Specifically, are the residuals plots improved? Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B E E005 Error C Error D Error E Model F Error G Error H Error J Model K Error AD Error AF Error BE Error DH Error DK Lenth's ME Lenth's SME As with the original analysis, factors F and K remain significant with a slight increase with the R. 848
145 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot arcsin Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J K: K Norm al % probability F K Effect Design Expert Output Response: arcsin ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant F K Residual Cor Total The Model Fvalue of.59 implies the model is significant. There is only a 0.09% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0. RSquared Mean.8 Adj RSquared C.V Pred RSquared PRESS 0.4 Adeq Precision 8.93 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept FF KK Final Equation in Terms of Coded Factors: arcsin = * F * K Final Equation in Terms of Actual Factors: arcsin = * F * K The inequality of variance has improved; however, there remain hints of inequality in the residuals versus predicted plot and the normal probability plot now appears to be irregular. 849
146 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Norm al % probability Residuals Residual Predicted (f) There is a modification to the arcsin square root transformation, proposed by Freeman and Tukey ( Transformations Related to the Angular and the Square Root, Annals of Mathematical Statistics, Vol., 950, pp ) that improves its performance in the tails. F&T s modification is: arcsin npˆ arcsin n npˆ n Rework parts (a) through (d) using this transformation and comment on the results. (For an interesting discussion and analysis of this experiment, refer to Analysis of Factorial Experiments with Defects or Defectives as the Response, by S. Bisgaard and H.T. Fuller, Quality Engineering, Vol. 7, 9945, pp ) Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B E E005 Error C Error D Error E Model F Error G Error H Error J Model K Error AD Error AF Error BE Error DH Error DK Lenth's ME Lenth's SME As with the prior analysis, factors F and K remain significant. 850
147 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot F&T Transform Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J K: K Norm al % probability F K Effect Design Expert Output Response: F&T Transform ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant F K Residual Cor Total The Model Fvalue of.67 implies the model is significant. There is only a 0.09% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0. RSquared Mean.7 Adj RSquared C.V Pred RSquared PRESS 0.3 Adeq Precision 8. Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept FF KK Final Equation in Terms of Coded Factors: F&T Transform = * F * K Final Equation in Terms of Actual Factors: F&T Transform = * F * K The residual plots appear as they did with the arcsin square root transformation. 85
148 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Norm al % probability Residuals Residual Predicted 8.6. A 6run fractional factorial experiment in nine factors was conducted by Chrysler Motors Engineering and described in the article Sheet Molded Compound Process Improvement, by P.I. Hsieh and D.E. Goodwin (Fourth Symposium on Taguchi Methods, American Supplier Institute, Dearborn, MI, 986, pp. 3). The purpose was to reduce the number of defects in the finish of sheetmolded grill opening panels. The design, and the resulting number of defects, c, observed on each run, is shown in Table P8.5. This is a resolution III fraction with generators E=BD, F=BCD, G=AC, H=ACD, and J=AB. Table P8.5 Run A B C D E F G H J c c F&T s Modification (a) Find the defining relation and the alias relationships in this design. I = ABJ = ACG = BDE = CEF = DGH = FHJ = ABFH = ACDH = ADEJ = AEFG = BCDF = BCGJ = BEGH = CEHJ = DFGJ = ABCEH = ABDFG = ACDFJ = ADEFH = AEGHJ = BCDHJ = BCFGH = BEFGJ = CDEGJ = ABCDEG = ABCEFJ = ABDGHJ = ACFGHJ = BDEFHJ = CDEFGH = ABCDEFGHJ (b) Estimate the factor effects and use a normal probability plot to tentatively identify the important factors. 85
149 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The effects are shown below in the Design Expert output. The normal probability plot of effects identifies factors A, D, F, and interactions AD, AF, BC, BG as important. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Model C Model D Error E Model F Model G Error H Error J Model AD Error AE Model AF Error AH Model BC Model BG Lenth's ME Lenth's SME DESIGNEXPERT Plot c Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J Normal % probability F BG D A AD BC AF Effect (c) Fit an appropriate model using the factors identified in part (b) above. The analysis of variance and corresponding model is shown below. Factors B, C, and G are included for hierarchal purposes. Design Expert Output Response: c ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C D F G
150 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY AD AF BC BG Residual Cor Total The Model Fvalue of 8.6 implies the model is significant. There is only a 0.09% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 0.06 Adj RSquared C.V Pred RSquared 0.80 PRESS Adeq Precision 7.77 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AA BB CC DD FF GG AD AF BC BG Final Equation in Terms of Coded Factors: c = * A * B .8 * C 7.9 * D 8.3 * F .06 * G +5.8 * A * D * A * F * B * C 6.3 * B * G Final Equation in Terms of Actual Factors: c = * A * B * C * D * F * G * A * D * A * F * B * C * B * G (d) Plot the residuals from this model versus the predicted number of defects. Also, prepare a normal probability plot of the residuals. Comment on the adequacy of these plots. 854
151 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Norm al % probability Residuals Residual Predicted There is a significant problem with inequality of variance. This is likely caused by the response variable being a count. A transformation may be appropriate. (e) In part (d) you should have noticed an indication that the variance of the response is not constant (considering that the response is a count, you should have expected this). The previous table also shows a transformation on c, the square root, that is a widely used variance stabilizing transformation for count data (refer to the discussion of variance stabilizing transformations in Chapter 3). Repeat parts (a) through (d) using the transformed response and comment on your results. Specifically, are the residual plots improved? Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Model B Error C Model D Error E Model F Model G Error H Error J Error AD Error AE Error AF Error AH Error BC Model BG Lenth's ME.7978 Lenth's SME The analysis of the data with the square root transformation identifies only D, F, the BG interaction as being significant. The original analysis identified factor A and several two factor interactions as being significant. 855
152 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot sq rt Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J Norm al % probability F D BG Effect Design Expert Output Response: sqrt ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant B D F G BG Residual Cor Total The Model Fvalue of 6.67 implies the model is significant. There is only a 0.56% chance that a "Model FValue" this large could occur due to noise. Std. Dev..3 RSquared Mean.3 Adj RSquared C.V Pred RSquared PRESS Adeq Precision 8.4 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept BB DD FF GG BG Final Equation in Terms of Coded Factors: sqrt = * B .08 * D .30 * F 0.9 * G * B * G Final Equation in Terms of Actual Factors: 856
153 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY sqrt = * B * D * F * G * B * G The residual plots are acceptable; although, there appears to be a slight u shape to the residuals versus predicted plot. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted (f) There is a modification to the square root transformation proposed by Freeman and Tukey ( Transformations Related to the Angular and the Square Root, Annals of Mathematical Statistics, Vol., 950, pp ) that improves its performance. F&T s modification to the square root transformation is: c c Rework parts (a) through (d) using this transformation and comment on the results. (For an interesting discussion and analysis of this experiment, refer to Analysis of Factorial Experiments with Defects or Defectives as the Response, by S. Bisgaard and H.T. Fuller, Quality Engineering, Vol. 7, 9945, pp ) Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Model B Error C Model D Error E Model F Model G Error H Error J Error AD Error AE Error AF
154 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Error AH Error BC Model BG Lenth's ME.400 Lenth's SME As with the square root transformation, factors D, F, and the BG interaction remain significant. DESIGNEXPERT Plot F&T Normal plot A: A B: B C: C D: D E: E F: F G: G H: H J: J Normal % probability F D BG Effect Design Expert Output Response: F&T ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant B D F G BG Residual Cor Total The Model Fvalue of 5.73 implies the model is significant. There is only a 0.95% chance that a "Model FValue" this large could occur due to noise. Std. Dev..3 RSquared 0.74 Mean.5 Adj RSquared 0.67 C.V Pred RSquared PRESS 44.7 Adeq Precision 7.86 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept BB DD FF GG BG Final Equation in Terms of Coded Factors: F&T = 858
155 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY * B .00 * D . * F 0.0 * G * B * G Final Equation in Terms of Actual Factors: F&T = * B * D .50 * F * G * B * G The following interaction plots appear as they did with the square root transformation; a slight u shape is observed in the residuals versus predicted plot. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 8.7. A spin coater is used to apply photoresist to a bare silicon wafer. This operation usually occurs early in the semiconductor manufacturing process, and the average coating thickness and the variability in the coating thickness has an important impact on downstream manufacturing steps. Six variables are used in the experiment. The variables and their high and low levels are as follows: Factor Low Level High Level Final Spin Speed 7350 rpm 6650 rpm Acceleration Rate 5 0 Volume of Resist Applied 3 cc 5 cc Time of Spin 4 s 6 s Resist Batch Variation Batch Batch Exhaust Pressure Cover Off Cover On The experimenter decides to use a 6 design and to make three readings on resist thickness on each test wafer. The data are shown in table P8.6. Table P
156 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A B C D E F Resist Thick ness Run Volume Batch Time Speed Acc. Cover Left Center Right Avg. Range Off Off Off Off Off Off On Off Off On On Off On On On On On Off Off Off On On Off On On Off On Off On On On Off (a) Verify that this is a 6 design. Discuss the alias relationships in this design. I=ABCDEF. This is a resolution VI design where main effects are aliased with fivefactor interactions and twofactor interactions are aliased with fourfactor interactions. (b) What factors appear to affect average resist thickness? Factors B, D, and E appear to affect the average resist thickness. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Model B Error C Model D Model E Error F Error AB Error AC Error AD Error AE Error AF Error BC Error BD Error BE Error BF Error CD Error CE Error CF
157 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Error DE Error DF Error EF Error ABC Error ABD Aliased Error ABE Error ABF Error ACD Error ACE Aliased Error ACF Aliased Error ADE Error ADF Aliased Error AEF Aliased Error BCD Error BCE Error BCF Aliased Error BDE Error BDF Error BEF Aliased Error CDE Aliased Error CDF Aliased Error CEF Error DEF Aliased Lenth's ME Lenth's SME DESIGNEXPERT Plot Thick Avg Normal plot A: Volume B: Batch C: Time D: Speed E: Acc F: Cover Normal % probability E B D Effect Design Expert Output Response: Thick Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model 6.540E E < significant B D 3.430E E < E.677E E < Residual Cor Total 7.3E The Model Fvalue of 79. implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared
158 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY C.V..8 Pred RSquared PRESS.006E+005 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept BBatch DSpeed EAcc Final Equation in Terms of Coded Factors: Thick Avg = * B * D * E Final Equation in Terms of Actual Factors: Batch Batch Thick Avg = * Speed * Acc Batch Batch Thick Avg = * Speed * Acc (c) Since the volume of resist applied has little effect on average thickness, does this have any important practical implications for the process engineers? Yes, less material could be used. (d) Project this design into a smaller design involving only the significant factors. Graphically display the results. Does this aid in interpretation? Cube Graph Thick Avg D Speed E+ Acc D E B B+ Batch The cube plot usually assists the experimenter in drawing conclusions. 86
159 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (e) Use the range of resist thickness as a response variable. Is there any indication that any of these factors affect the variability in resist thickness? Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Model A Model B Error C Error D Model E Model F Model AB Error AC Error AD Error AE Model AF Error BC Error BD Error BE Model BF Error CD Error CE Error CF Error DE Error DF Model EF Error ABC Aliased Error ABD Aliased Error ABE Model ABF Error ACD Error ACE Aliased Error ACF Aliased Error ADE Error ADF Error AEF Error BCD Aliased Error BCE Aliased Error BCF Aliased Error BDE Error BDF Error BEF Aliased Error CDE Aliased Error CDF Aliased Error CEF Error DEF Lenth's ME Lenth's SME
160 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Range Normal plot A: Volume B: Batch C: Time D: Speed E: Acc F: Cover EF F ABF Normal %probabi li ty A BF B AF AB 0 5 E Effect Design Expert Output Response: Thick Range ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B E F AB AF BF EF ABF Residual Cor Total The Model Fvalue of.48 implies the model is significant. There is only a 4.00% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 6.56 Adj RSquared C.V Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AVolume BBatch EAcc FCover AB AF BF EF ABF Final Equation in Terms of Coded Factors: Thick Range =
161 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 0.3 * A +.06 * B .69 * E * F +0.3 * A * B * A * F +.63 * B * F * E * F * A * B * F Final Equation in Terms of Actual Factors: Batch Batch Cover Off Thick Range = * Volume * Acc Batch Batch Cover Off Thick Range = * Volume * Acc Batch Batch Cover On Thick Range = * Volume * Acc Batch Batch Cover On Thick Range = * Volume * Acc The model for thickness range is not very strong. Notice the small value of R, and in particular, the adjusted R. Often we find that obtaining a good model for a response that expresses variability is not as easy as finding a satisfactory model for a response that measures the mean. (f) Where would you recommend that the process engineers run the process? Considering only the average thickness results, the engineers could use factors B, D and E to put the process mean at target. Then the engineer could consider the other factors on the range model to try to set the factors to reduce the variation in thickness at that mean A 6run experiment was performed in a semiconductor manufacturing plant to study the effects of six factors on the curvature or camber of the substrate devices produced. The six variables and their levels are shown in Table P8.7. Table P8.7 Lamination Lamination Lamination Firing Firing Firing Temperature Time Pressure Temperature Cycle Time Dew Point Run (c) (s) (tn) (c) (h) (c)
162 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Each run was replicated four times, and a camber measurement was taken on the substrate. The data are shown in Table P8.8. Table P8.8 Camber for Replicate (in/in) Total Mean Standard Run 3 4 (04 in/in) (04 in/in) Deviation (a) What type of design did the experimenters use? The 6 IV, a 6run design. (b) What are the alias relationships in this design? The defining relation is I=ABCE=ACDF=BDEF. The aliases are shown below. A(ABCE)= BCE A(ACDF)= CDF A(BDEF)= ABDEF A=BCE=CDF=ABDEF B(ABCE)= ACE B(ACDF)= ABCDF B(BDEF)= DEF B=ACE=ABCDF=DEF C(ABCE)= ABE C(ACDF)= ADF C(BDEF)= BCDEF C=ABE=ADF=BCDEF D(ABCE)= ABCDE D(ACDF)= ACF D(BDEF)= BEF D=ABCDE=ACF=BEF E(ABCE)= ABC E(ACDF)= ACDEF E(BDEF)= BDF E=ABC=ACDEF=BDF F(ABCE)= ABCEF F(ACDF)= ACD F(BDEF)= BDE F=ABCEF=ACD=BDE AB(ABCE)= CE AB(ACDF)= BCDF AB(BDEF)= ADEF AB=CE=BCDF=ADEF AC(ABCE)= BE AC(ACDF)= DF AC(BDEF)= ABCDEF AC=BE=DF=ABCDEF AD(ABCE)= BCDE AD(ACDF)= CF AD(BDEF)= ABEF AD=BCDE=CF=ABEF AE(ABCE)= BC AE(ACDF)= CDEF AE(BDEF)= ABDF AE=BC=CDEF=ABDF AF(ABCE)= BCEF AF(ACDF)= CD AF(BDEF)= ABDE AF=BCEF=CD=ABDE BD(ABCE)= ACDE BD(ACDF)= ABCF BD(BDEF)= EF BD=ACDE=ABCF=EF BF(ABCE)= ACEF BF(ACDF)= ABCD BF(BDEF)= DE BF=ACEF=ABCD=DE (c) Do any of the process variables affect average camber? Yes, per the analysis below, variables A, C, E, and F affect average camber. Design Expert Output 866
163 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Term Effect SumSqr % Contribtn Model ALamination Temperature Error BLamination Time Model CLamination Pressure Error DFiring Temperature Model EFiring Cycle Time Model FFiring Dew Point Error AB Error AC Error AD Error AE Error AF Aliased BC Aliased Error BD Aliased BE Aliased Error BF Aliased CD Aliased Aliased CE Aliased Aliased CF Aliased Aliased DE Aliased Aliased DF Aliased Aliased EF Aliased Lenth's ME Lenth's SME DesignExpert Software Average Camber Normal Plot ShapiroWilk test Wvalue = pvalue = 0.30 A: Lamination Temperature B: Lamination Time C: Lamination Pressure D: Firing Temperature E: Firing Cycle Time F: Firing Dew Point Positive Effects Negative Effects Normal % Probability E A C 5 F Standardized Effect Design Expert Output Response Average Camber ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ALamination Temperature CLamination Pressure EFiring Cycle Time FFiring Dew Point Residual Cor Total The Model Fvalue of.43 implies the model is significant. There is only a 0.07% chance that a "Model FValue" this large could occur due to noise. 867
164 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Std. Dev. 3.8 RSquared Mean Adj RSquared C.V. % Pred RSquared PRESS Adeq Precision.499 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ALamination Temperature CLamination Pressure EFiring Cycle Time FFiring Dew Point Final Equation in Terms of Coded Factors: Average Camber = * A * C 7.0 * E * F Final Equation in Terms of Actual Factors: Average Camber = * Lamination Temperature * Lamination Pressure * Firing Cycle Time * Firing Dew Point (d) Do any of the process variables affect the variability in camber measurements? Yes, A, B, F, and AF interaction affect the variability in camber measurements. Design Expert Output Term Effect SumSqr % Contribtn Model ALamination Temperature Model BLamination Time Error CLamination Pressure Error DFiring Temperature Error EFiring Cycle Time Model FFiring Dew Point Error AB Error AC Error AD Error AE Model AF Aliased BC Aliased Error BD Aliased BE Aliased Error BF Aliased CD Aliased Aliased CE Aliased Aliased CF Aliased Aliased DE Aliased Aliased DF Aliased Aliased EF Aliased Lenth's ME Lenth's SME
165 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Average Camber Normal Plot ShapiroWilk test Wvalue = pvalue = 0.30 A: Lamination Temperature B: Lamination Time C: Lamination Pressure D: Firing Temperature E: Firing Cycle Time F: Firing Dew Point Positive Effects Negative Effects Normal % Probability E A C 5 F Standardized Effect Design Expert Output Response Std. Dev. ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ALamination Temperature BLamination Time FFiring Dew Point AF Residual Cor Total The Model Fvalue of.0 implies the model is significant. There is only a 0.08% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 8.5 RSquared Mean 5.34 Adj RSquared C.V. % 3.7 Pred RSquared PRESS Adeq Precision 9.5 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ALamination Temperature BLamination Time FFiring Dew Point AF Final Equation in Terms of Coded Factors: Std. Dev. = * A 8.7 * B 4.6 * F * A * F 869
166 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Std. Dev. = * Lamination Temperature * Lamination Time * Firing Dew Point * Lamination Temperature * Firing Dew Point (e) If it is important to reduce camber as much as possible, what recommendations would you make? One Factor One Factor Average Camber 30 Average Camber A: Lamination Temperature C: Lamination Pressure One Factor One Factor Average Camber 30 Average Camber E: Firing Cycle Time F: Firing Dew Point 870
167 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Std. Dev. F F X = A: Lamination Temperature X = F: Firing Dew Point Actual Factors B: Lamination Time = 5.00 C: Lamination Pressure = 5.00 D: Firing Temperature = E: Firing Cycle Time = Interaction F: Firing Dew Point DesignExpert Software Std. Dev. X = B: Lamination Time Actual Factors A: Lamination Temperature = C: Lamination Pressure = 5.00 D: Firing Temperature = E: Firing Cycle Time = 9.00 F: Firing Dew Point = One Factor Std. Dev Std. Dev A: Lamination Temperature B: Lamination Time Run A and C at the low level and E and F at the high level. B at the higher level enables a lower variation without affecting the average camber Carbon anodes used in a smelting process are baked in a ring furnace. An experiment is run in the furnace to determine which factors influence the weight of packing material that is stuck to the anodes after baking. Six variables are of interest, each at two levels: A = pitch/fines ratio (0.45, 0.55); B = packing material type (, ); C = packing material temperature (ambient, 35 C); D = flue location (inside, outside); E = pit temperature (ambient, 95 C); and F = delay time before packing (zero, 4 hours). A 63 design is run, and three replicates are obtained at each of the design points. The weight of packing material stuck to the anodes is measured in grams. The data in run order are as follows: abd = (984, 86, 936); abcdef = (75, 976, 457); be = (7, 0, 890); af = (474, 64, 54); def = (30, 56, 93); cd = (765, 705, 8); ace = (338, 54, 94); and bcf = (35, 99, 53). We wish to minimize the amount stuck packing material. (a) Verify that the eight runs correspond to a 6 3 III design. What is the alias structure? A B C D=AB E=AC F=BC def af be abd cd ace bcf abcdef I=ABD=ACE=BCF=BCDE=ACDF=ABEF=DEF, Resolution III A=BD=CE=CDF=BEF B=AD=CF=CDE=AEF C=AE=BF=BDE=ADF D=AB=EF=BCE=ACF E=AC=DF=BCD=ABF F=BC=DE=ACD=ABE CD=BE=AF=ABC=ADE=BDF=CEF 87
168 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (b) Use the average weight as a response. What factors appear to be influential? Design Expert Output Term Effect SumSqr % Contribtn Model A Error B Error C Model D Model E Model F Error AF Lenth's ME Lenth's SME Half Normal plot 99 Half Normal % probability E A F D Effect Factors A, D, E and F (and their aliases) are apparently important. (c) Use the range of the weights as a response. What factors appear to be influential? Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Error B Model C Error D Model E Model F Model AF Lenth's ME Lenth's SME Factors C, E, F and the AF interaction (and their aliases) appear to be large. 87
169 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Half Normal plot 99 Half Normal % probability C AF E F Effect (d) What recommendations would you make to the process engineers? It is not known exactly what to do here, since A, D, E and F are large effects, and because the design is resolution III, the main effects are aliased with twofactor interactions. Note, for example, that D is aliased with EF and the main effect could really be a EF interaction. If the main effects are really important, then setting all factors at the low level would minimize the amount of material stuck to the anodes. It would be necessary to run additional experiments to confirm these findings In an article in Quality Engineering ( An Application of Fractional Factorial Experimental Designs, 988, Vol. pp. 93) M.B. Kilgo describes an experiment to determine the effect of CO pressure (A), CO temperature (B), peanut moisture (C), CO flow rate (D), and peanut particle size (E) on the total yield of oil per batch of peanuts (y). The levels she used for these factors are shown in Table P8.9. She conducted the 6run fractional factorial experiment shown in Table P8.0. Table P8.9 A B C D E Coded Pressure Temp Moisture Flow Particle Size Level (bar) (C) (% by weight) (liters/min) (mm) Table P8.0 A B C D E y
170 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) What type of design has been used? Identify the defining relation and the alias relationships. A 5 V, 6run design, with I= ABCDE. A(ABCDE)= BCDE A= BCDE B(ABCDE)= ACDE B= ACDE C(ABCDE)= ABDE C= ABDE D(ABCDE)= ABCE D= ABCE E(ABCDE)= ABCD E= ABCD AB(ABCDE)= CDE AB= CDE AC(ABCDE)= BDE AC= BDE AD(ABCDE)= BCE AD= BCE AE(ABCDE)= BCD AE= BCD BC(ABCDE)= ADE BC= ADE BD(ABCDE)= ACE BD= ACE BE(ABCDE)= ACD BE= ACD CD(ABCDE)= ABE CD= ABE CE(ABCDE)= ABD CE= ABD DE(ABCDE)= ABC DE= ABC (b) Estimate the factor effects and use a normal probability plot to tentatively identify the important factors. Design Expert Output Term Effect SumSqr % Contribtn Model Intercept Error A Model B Error C Error D Model E Error AB Error AC Error AD Error AE Error BC Error BD Error BE Error CD Error CE Error DE Lenth's ME.5676 Lenth's SME
171 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Yield Half Normal plot A: Pressure B: Temperature C: Moisture D: Flow E: Particle Size Half Normal % probability B E Effect (c) Perform an appropriate statistical analysis to test the hypothesis that the factors identified in part above have a significant effect on the yield of peanut oil. Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant B E < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 8.4 RSquared Mean 54.5 Adj RSquared C.V. 5.8 Pred RSquared 0.87 PRESS Adeq Precision 8.07 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept BTemperature EParticle Size (d) Fit a model that could be used to predict peanut oil yield in terms of the factors that you have identified as important. Design Expert Output Final Equation in Terms of Coded Factors: Yield = * B +.5 * E Final Equation in Terms of Actual Factors: 875
172 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Yield = * Temperature * Particle Size (e) Analyze the residuals from this experiment and comment on model adequacy. The residual plots are satisfactory. There is a slight tendency for the variability of the residuals to increase with the predicted value of y. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Residuals vs. Temperature Residuals vs. Particle Size Residuals Residuals Temperature Particle Size 8.3. An article by L.B. Hare ( In the Soup: A Case Study to Identify Contributors to Filling Variability, Journal of Quality Technology, Vol. 0, pp ) describes a factorial experiment used to study the filling variability of dry soup mix packages. The factors are A = number of mixing ports through which the vegetable oil was added (, ), B = temperature surrounding the mixer (cooled, ambient), C = mixing time (60, 80 sec), D = batch weight (500, 000 lb), and E = number of days between mixing and packaging (,7). Between 5 and 50 packages of soup were sampled over an eight hour period for each run in the 876
173 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY design and the standard deviation of package weight was used as the response variable. The design and resulting data are shown in Table P8.. Table P8. Std Order A  Mixer Ports B  Temp C  Time D  Batch Weight E  Delay y  Std Dev (a) What is the generator for this design? The design generator is I = ABCDE. (b) What is the resolution of this design? This design is Resolution V. (c) Estimate the factor effects. Which effects are large? Design Expert Output Term Effect SumSqr % Contribtn Require Intercept Error AMixer Ports Model BTemp Error CTime Model DBatch Weight Model EDelay Error AB Error AC Error AD Error AE Error BC Error BD Model BE Error CD Error CE Model DE Lenth's ME Lenth's SME
174 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Factor E and the two factor interactions BE and DE appear to be significant. Factors B, and D, are included to satisfy model hierarchy. The analysis of variance and model can be found in the Design Expert Output below. DesignExpert Software Std. Dev. HalfNormal Plot ShapiroWilk test Wvalue = 0.93 pvalue = 0.38 A: Mixer Ports B: Temp C: Time D: Batch Weight E: Delay Positive Effects Negative Effects HalfNormal % Probability B DE BE E D Standardized Effect Design Expert Output Response Std. Dev. ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant BTemp DBatch Weight 4.556E E EDelay BE DE Residual Cor Total.4 5 The Model Fvalue of 8.96 implies the model is significant. There is only a 0.9% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0. RSquared Mean.3 Adj RSquared C.V. % 7. Pred RSquared PRESS.3 Adeq Precision 9.7 Final Equation in Terms of Coded Factors: Std. Dev. = * B * D 0.4 * E 0.0 * B * E 0.6 * D * E 878
175 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Std. Dev. = * Temp * Batch Weight * Delay * Temp * Delay * Batch Weight * Delay (d) Does a residual analysis indicate any problems with the underlying assumptions? Often a transformation such as the natural log is required for the standard deviation response; however, the following residuals appear to be acceptable without the transformation. Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. Mixer Ports Residuals vs. Temp Residuals Residuals Mixer Ports B:Temp 879
176 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Time Residuals vs. Batch Weight Residuals Residuals C:Time D:Batch Weight Residuals vs. Delay Residuals E:Delay (e) Draw conclusions about this filling process. From the interaction plots below, the lowest standard deviation can be achieved with the Temperature at ambient, Batch Weight at 000 lbs, and a Delay of 7 days. 880
177 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Std. Dev. Design Points. Interaction B: Temp DesignExpert Software Std. Dev. Design Points. Interaction E: Delay B cooled B ambient E.000 E X = E: Delay X = B: Temp.675 X = D: Batch Weight X = E: Delay.675 Actual Factors A: Mixer Ports =.00 C: Time = D: Batch Weight = Std. Dev..5 Actual Factors A: Mixer Ports =.00 B: Temp = ambient C: Time = Std. Dev E: Delay D: Batch Weight 8.3. Heat treating is often used to carbonize metal parts, such as gears. The thickness of the carbonized layer is a critical output variable from this process, and it is usually measured by performing a carbon 6 analysis on the gear pitch (top of the gear tooth). Six factors were studied on a IV design: A = furnace temperature, B = cycle time, C = carbon concentration, D = duration of the carbonizing cycle, E = carbon concentration of the diffuse cycle, and F = duration of the diffuse cycle. The experiment is shown in Table P8.. Table P8. Standard Run Order Order A B C D E F Pitch (a) Estimate the factor effects and plot them on a normal probability plot. Select a tentative model. 88
178 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Term Effect SumSqr % Contribtn Require Intercept Model ATemp Error BCycle Time Model CCarbon Conc Model DDuration of Carbon Cycle Model ECarbon Conc.Diffuse Cycle Error FDuration of Diffuse cycle Error AB Error AC Error AD Error AE Error BD Model CD Model DE Error ABD \Error ABF Lenth's ME Lenth's SME Factors A, C, D, E and the two factor interactions CD and DE appear to be significant. The CD and DE interactions are aliased with BF and AF interactions respectively. Because factors B and F are not significant, CD and DE were included in the model. The model can be found in the Design Expert Output below. DesignExpert Software Pitch Normal Plot ShapiroWilk test Wvalue = 0.9 pvalue = A: Temp B: Cycle Time C: Carbon Conc D: Duration of Carbon Cycle E: Carbon Conc.Diffuse Cycle F: Duration of Diffuse cycle Positive Effects Negative Effects Normal % Probability C CD D E A 5 DE Standardized Effect 88
179 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (b) Perform appropriate statistical tests on the model. Design Expert Output Response Pitch ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ATemp < CCarbon Conc DDuration of Carbon Cycle < ECarbon Conc.Diffuse Cycle < CD DE Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 7.6 RSquared Mean Adj RSquared C.V. % 5.35 Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATemp CCarbon Conc DDuration of Carbon Cycle ECarbon Conc.Diffuse Cycle CD DE Final Equation in Terms of Coded Factors: Pitch = * A * C * D * E * C * D * D * E Final Equation in Terms of Actual Factors: Pitch = * Temp * Carbon Conc * Duration of Carbon Cycle * Carbon Conc.Diffuse Cycle * Carbon Conc * Duration of Carbon Cycle * Duration of Carbon Cycle * Carbon Conc.Diffuse Cycle 883
180 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (c) Analyze the residuals and comment on model adequacy. The residual plots are acceptable. The normality and equality of variance assumptions are verified. There does not appear to be any trends or interruptions in the residuals versus run order plot. The plots of the residuals versus factors C and E identify reduced variation at the lower level of both variables while the plot of residuals versus factor F identifies reduced variation at the upper level. Because C and E are significant factors in the model, this might not affect the decision on the optimum solution for the process. However, factor F is not included in the model and may be set at the upper level to reduce variation. Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Residuals Residual Predicted Residuals vs. Run Residuals vs. Temp Residuals .5 Residuals Run Number Temp 884
181 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Cycle Time Residuals vs. Carbon Conc Residuals .5 Residuals B:Cycle Time C:Carbon Conc Residuals vs. Duration of Carbon Cycle Residuals vs. Carbon Conc.Diffuse Cycle Residuals .5 Residuals D:Duration of Carbon Cycle E:Carbon Conc.Diffuse Cycle Residuals vs. Duration of Diffuse cycle Residuals F:Duration of Diffuse cycle 885
182 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (d) Interpret the results of this experiment. Assume that a layer thickness of between 40 and 60 is desirable. The graphs below identify a region that is acceptable between 40 and 60. DesignExpert Software Pitch D D Interaction D: Duration of Carbon Cycle DesignExpert Software Pitch E E Interaction E: Carbon Conc.Diffuse Cycle X = C: Carbon Conc X = D: Duration of Carbon Cycle Actual Factors A: Temp = 0.00 B: Cycle Time = 0.00 E: Carbon Conc.Diffuse Cycle = 0.00 F: Duration of Diffuse cycle = 0.00 Pitch X = D: Duration of Carbon Cycle X = E: Carbon Conc.Diffuse Cycle Actual Factors A: Temp = 0.00 B: Cycle Time = 0.00 C: Carbon Conc = 0.00 F: Duration of Diffuse cycle = 0.00 Pitch C: Carbon Conc D: Duration of Carbon Cycle DesignExpert Software Overlay Plot Pitch X = C: Carbon Conc X = D: Duration of Carbon Cycle Actual Factors A: Temp = 0.50 B: Cycle Time = 0.00 E: Carbon Conc.Diffuse Cycle = 0.00 F: Duration of Diffuse cycle = 0.00 D: Duration of Carbon Cycle Overlay Plot Pitch: 60 Pitch: C: Carbon Conc An experiment is run in a semiconductor factory to investigate the effect of six factors on transistor 6 gain. The design selected is the IV shown in Table P8.3. Table P8.3 Standard Run Order Order A B C D E F Gain
183 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) Use a normal plot of the effects to identify the significant factors. Design Expert Output Term Effect SumSqr % Contribtn Model AA Model BB Model CC Error DD Error EE Error FF Model AB Model AC Error AD Error AE Error AF Error BD Error BF Error ABD Error ABF Lenth's ME Lenth's SME.569 (b) Conduct appropriate statistical tests for the model identified in part (a). Design Expert Output Response Gain ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA < BB < CC < AB < AC Residual Cor Total The Model Fvalue of 5.74 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared 0.99 Mean Adj RSquared C.V. % 0.4 Pred RSquared PRESS Adeq Precision
184 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AA BB CC AB AC Final Equation in Terms of Coded Factors: Gain = * A * B 5.5 * C +3.5 * A * B * A * C Final Equation in Terms of Actual Factors: Gain = * A * B * C * A * B * A * C (c) Analyze the residuals and comment on your findings. The residual plots are acceptable. The normality and equality of variance assumptions are verified. There does not appear to be any trends or interruptions in the residuals versus run order plot. Normal Plot of Residuals Residuals vs. Predicted.5 99 Normal % Probability Residuals Residual Predicted 888
185 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Run Residuals vs. A Residuals.875 Residuals Run Number A Residuals vs. B Residuals vs. C Residuals.875 Residuals B:B C:C 889
186 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. D Residuals vs. E Residuals.875 Residuals D:D E:E Residuals vs. F Residuals F:F (d) Can you find a set of operating conditions that produce gain of 500 5? Yes, see the graphs below. 890
187 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Gain B B Interaction B: B DesignExpert Software Gain C C Interaction C: C X = A: A X = B: B Actual Factors C: C = 0.00 D: D = 0.00 E: E = 0.00 F: F = 0.00 Gain X = A: A X = C: C Actual Factors B: B = 0.00 D: D = 0.00 E: E = 0.00 F: F = 0.00 Gain A: A A: A DesignExpert Software Overlay Plot.00 Overlay Plot Gain X = A: A X = B: B Actual Factors C: C = .00 D: D = 0.00 E: E = 0.00 F: F = 0.00 B: B Gain: 55 Gain: A: A Consider the 6 IV design. (a) Suppose that the design had been folded over by changing the signs in the column B instead of column A. What changes would have resulted in the effects that can be estimated from the combined design? Minitab Output Fold Over on A Fractional Factorial Design Factors: 6 Base Design: 6, 6 Resolution: IV Runs: 3 Replicates: Fraction: / Blocks: Center pts (total): 0 Design Generators (before folding): E = ABC, F = BCD Folded on Factors: A Alias Structure 89
188 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY I + BCDF A + ABCDF B + CDF C + BDF D + BCF E + BCDEF F + BCD AB + ACDF AC + ABDF AD + ABCF AE + ABCDEF AF + ABCD BC + DF BD + CF BE + CDEF BF + CD CE + BDEF DE + BCEF EF + BCDE ABC + ADF ABD + ACF ABE + ACDEF ABF + ACD ACE + ABDEF ADE + ABCEF AEF + ABCDE BCE + DEF BDE + CEF BEF + CDE ABCE + ADEF ABDE + ACEF ABEF + ACDE Minitab Output Fold Over on B Fractional Factorial Design Factors: 6 Base Design: 6, 6 Resolution: IV Runs: 3 Replicates: Fraction: / Blocks: Center pts (total): 0 Design Generators (before folding): E = ABC, F = BCD Folded on Factors: B Alias Structure I + ADEF A + DEF B + ABDEF C + ACDEF D + AEF E + ADF F + ADE AB + BDEF AC + CDEF AD + EF AE + DF AF + DE BC + ABCDEF BD + ABEF BE + ABDF BF + ABDE CD + ACEF CE + ACDF CF + ACDE ABC + BCDEF ABD + BEF ABE + BDF 89
189 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY ABF + BDE ACD + CEF ACE + CDF ACF + CDE BCD + ABCEF BCE + ABCDF BCF + ABCDE ABCD + BCEF ABCE + BCDF ABCF + BCDE Both combined designs are still resolution IV, there are some twofactor interactions aliased with other twofactor interactions. In the combined design folded on A, all twofactor interactions with A are now aliased with fourfactor or higher interactions. The sparsity of effects principle would tell us that the higher order interactions are highly unlikely to occur. In the combined design folded on B, all twofactor interactions with B are now aliased with fourfactor or higher interactions. (b) Suppose that the design had been folded over by changing the sign in the column E instead of column A. What changes would have resulted in the effects that can be estimated from the combined design? Minitab Output Fold Over on B Fractional Factorial Design Factors: 6 Base Design: 6, 6 Resolution: IV Runs: 3 Replicates: Fraction: / Blocks: Center pts (total): 0 Design Generators (before folding): E = ABC, F = BCD Folded on Factors: E Alias Structure I + BCDF A + ABCDF B + CDF C + BDF D + BCF E + BCDEF F + BCD AB + ACDF AC + ABDF AD + ABCF AE + ABCDEF AF + ABCD BC + DF BD + CF BE + CDEF BF + CD CE + BDEF DE + BCEF EF + BCDE ABC + ADF ABD + ACF ABE + ACDEF ABF + ACD ACE + ABDEF ADE + ABCEF AEF + ABCDE BCE + DEF BDE + CEF BEF + CDE ABCE + ADEF ABDE + ACEF ABEF + ACDE In the combined design folded on E, all twofactor interactions with E are now aliased with fourfactor or higher interactions. 893
190 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Consider the 7 3 IV design. (a) Suppose that a partial fold over of this design is run using column A (+ signs only). Determine the alias relationship in the combined design. By choosing a fold over design in Design Expert, sorting on column A, and deleting the rows with a minus sign for A in the second block, the alias structures are identified below. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept + ABCE + ABFG + ACDG + ADEF + BCDF + BDEG + CEFG [A] = A  ABCE  ABFG  ACDG  ADEF + ABCDF + ABDEG + ACEFG [B] = B + ACE + AFG + CDF + DEG + ABCDG + ABDEF + BCEFG [C] = C + ABE + ADG + BDF + EFG + ABCFG + ACDEF + BCDEG [D] = D + ACG + AEF + BCF + BEG + ABCDE + ABDFG + CDEFG [E] = E + ABC + ADF + BDG + CFG + ABEFG + ACDEG + BCDEF [F] = F + ABG + ADE + BCD + CEG + ABCEF + ACDFG + BDEFG [G] = G + ABF + ACD + BDE + CEF + ABCEG + ADEFG + BCDFG [AB] = AB  ACE  AFG + ACDF + ADEG  ABCDG  ABDEF + ABCEFG [AC] = AC  ABE  ADG + ABDF + AEFG  ABCFG  ACDEF + ABCDEG [AD] = AD  ACG  AEF + ABCF + ABEG  ABCDE  ABDFG + ACDEFG [AE] = AE  ABC  ADF + ABDG + ACFG  ABEFG  ACDEG + ABCDEF [AF] = AF  ABG  ADE + ABCD + ACEG  ABCEF  ACDFG + ABDEFG [AG] = AG  ABF  ACD + ABDE + ACEF  ABCEG  ADEFG + ABCDFG [BC] = BC + DF + ABC + ADF + BEFG + CDEG + ABEFG + ACDEG [BD] = BD + CF + EG + ABCG + ABEF + ACDE + ADFG + BCDEFG [BE] = BE + DG + ABE + ADG + BCFG + CDEF + ABCFG + ACDEF [BF] = BF + CD + ABF + ACD + BCEG + DEFG + ABCEG + ADEFG [BG] = BG + DE + ABG + ADE + BCEF + CDFG + ABCEF + ACDFG [CE] = CE + FG + ACE + AFG + BCDG + BDEF + ABCDG + ABDEF [CG] = CG + EF + ACG + AEF + BCDE + BDFG + ABCDE + ABDFG [ABD] = ABD + ACF + AEG  ABCG  ABEF  ACDE  ADFG + ABCDEFG [BCE] = BCE + BFG + CDG + DEF + ABCE + ABFG + ACDG + ADEF [BCG] = BCG + BEF + CDE + DFG + ABCG + ABEF + ACDE + ADFG Factorial Effects Defining Contrast I = BCDF = BDEG = CEFG (b) Rework part (a) using the negative signs to define the partial fold over. Does it make any difference which signs are used to define the partial fold over? Both partial fold over designs produce the same alias relationships as shown below. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept + ABCE + ABFG + ACDG + ADEF + BCDF + BDEG + CEFG [A] = A  ABCE  ABFG  ACDG  ADEF + ABCDF + ABDEG + ACEFG [B] = B + ACE + AFG + CDF + DEG + ABCDG + ABDEF + BCEFG [C] = C + ABE + ADG + BDF + EFG + ABCFG + ACDEF + BCDEG [D] = D + ACG + AEF + BCF + BEG + ABCDE + ABDFG + CDEFG [E] = E + ABC + ADF + BDG + CFG + ABEFG + ACDEG + BCDEF [F] = F + ABG + ADE + BCD + CEG + ABCEF + ACDFG + BDEFG [G] = G + ABF + ACD + BDE + CEF + ABCEG + ADEFG + BCDFG [AB] = AB  ACE  AFG + ACDF + ADEG  ABCDG  ABDEF + ABCEFG [AC] = AC  ABE  ADG + ABDF + AEFG  ABCFG  ACDEF + ABCDEG [AD] = AD  ACG  AEF + ABCF + ABEG  ABCDE  ABDFG + ACDEFG [AE] = AE  ABC  ADF + ABDG + ACFG  ABEFG  ACDEG + ABCDEF [AF] = AF  ABG  ADE + ABCD + ACEG  ABCEF  ACDFG + ABDEFG [AG] = AG  ABF  ACD + ABDE + ACEF  ABCEG  ADEFG + ABCDFG 894
191 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY [BC] = BC + DF + ABC + ADF + BEFG + CDEG + ABEFG + ACDEG [BD] = BD + CF + EG + ABCG + ABEF + ACDE + ADFG + BCDEFG [BE] = BE + DG + ABE + ADG + BCFG + CDEF + ABCFG + ACDEF [BF] = BF + CD + ABF + ACD + BCEG + DEFG + ABCEG + ADEFG [BG] = BG + DE + ABG + ADE + BCEF + CDFG + ABCEF + ACDFG [CE] = CE + FG + ACE + AFG + BCDG + BDEF + ABCDG + ABDEF [CG] = CG + EF + ACG + AEF + BCDE + BDFG + ABCDE + ABDFG [ABD] = ABD + ACF + AEG  ABCG  ABEF  ACDE  ADFG + ABCDEFG [BCE] = BCE + BFG + CDG + DEF + ABCE + ABFG + ACDG + ADEF [BCG] = BCG + BEF + CDE + DFG + ABCG + ABEF + ACDE + ADFG Factorial Effects Defining Contrast I = BCDF = BDEG = CEFG 5 III Consider a partial fold over for the design. Suppose that the partial fold over of this design is constructed using column A (+ signs only). Determine the alias relationship for the combined design. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept + ABCE + ABFG + ACDG + ADEF + BCDF + BDEG + CEFG [Intercept] = Intercept  BD  CE + BCDE [Block ] = Block + BD + CE + ABD + ACE [Block ] = Block  BD  CE  ABD  ACE [A] = A + BD + CE + ABCDE [B] = B + AD + CDE + ABCE [C] = C + AE + BDE + ABCD [D] = D + AD + BCE + ABCE [E] = E + AE + BCD + ABCD [AB] = AB  AD  ABCE + ACDE [AC] = AC  AE  ABCD + ABDE [BC] = BC + DE + ABE + ACD [BE] = BE + CD + ABE + ACD [ABC] = ABC  ABE  ACD + ADE Factorial Effects Defining Contrast I = BCDE Consider a partial fold over for the IV design. Suppose that the signs are reversed in column A, but the eight runs are retained are the runs that have positive signs in column C. Determine the alias relationship in the combined design. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept  ABE + BCDF  ACDEF [Block ] = Block + ABE + ABCE + ADEF + ACDEF [Block ] = Block  ABE  ABCE  ADEF  ACDEF [A] = A + BCE + DEF + ABCDF [B] = B + ACE + CDF + ABDEF [C] = C + ABE + BDF + ACDEF [D] = D + BCF  ABDE  ACEF [E] = E + ABC + ADF + BCDEF [F] = F + BCD  ABEF  ACDE [AB] = AB + ABC + ADF + ACDF [AC] = AC  BCE  DEF + ABDF [AD] = AD  BDE  CEF + ABCF [AE] = AE + ACE + ABDEF + ABCDEF [AF] = AF  BEF  CDE + ABCD [BC] = BC + DF  ACE  ABDEF [BD] = BD + CF + ABEF + ACDE [BE] = BE + BCE + DEF + CDEF 895
192 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY [BF] = BF + CD + ABDE + ACEF [CE] = CE  ABC  ADF + BDEF [DE] = DE + BEF + CDE + BCEF [EF] = EF + BDE + CEF + BCDE [ABD] = ABD + ACF + BEF + CDE [ABF] = ABF + ACD + BDE + CEF [ADE] = ADE + ABEF + ACDE + ABCEF [AEF] = AEF + ABDE + ACEF + ABCDE Factorial Effects Defining Contrast I = BCDF Consider the 7 3 IV design. Suppose that fold over of this design is run by changing the signs in column A. Determine the alias relationship in the combined design. Minitab Output Fractional Factorial Design Factors: 7 Base Design: 7, 6 Resolution: IV Runs: 3 Replicates: Fraction: /4 Blocks: Center pts (total): 0 Design Generators (before folding): E = ABC, F = BCD, G = ACD Folded on Factors: A Alias Structure I + BCDF + BDEG + CEFG A + ABCDF + ABDEG + ACEFG B + CDF + DEG + BCEFG C + BDF + EFG + BCDEG D + BCF + BEG + CDEFG E + BDG + CFG + BCDEF F + BCD + CEG + BDEFG G + BDE + CEF + BCDFG AB + ACDF + ADEG + ABCEFG AC + ABDF + AEFG + ABCDEG AD + ABCF + ABEG + ACDEFG AE + ABDG + ACFG + ABCDEF AF + ABCD + ACEG + ABDEFG AG + ABDE + ACEF + ABCDFG BC + DF + BEFG + CDEG BD + CF + EG + BCDEFG BE + DG + BCFG + CDEF BF + CD + BCEG + DEFG BG + DE + BCEF + CDFG CE + FG + BCDG + BDEF CG + EF + BCDE + BDFG ABC + ADF + ABEFG + ACDEG ABD + ACF + AEG + ABCDEFG ABE + ADG + ABCFG + ACDEF ABF + ACD + ABCEG + ADEFG ABG + ADE + ABCEF + ACDFG ACE + AFG + ABCDG + ABDEF ACG + AEF + ABCDE + ABDFG BCE + BFG + CDG + DEF BCG + BEF + CDE + DFG ABCE + ABFG + ACDG + ADEF ABCG + ABEF + ACDE + ADFG Reconsider the 73 IV design in Problem (a) Suppose that a fold over of this design is run by changing the signs in column B. Determine the alias relationship in the combined design. 896
193 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Minitab Output Fractional Factorial Design Factors: 7 Base Design: 7, 6 Resolution: IV Runs: 3 Replicates: Fraction: /4 Blocks: Center pts (total): 0 Design Generators (before folding): E = ABC, F = BCD, G = ACD Folded on Factors: B Alias Structure I + ACDG + ADEF + CEFG A + CDG + DEF + ACEFG B + ABCDG + ABDEF + BCEFG C + ADG + EFG + ACDEF D + ACG + AEF + CDEFG E + ADF + CFG + ACDEG F + ADE + CEG + ACDFG G + ACD + CEF + ADEFG AB + BCDG + BDEF + ABCEFG AC + DG + AEFG + CDEF AD + CG + EF + ACDEFG AE + DF + ACFG + CDEG AF + DE + ACEG + CDFG AG + CD + ACEF + DEFG BC + ABDG + BEFG + ABCDEF BD + ABCG + ABEF + BCDEFG BE + ABDF + BCFG + ABCDEG BF + ABDE + BCEG + ABCDFG BG + ABCD + BCEF + ABDEFG CE + FG + ACDF + ADEG CF + EG + ACDE + ADFG ABC + BDG + ABEFG + BCDEF ABD + BCG + BEF + ABCDEFG ABE + BDF + ABCFG + BCDEG ABF + BDE + ABCEG + BCDFG ABG + BCD + ABCEF + BDEFG ACE + AFG + CDF + DEG ACF + AEG + CDE + DFG BCE + BFG + ABCDF + ABDEG BCF + BEG + ABCDE + ABDFG ABCE + ABFG + BCDF + BDEG ABCF + ABEG + BCDE + BDFG (b) Compare the aliases from this combined design to those from the combined design from Problem What differences resulted by changing the signs in a different column? Both combined designs are still resolution IV, there are some twofactor interactions aliased with other twofactor interactions. In the combined design folded on A, all twofactor interactions with A are now aliased with fourfactor or higher interactions. The sparsity of effects principle would tell us that the higher order interactions are highly unlikely to occur. In the combined design folded on B, all twofactor interactions with B are now aliased with fourfactor or higher interactions Consider a partial fold over for the III design. Suppose that the partial fold over of this design is constructed using column A (+ signs only). Determine the alias relationship for the combined design. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept  BD  CE  FG  ABCF + ABCG + ABEF  ABEG + ACDF  ACDG  ADEF + ADEG + BCDE + BDFG + CEFG  BCDEFG [Block ] = Block + BD + CE + FG + ABD + ACE + AFG + BCF + BEG + CDG + DEF 897
194 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY + ABCF + ABEG + ACDG + ADEF + BCDEFG + ABCDEFG [Block ] = Block  BD  CE  FG  ABD  ACE  AFG  BCF  BEG  CDG  DEF  ABCF  ABEG  ACDG  ADEF  BCDEFG  ABCDEFG [A] = A + BD + CE + FG + BCG + BEF + CDF + DEG + ABCF + ABEG + ACDG + ADEF + ABCDE + ABDFG + ACEFG + BCDEFG [B] = B + AD + CF + EG + ACG + AEF + CDE + DFG + ABCE + ABFG + BCDG + BDEF + ABCDF + ABDEG + BCEFG + ACDEFG [C] = C + AE + BF + DG + ABG + ADF + BDE + EFG + ABCD + ACFG + BCEG + CDEF + ABCEF + ACDEG + BCDFG + ABDEFG [D] = D + AD + CF + EG + ACF + AEG + BCE + BFG + ABCE + ABFG + BCDG + BDEF + ABCDG + ABDEF + CDEFG + ACDEFG [E] = E + AE + BF + DG + ABF + ADG + BCD + CFG + ABCD + ACFG + BCEG + CDEF + ABCEG + ACDEF + BDEFG + ABDEFG [F] = F + AG + BC + DE + ABE + ACD + BDG + CEG + ABDF + ACEF + BEFG + CDFG + ABCFG + ADEFG + BCDEF + ABCDEG [G] = G + AG + BC + DE + ABC + ADE + BDF + CEF + ABDF + ACEF + BEFG + CDFG + ABEFG + ACDFG + BCDEG + ABCDEG [AB] = AB  AD  CF + CG + EF  EG  ABCE  ABFG + ACDE + ADFG + BCDF  BCDG  BDEF + BDEG + ABCEFG  ACDEFG [AC] = AC  AE  BF + BG + DF  DG  ABCD + ABDE  ACFG + AEFG + BCEF  BCEG  CDEF + CDEG + ABCDFG  ABDEFG [AF] = AF  AG  BC + BE + CD  DE  ABDF + ABDG  ACEF + ACEG + BCFG  BEFG  CDFG + DEFG + ABCDEF  ABCDEG Factorial Effects Defining Contrast I = ABCG = ABEF = ACDF = ADEG = BCDE = BDFG = CEFG 8.4. Reconsider the 4 design in Example 8.. The significant factors are A, C, D, AC + BD, and AD + BC. Find a partial fold over design that will allow the AC, BD, AD, and BD interactions to be estimated. By constructing a partial fold over reversing the signs in column A and using column A (+ signs only), the AC, BD, AD, and BD interactions can be estimated as shown below. This could also be accomplished by reversing the signs of any one of the factors A, B, C, and D. Design Expert Output Factorial Effects Aliases [Est Terms] Aliased Terms [Intercept] = Intercept  BCD [Block ] = Block + BCD + ABCD [Block ] = Block  BCD  ABCD [A] = A + BCD [B] = B + ACD [C] = C + ABD [D] = D + ABC [AB] = AB  ACD [AC] = AC  ABD [AD] = AD  ABC [BC] = BC + ABC [BD] = BD + ABD [CD] = CD + ACD 8.4. How could an optimal design approach be used to augment a fractional factorial design to dealias effects of potential interest? Both Design Expert and JMP software packages have the capability to augment designs and dealias effects of potential interest. Upon completing the data entry and analysis in Design Expert for the fractional factorial experiment, design augmentation can be selected with the options of central composite, foldover, and fractional factorial. By selecting fractional factorial, and choosing the additional model term(s) to be dealiased, the Doptimal algorithm will insert additional runs as a new block for the experiment. 898
195 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Construct a supersaturated design for k = 8 factors in N = 6 runs. We have chosen the Hadamard matrix design approach using the following PlackettBurman design for N = runs and k = factors. By sorting on the th factor, L, deleting the rows with the positive levels on L, and removing the last 3 factors, the design is reduced to 6 runs with 8 factors. Either the positive or negative levels on L could have been chosen. Also, the JMP statistics software package has the capability to generate supersaturated designs. Run I A B C D E F G H J K L Original Run Run I A B C D E F G H Construct a supersaturated design for k = factors in N = 0 runs. We have chosen the Hadamard matrix design approach using the following PlackettBurman design for N = 0 runs and k = 9 factors. By sorting on the 9 th factor, T, deleting the rows with the negative levels on T, and removing the last 7 factors, the design is reduced to 0 runs with factors. Either the positive or negative levels on T could have been chosen. Also, the JMP statistics software package has the capability to generate supersaturated designs. Run I A B C D E F G H J K L M N O P Q R S T
196 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Original Run Run I A B C D E F G H J K L M
197 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter 9 ThreeLevel and MixedLevel Factorial and Fractional Factorial Design Solutions 9.. The effects of developer strength (A) and developer time (B) on the density of photographic plate film are being studied. Three strengths and three times are used, and four replicates of a 3 factorial experiment are run. The data from this experiment follow. Analyze the data using the standard methods for factorial experiments. Development Time (minutes) Developer Strength Design Expert Output Response: Data ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B AB Pure Error Cor Total The Model Fvalue of 0.66 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. Strength and time are significant. The quadratic and interaction effects are not significant. By treating both A and B as numerical factors, the analysis can be performed as follows: Design Expert Output Response: Data ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. 9
198 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 9.. Compute the I and J components of the twofactor interaction in Problem 9.. AB Totals = 77, 78, 7; IAB AB Totals = 78, 74, 74; JAB SS AB B 0 7 A I AB J AB A medical researcher is studying the effect of lidocaine on the enzyme level in the heart muscle of beagle dogs. Three different commercial brands of lidocaine (A), three dosage levels (B), and three dogs (C) are used in the experiment, and two replicates of a 3 3 factorial design are run. The observed enzyme levels follow. Analyze the data from this experiment. Replicate I Replicate II Lidocaine Dosage Dog Dog Brand Strength
199 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response Enzyme Level ANOVA for selected factorial model Analysis of variance table [Classical sum of squares  Type II] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ALidocaine Brand BDosage < CDog AB AC BC ABC Pure Error Cor Total The Model Fvalue of.9 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B are significant model terms. The factor B, dosage strength, is significant Compute the I and J components of the twofactor interactions for Example 9.. A B I totals = 74, 75, 6 J totals = 8, 3, 8 I(AB) = 6.78 J(AB) = 674. SS AB = A C I totals = 00, 34, 77 J totals = 5, 4,  I(AC) = J(AC) = 635. SS AC = B C I totals = 5, 79, 38 J totals =53, 87, 3 I(BC) = J(BC) = SS BC =
200 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 9.5. An experiment was performed to study the effect of three different types of 3ounce bottles (A) and three different shelf types (B)  smooth permanent shelves, endaisle displays with grilled shelves, and beverage coolers  on the time it takes to stock ten bottle cases on the shelves. Three workers (factor C) were employed in this experiment, and two replicates of a 3 3 factorial design were run. The observed time data are shown in the following table. Analyze the data and draw conclusions. Replicate I Replicate II Worker Bottle Type Permanent End Aisle Cooler Permanent End Aisle Cooler Plastic mm glass mm glass Plastic mm glass mm glass Plastic mm glass mm glass Design Expert Output Response Stock Time ANOVA for selected factorial model Analysis of variance table [Classical sum of squares  Type II] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ABottle Type BShelf Type < CWorker < AB AC BC ABC Pure Error Cor Total The Model Fvalue of 4.45 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, C, BC are significant model terms. Factors B and C, shelf type and worker, and the BC interaction are significant. For the shortest time regardless of worker chose the permanent shelves. This can easily be seen in the interaction plot below. 94
201 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DesignExpert Software Stock Time Design Points B Permanent B End Aisle B3 Cooler X = C: Worker X = B: Shelf Type Interaction B: Shelf Type Actual Factor A: Bottle Type = Plastic Stock Time C: Worker 9.6. An experiment is run in a chemical process using a 3 factorial design. The design factors are temperature and pressure, and the response variable is yield. The data that result from this experiment are shown below. Pressure, psig Temperature, C , , , , , , , , , (a) Analyze the data from this experiment by conducting an analysis of variance. What conclusions can you draw? Design Expert Output Response: Yield ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B AB Pure Error Cor Total The Model Fvalue of 4.37 implies the model is significant. There is only a.05% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B are significant model terms. Temperature and pressure are significant. Their interaction is not. An alternate analysis is performed below with A and B treated as numeric factors: Design Expert Output Response: Yield ANOVA for Selected Factorial Model 95
202 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B A B AB Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.90 implies the model is significant. There is only a 0.7% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case A, B, AB are significant model terms. (b) Graphically analyze the residuals. Are there any concerns about underlying assumptions or model adequacy? The following residual plots are based on the first analysis shown above with the A and B treated as categorical factors. The plot of residuals versus pressure shows a decreasing funnel shape indicating a nonconstant variance. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 96
203 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Pressure Residuals vs. Temperature Residuals 0 Residuals Pressure Temperature (c) Verify that if we let the low, medium and high levels of both factors in this experiment take on the levels , 0, and +, then a least squares fit to a second order model for yield is yˆ x 8.4x 7.7x 7.84x 7.69x x The coefficients can be found in the following table of computer output. Design Expert Output Final Equation in Terms of Coded Factors: Yield = * A * B * A * B * A * B (d) Confirm that the model in part (c) can be written in terms of the natural variables temperature (T) and pressure (P) as y T P0. 07T P TP The coefficients can be found in the following table of computer output. Design Expert Output Final Equation in Terms of Actual Factors: Yield = * Pressure * Temperature * Pressure * Temperature * Pressure * Temperature (e) Construct a contour plot for yield as a function of pressure and temperature. Based on the examination of this plot, where would you recommend running the process? 97
204 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Yield B: Temperature A: Pres s ure Run the process in the oval region indicated by the yield of Confound a 3 4 design in three blocks using the AB CD component of the fourfactor interaction. The three blocks are shown below with L = X + X + X 3 + X 4 Block Block Block Consider the data from the first replicate of Problem 9.5. Assuming that all 7 observations could not be run on the same day, set up a design for conducting the experiment over three days with AB C confounded with blocks. Analyze the data. Block Block Block = = = = = = = 5. = 5.5 = = = = = = = 4.5 = = 3.94 =
205 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY = 6.5 = = 6. 0 = 5.4 = 4.99 = = = = 5.48 Totals = The analysis of variance below identifies factors B and C as significant. Design Expert Output Response Stock Time ANOVA for selected factorial model Analysis of variance table [Classical sum of squares  Type II] Sum of Mean F pvalue Source Squares df Square Value Prob > F Block Model significant ABottle Type BShelf Type CWorker AB AC BC Residual Cor Total The Model Fvalue of 9.56 implies the model is significant. There is only a 0.5% chance that a "Model FValue" this large could occur due to noise. Values of "Prob > F" less than indicate model terms are significant. In this case B, C are significant model terms (a) Confound a 3 3 design in three blocks using the ABC component of the threefactor interaction. Compare your results with the design in Figure 9.7. L = X + X + X 3 Block Block Block The new design is a 80 rotation around the Factor B axis. (b) Confound a 3 3 design in three blocks using the AB C component of the threefactor interaction. Compare your results with the design in Figure 9.7. L = X + X + X 3 Block Block Block 3 99
206 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The new design is a 80 rotation around the Factor C axis. (c) Confound a 3 3 design three blocks using the ABC component of the threefactor interaction. Compare your results with the design in Figure 9.7. L = X + X + X 3 Block Block Block The new design is a 90 rotation around the Factor C axis along with switching layer 0 and layer in the C axis. (d) After looking at the designs in parts (a), (b), and (c) and Figure 9.7, what conclusions can you draw? All four designs are relatively the same. The only differences are rotations and swapping of layers Consider the data in Problem 9.5. If ABC is confounded in replicate I and ABC is confounded in replicate II, perform the analysis of variance. L = X + X + X 3 L = X + X + X 3 Block Block Block 3 Block Block Block = = = = = = 3.68 = 5.5 = = = = = 4.40 = = 5.67 = = 4.70 = 4.65 = = 5.48 = 6.5 = 4.99 = = 5.88 = = = = = = = = 4.6 = 5.7 = = 4.49 = 4.59 = = = = = = = = = = 5. 0 = = = = 6. 0 = = 5.80 = 4.08 = = 4.08 The sums of squares for A, B, C, AB, AC, and BC are calculated as usual. The only sums of squares presenting difficulties with calculations are the four components of the ABC interaction (ABC, ABC, AB C, and AB C ). ABC is computed using replicate I and ABC is computed using replicate II. AB C and AB C are computed using data from both replicates. 90
207 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY We will show how to calculate AB C and AB C from both replicates. Form a twoway table of A x B at each level of C. Find the I(AB) and J(AB) totals for each third of the A x B table. A C B 0 I J The I and J components for each third of the above table are used to form a new table of diagonal totals. C I(AB) J(AB) I Totals: I Totals: 85.06, 85.36, , 85.3, 83. J Totals: J Totals: 85.73, 83.70, , 85.6, 85.3 (85.06) (85.36) (83.3) (53.74) Now, AB C = I[C x I(AB)] = (85.73) (83.70) (84.3) (53.74) and, AB C = J[C x I(AB)]= If it were necessary, we could find ABC as ABC = I[C x J(AB)] and ABC as J[C x J(AB)]. However, these components must be computed using the data from the appropriate replicate. The analysis of variance table: 9
208 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Source SS DF MS F 0 Replicates.0980 Blocks within Replicates A B C AB < AC < BC ABC (rep I) ABC (rep II) < AB C AB C Error Total Outline the analysis of variance table for the 3 4 design in nine blocks. Is this a practical design? Source DF A B C D AB 4 AC 4 AD 4 BC 4 BD 4 CD 4 ABC (AB C,ABC,AB C ) 6 ABD (ABD,AB D,ABD ) 6 ACD (ACD,ACD,AC D ) 6 BCD (BCD,BC D,BCD ) 6 ABCD 6 Blocks (ABC,AB C,AC D,BC D ) 8 Total 80 Any experiment with 8 runs is large. Instead of having three full levels of each factor, if two levels of each factor could be used, then the overall design would have 6 runs plus some center points. This twolevel design could now probably be run in or 4 blocks, with center points in each block. Additional curvature effects could be determined by augmenting the experiment with the axial points of a central composite design and additional enter points. The overall design would be less than 8 runs. 9.. Construct a 3 4 IV The 7 runs for this design are as follows: design with I=ABCD. Write out the alias structure for this design
209 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A = AB C D = BCD B = AB CD = ACD C = ABC D = ABD D = ABCD = ABC AB = ABC D = CD AB = AC D = BC D AC = AB CD = BD AC = AB D = BC D BC = AB C D = AD BC = AB D = AC D BD = AB C = ACD CD = ABC = ABD AD = AB C = BCD 9.3. Verify that the design in Problem 9. is a resolution IV design. The design in Problem 9. is a Resolution IV design because no main effect is aliased with a component of a twofactor interaction, but some twofactor interaction components are aliased with each other From examining Figure 9.9, what type of design would remain if after completing the first 9 runs, one of the three factors could be dropped? The remaining design is a full 3 factorial Consider the data from replicate I in Problem 9.5. Suppose that only a onethird fraction of this design with I=ABC is run. Construct the design, determine the alias structure, and analyze the design. The design is 000, 0, 0, 0, 0,, 0, 0,. The alias structure is: A = BC = AB C B = AC = AB C C = AB = ABC AB = AC = BC B C A Permanent End Aisle Cooler Plastic mm glass mm glass 4.6 Plastic 6. 8mm glass mm glass 4.96 Plastic mm glass mm glass 4.85 Source SS DF A B.806 C.49 AB Total
210 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 9.6. Construct a design, and verify that it is a resolution III design. Use the generators I = AC D, I = AB C E, I = BC F, I = AB CG, I = ABCH, and I = ABJ To find the alias of any effect, multiply the effect by I and I. For example, the alias of C is: C = C(BC F ) = BF, At least one main effect is aliased with a component of a twofactor interaction Construct a 3 5 design with I=ABC and I=CDE. Write out the alias structure for this design. What is the resolution of this design? The complete defining relation for this design is: I = ABC = CDE = ABC DE = ABD E This is a resolution III design with x 4 = x + x + x 3 and x 5 = x + x + x 4 (mod 3) To find the alias of any effect, multiply the effect by I and I. For example, the alias of A is: A = AB C = ACDE = AB CDE = AB DE = BC = AC D E = BC DE = BD E 9.8. Construct a 4 x 3 design confounded in two blocks of 6 observations each. Outline the analysis of variance for this design. Design is a 4 x 3, with ABC at two levels, and Z at 4 levels. Represent Z with two pseudofactors D and E as follows: Factor Pseudo Factors Z D E Z 0 0 = () Z 0 = d Z 3 0 = e Z 4 = de 94
211 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The 4 x 3 is now a 5 in the factors A, B, C, D and E. Confound ABCDE with blocks. We have given both the letter notation and the digital notation for the treatment combinations. Block Block () = 0000 a = 000 ab = 00 b = 000 ac = 00 c = 000 bc = 00 abc = 0 abcd = bcd = 0 abce = bce = 0 cd = 00 acd = 0 ce = 00 ace = 0 de = 0003 ade = 003 abde = 03 bde = 003 bcde = 03 abcde = 3 be = 00 abd = 0 ad = 00 abe = 0 ae = 00 d = 000 acde = 03 e = 000 bd = 00 cde = 003 Source DF A B C Z (D+E+DE) 3 AB AC AZ (AD+AE+ADE) 3 BC BZ (BD+BE+BDE) 3 CZ (CD+CE+CDE) 3 ABC ABZ (ABD+ABE+ABDE) 3 ACZ (ACD+ACE+ACDE) 3 BCZ (BCD+BCE+BCDE) 3 ABCZ (ABCD+ABCE) Blocks (or ABCDE) Total Outline the analysis of variance table for a 3 factorial design. Discuss how this design may be confounded in blocks. Suppose we have n replicates of a 3 factorial design. A and B are at levels, and C and D are at 3 levels. Source DF Components for Confounding A A B B C C D D AB AB AC AC AD AD 95
212 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY BC BD BD CD,CD CD 4 ABC ABC ABD ABD ACD,ACD ACD 4 BCD,BCD BCD 4 ABCD,ABCD ABCD 4 Error 36(n) Total 36n Confounding in this series of designs is discussed extensively by Margolin (967). The possibilities for a single replicate of the 3 design are: blocks of 8 observations 4 blocks of 9 observations 9 blocks of 4 observations 3 blocks of observations 6 blocks of 6 observations For example, one component of the fourfactor interaction, say ABCD, could be selected to confound the design in 3 blocks of observations each, while to confound the design in blocks of 8 observations each we would select the AB interaction. Cochran and Cox (957) and Anderson and McLean (974) discuss confounding in these designs Starting with a 6run 4 design, show how one threelevel factor and three twolevel factors can be accommodated and still allow the estimation of twofactor interactions. Use columns A and B for the threelevel factor, and columns C and D and ABCD for the three twolevel factors. This design will be of resolution V. 9.. Starting with a 6run 4 design, show how two threelevel factors can be incorporated in this experiment. How many twolevel factors can be included if we want some information on twofactor interactions? Use column A and B for one threelevel factor and columns C and D for the other. Use the AC and BD columns for the two, twolevel factors. The design will be of resolution V. 9.. An article by W.D. Baten in the 956 volume of Industrial Quality Control described an experiment to study the effect of three factors on the lengths of steel bars. Each bar was subjected to one of two heat treatment processes and was cut on one of four machines at one of three times during the day (8 am, am, or 3 pm). The coded length data are shown below Time of Day Heat Treatment Process Machine am am pm
213 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) Analyze the data from this experiment assuming that the four observations in each cell are replicates. The Heat Treat Process effect (B) and Machine effect (C) are significant, while the Time of Day (A) is not significant. None of the interactions are significant. Design Expert Output Response: Length ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B C < AB AC BC ABC Pure Error Cor Total The Model Fvalue of 4.3 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..49 RSquared Mean 3.96 Adj RSquared 0.43 C.V Pred RSquared PRESS Adeq Precision 7.00 Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept A[] A[] BProcess C[] C[] C[3] A[]B A[]B A[]C[] A[]C[] A[]C[] A[]C[] A[]C[3] A[]C[3] BC[] BC[] BC[3] A[]BC[] A[]BC[] A[]BC[] A[]BC[] A[]BC[3] A[]BC[3]
214 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Length = * A[] * A[] .0 * B * C[] +.9 * C[] * C[3] +0.8 * A[]B * A[]B +0.5 * A[]C[] .58 * A[]C[] 0.0 * A[]C[] 0.4 * A[]C[] +0.8 * A[]C[3] * A[]C[3] +0.0 * BC[] 0.0 * BC[] +0.5 * BC[3] 0.6 * A[]BC[] +0. * A[]BC[] * A[]BC[] * A[]BC[] 0.8 * A[]BC[3] +0.4 * A[]BC[3] (b) Analyze the residuals from this experiment. Is there any indication that there is an outlier in one cell? If you find an outlier, remove it and repeat the analysis from part (a). What are your conclusions? Standard Order 84, Time of Day at 3:00pm, Heat Treat #, Machine #, and length of 0, appears to be an outlier. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 98
215 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Time Residuals vs. Process Residuals Residuals Time Process Residuals vs. Machine Residuals Machine The following analysis was performed with the outlier described above removed. As with the original analysis, Heat Treat Process (B) and Machine (C) are significant, but now the AC interaction appears to be significant. Design Expert Output Response: Length ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B C < AB AC BC ABC Pure Error Cor Total
216 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The Model Fvalue of 4.89 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..36 RSquared 0.63 Mean 4.00 Adj RSquared C.V Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept A[] A[] BProcess C[] C[] C[3] A[]B A[]B A[]C[] A[]C[] A[]C[] A[]C[] A[]C[3] A[]C[3] BC[] BC[] BC[3] A[]BC[] A[]BC[] A[]BC[] A[]BC[] A[]BC[3] A[]BC[3] Final Equation in Terms of Coded Factors: Length = * A[] 0.4 * A[] * B * C[] +.8 * C[] 3.7 * C[3] * A[]B 0.3 * A[]B * A[]C[] .50 * A[]C[] * A[]C[] * A[]C[] +0.6 * A[]C[3] * A[]C[3] * BC[] +0.6 * BC[] * BC[3] 0.7 * A[]BC[] * A[]BC[] 0.3 * A[]BC[] 0.7 * A[]BC[] * A[]BC[3] * A[]BC[3] The following residual plots are acceptable. Both the normality and constant variance assumptions are satisfied 90
217 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Residuals vs. Time Residuals vs. Process Residuals Residuals Time Process Residuals vs. Machine Residuals Machine 9
218 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (c) Suppose that the observations in the cells are the lengths (coded) of bars processed together in heat treating and then cut sequentially (that is, in order) on the four machines. Analyze the data to determine the effects of the three factors on mean length. The analysis with all effects and interactions included is shown below. Design Expert Output Response: Length ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model A 3..6 B C AB AC BC ABC Pure Error Cor Total By removing the three factor interaction from the model and applying it to the error, the analysis identifies factors B and C as being significant as well as the AC interaction. Design Expert Output Response: Length ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C < AB AC BC Residual Cor Total When removing the remaining insignificant effects from the model, factors A, B, C, and the AC interaction are significant. Design Expert Output Response: Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B < C < AC Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared
219 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Mean 3.96 Adj RSquared C.V. 3.7 Pred RSquared PRESS 5.4 Adeq Precision Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept A[] A[] BProcess C[] C[] C[3] A[]C[] A[]C[] A[]C[] A[]C[] A[]C[3] A[]C[3] Final Equation in Terms of Coded Factors: Avg = * A[] * A[] .0 * B * C[] +.9 * C[] * C[3] +0.5 * A[]C[] .58 * A[]C[] 0.0 * A[]C[] 0.4 * A[]C[] +0.8 * A[]C[3] * A[]C[3] The following residual plots are acceptable. Both the normality and uniformity of variance assumptions are verified. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 93
220 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Time Residuals vs. Process Residuals 0 Residuals Time Process Residuals vs. Machine Residuals Machine (d) Calculate the log variance of the observations in each cell. Analyze the average length and the log variance of the length for each of the bars cut at each machine/heat treatment process combination. What conclusions can you draw? Factors A and C, are significant as well as the AB and BC interactions. Factor B remains in the model for hierarchal purposes. 94
221 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response: Var Transform: Base 0 log Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C AB AC Residual Cor Total The Model Fvalue of 9.44 implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0.7 RSquared Mean 0.65 Adj RSquared C.V Pred RSquared PRESS.8 Adeq Precision.45 Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept A[] A[] BProcess C[] C[] C[3] A[]B A[]B A[]C[] A[]C[] A[]C[] A[]C[] A[]C[3] A[]C[3] Final Equation in Terms of Coded Factors: Log 0 (Var) = * A[] * A[] * B +0. * C[] * C[] 0.9 * C[3] 0. * A[]B * A[]B * A[]C[] * A[]C[] * A[]C[] * A[]C[] * A[]C[3] 0.8 * A[]C[3] The following residual plots are acceptable, although the residuals versus time and residuals versus machine plots identify moderate inconsistency with variance. 95
222 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Residuals vs. Time Residuals vs. Process Residuals 0 Residuals Time Process Residuals vs. Machine Residuals Machine 96
223 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (e) Suppose the time at which a bar is cut really cannot be controlled during routine production. Analyze the average length and the log variance of the length for each of the bars cut at each machine/heat treatment process combination. What conclusions can you draw? The analysis of the average length is as follows: Design Expert Output Response: Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model A B AB Pure Error Cor Total Because the Means Square of the AB interaction is much less than the main effects, it is removed from the model and placed in the error. The average length is significantly affected by factor A, Heat Treat Process, and factor B, Machine. Design Expert Output Response: Avg ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.04% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0. RSquared Mean 3.96 Adj RSquared C.V. 5.3 Pred RSquared PRESS 0.9 Adeq Precision Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept AProcess B[] B[] B[3] Final Equation in Terms of Coded Factors: Avg = * A * B[] +.9 * B[] * B[3] The following residual plots are acceptable. Both the normality and uniformity of variance assumptions are verified. 97
224 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted Residuals vs. Process Residuals vs. Machine Residuals 0 Residuals Process Machine The Log(Var) analysis is shown below. Design Expert Output Response: Log(Var) ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model A.070E E003 B AB E003 Pure Error Cor Total Because the factor A and the AB interaction small Mean Squares, they were removed from the model and placed in the error. From the following analysis of variance, Machine (B) is significant. 98
225 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response: Var Transform: Base 0 log Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant B Residual E003 Cor Total The Model Fvalue of 7.58 implies the model is significant. There is only a 3.98% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 0.77 Adj RSquared C.V. 0.7 Pred RSquared PRESS Adeq Precision 6.5 Coefficient Standard 95% CI 95% CI Term Estimate DF Error Low High VIF Intercept B[] B[] B[3] Final Equation in Terms of Coded Factors: Log 0(Var) = * B[] * B[] 0.8 * B[3] The following residual plots are acceptable. Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted 99
226 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals vs. Process Residuals vs. Machine Residuals 0 Residuals Process Machine 9.3. In Problem 8.4, you met Harry and Judy PetersonNedry, two friends of the author who have a winery and vineyard in Newberg, Oregon. That problem described the application of twolevel fractional factorial designs to their 985 Pinor Noir product. In 987, they wanted to conduct another Pinot Noir experiment. The variables for this experiment were Variable Levels Clone of Pinot Noir Wadenswil, Pommard Berry Size Small, Large Fermentation temperature 80F, 85F, 90/80F, 90F Whole Berry None, 0% Maceration Time 0 days, days Yeast Type Assmanhau, Champagne Oak Type Troncais, Allier Harry and Judy decided to use a 6run twolevel fractional factorial design, treating the four levels of fermentation temperature as two twolevel variables. As in Problem 8.4, they used the rankings from a tastetest panel as the response variable. The design and the resulting average ranks are shown below: Run Clone Size Berry Temp. Ferm. Berry Whole Time Macer Type Yeast Type Oak Rank Average
227 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) Describe the aliasing in this design. The design is a resolution IV design such that the main effects are aliased with three factor interactions. Design Expert Output Term Aliases Intercept ABCG ABDH ABEF ACDF ACEH ADEG AFGH BCDE BCFH BDFG BEGH CDGH CEFG DEFH A BCG BDH BEF CDF CEH DEG FGH ABCDE B ACG ADH AEF CDE CFH DFG EGH C ABG ADF AEH BDE BFH DGH EFG D ABH ACF AEG BCE BFG CGH EFH E ABF ACH ADG BCD BGH CFG DFH F ABE ACD AGH BCH BDG CEG DEH G ABC ADE AFH BDF BEH CDH CEF H ABD ACE AFG BCF BEG CDG DEF AB CG DH EF ACDE ACFH ADFG AEGH BCDF BCEH BDEG BFGH AC BG DF EH ABDE ABFH ADGH AEFG BCDH BCEF CDEG CFGH AD BH CF EG ABCE ABFG ACGH AEFH BCDG BDEF CDEH DFGH AE BF CH DG ABCD ABGH ACFG ADFH BCEG BDEH CDEF EFGH AF BE CD GH ABCH ABDG ACEG ADEH BCFG BDFH CEFH DEFG AG BC DE FH ABDF ABEH ACDH ACEF BDGH BEFG CDFG CEGH AH BD CE FG ABCF ABEG ACDG ADEF BCGH BEFH CDFH DEGH (b) Analyze the data and draw conclusions. All of the main effects except Yeast and Oak are significant. The Macer Time is the most significant. None of the interactions were significant. DESIGNEXPERT Plot Rank Normal plot A: Clone B: Berry Size C: Ferm Temp D: Ferm Temp E: Whole Berry F: Macer Time G: Yeast H: Oak Normal % probability A E D C B F Effect Design Expert Output Response: Rank ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A B C D E
228 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY F < Residual Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev.. RSquared Mean 8.50 Adj RSquared C.V. 3.5 Pred RSquared PRESS Adeq Precision 9.70 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AClone BBerry Size CFerm Temp DFerm Temp EWhole Berry FMacer Time Final Equation in Terms of Coded Factors: Rank = * A * B * C * D * E * F (c) What comparisons can you make between this experiment and the 985 Pinot Noir experiment from Problem 87? The experiment from Problem 8.4 indicates that yeast, barrel, whole cluster and the clone x yeast interactions were significant. This experiment indicates that maceration time, whole berry, clone and fermentation temperature are significant Suppose there are four threelevel categorical factors and a style twolevel continuous factor. What is the minimum number of runs required to estimate all main effects and twofactor interactions? Construct this design. A minimum of 4 runs are required to estimate all main effects and two factor interactions. Below is a D optimal design with 4 runs. Run Factor A Factor B Factor C Factor D Factor E
229 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Reconsider the experiment in Problem 9.4. Construct a design with N = 48 runs and compare it to the design you constructed in Problem 9.4. When generating the Doptimal design in Design Expert software, an additional six runs were included for lack of fit. The resulting design is shown below. As with the design created in Problem 9.4, the three level factors are balanced with equal runs at each individual factor level, and the two level factor is very slightly unbalanced with 5 low level runs and 3 high level runs. Run Factor A Factor B Factor C Factor D Factor E
230 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Reconsider the experiment in Problem 9.4. Suppose that you are only interested in main effects. Construct a design with N = runs for this experiment. Run Factor A Factor B Factor C Factor D Factor E Reconsider the experiment in Problem 9.. Suppose that it was necessary to estimate all main effects and twofactor interactions, but the full factorial with 4 runs (not counting replication) was too expensive. Recommend an alternative design. 934
231 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The design below is a Doptimal design generated by Design Expert software. For this design, zero additional runs were selected for lack of fit and replication. This is the minimum number of runs required to estimate all main effects and twofactor interactions. Additional runs could have been selected for lack of fit or replication. Run Heat Treatment Machine Time of Day
232 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter Response Surface Methods and Designs Solutions.. For the firstorder model ŷ 60. 5x 0. 8x. 0x3 find the path of steepest ascent. The variables are coded as x. ˆ Let the basic step size be x 3. 3 x3 or.0 = ˆ. x Therefore, in the coded variables we have i x ˆ Then. 0 Coded Variables x x x 3 Origin Origin Origin The region of experimentation for three factors are time ( 40 T 80 min), temperature ( 00 T 300 C), and pressure ( 0 P 50 psig). A firstorder model in coded variables has been fit to yield data from a 3 design. The model is ŷ 30 5x x. 5x 3. 5 Is the point T = 85, T = 35, P=60 on the path of steepest ascent? The coded variables are found with the following: 60 T 50 x T P 35 x x T 5 5 x ˆ 5 x or 0.5 = where 0 ˆ. 5 x ˆ x Coded Variables Natural Variables 3 
233 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY x x x 3 T T P Origin Origin Origin The point T =85, T =35, and P=60 is not on the path of steepest ascent..3. The region of experimentation for two factors are temperature ( 00 T 300 F) and catalyst feed rate ( 0 C 30 lb/h). A first order model in the usual coded variables has been fit to a molecular weight response, yielding the following model. (a) Find the path of steepest ascent. ŷ 000 5x 40x 00 T 0 x x C T 00 x 00 ˆ 5 x or 5 ˆ 40 x Coded Variables Natural Variables x x T C Origin Origin Origin (b) It is desired to move to a region where molecular weights are above 500. Based on the information you have from the experimentation in this region, about how may steps along the path of steepest ascent might be required to move to the region of interest? ŷ x ˆ ˆ x. # Steps A chemical plant produces oxygen by liquefying air and separating it into its component gases by fractional distillation. The purity of the oxygen is a function of the main condenser temperature and the pressure ratio between the upper and lower columns. Current operating conditions are temperature ( ) 0 C and pressure ratio ( ).. Using the following data find the path of steepest ascent. Temperature (x ) Pressure Ratio (x ) Purity
234 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response: Purity ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATemperature BPressure Ratio Curvature not significant Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 7.30 implies the model is significant. There is only a 0.47% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared C.V. % Pred RSquared PRESS Adeq Precision.9749 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATemperature BPressure Ratio Center Point Final Equation in Terms of Coded Factors: Purity = * A * B Final Equation in Terms of Actual Factors: Purity = * Temperature * Pressure Ratio From the computer output use the model ŷ x 0. 5x as the equation for steepest ascent. Suppose we use a one degree change in temperature as the basic step size. Thus, the path of steepest ascent passes through the point (x =0, x =0) and has a slope 0.5/0.85. In the coded variables, one degree of temperature is equivalent to a step of x /5=0.. Thus, x (0.5/0.85)0.= The path of steepest ascent is: 3
235 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Coded Variables Natural Variables x x Origin Origin Origin Origin An industrial engineer has developed a computer simulation model of a twoitem inventory system. The decision variables are the order quantity and the reorder point for each item. The response to be minimized is the total inventory cost. The simulation model is used to produce the data shown in the Table P.. Identify the experimental design. Find the path of steepest descent. Table P. Item Item Order Quantity (x) Reorder Point (x) Order Quantity (x3) Reorder Point (x4) Total Cost The design is a 4 fractional factorial with generator I=ABCD, and three center points. Design Expert Output Response: Total Cost ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A C D AC AD CD Curvature significant Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 63.6 implies the model is significant. There is only a 0.30% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 3.0 RSquared 0.99 Mean Adj RSquared C.V Pred RSquared
236 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY PRESS 9.50 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AItem QTY CItem QTY DItem Reorder AC AD CD Center Point Final Equation in Terms of Coded Factors: Total Cost = * A +3.5 * C * D * A * C * A * D +9.5 * C * D Final Equation in Terms of Actual Factors: Total Cost = * Item QTY * Item QTY * Item Reorder * Item QTY * Item QTY * Item QTY * Item Reorder * Item QTY * Item Reorder The equation used to compute the path of steepest ascent is ŷ x 3. 5x x4. Notice that even though the model contains interaction, it is relatively common practice to ignore the interactions in computing the path of steepest ascent. This means that the path constructed is only an approximation to the path that would have been obtained if the interactions were considered, but it s usually close enough to give satisfactory results. It is helpful to give a general method for finding the path of steepest ascent. Suppose we have a firstorder model in k variables, say ŷ ˆ k ˆ 0 i x i i The path of steepest ascent passes through the origin, x=0, and through the point on a hypersphere of radius, R where ŷ is a maximum. Thus, the x s must satisfy the constraint k i x i R To find the set of x s that maximize ŷ subject to this constraint, we maximize L ˆ k k ˆ 0 i xi xi i i R 5
237 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY where is a LaGrange multiplier. From L / xi L / 0, we find ˆ i xi It is customary to specify a basic step size in one of the variables, say x j, and then calculate as = ˆ j / x j. Then this value of can be used to generate the remaining coordinates of a point on the path of steepest ascent. We demonstrate using the data from this problem. Suppose that we use 0 units in as the basic step size. Note that a decrease in is called for, because we are looking for a path of steepest decent. Now 0 units in is equal to 0/0 = 0.5 units change in x. Thus, = ˆ / x = 9.5/(0.5) = Consequently, ˆ x ˆ x are the remaining coordinates of points along the path of steepest decent, in terms of the coded variables. The path of steepest decent is shown below: Coded Variables Natural Variables x x x 3 x Origin Origin Origin Verify that the following design is a simplex. Fit the firstorder model and find the path of steepest ascent. Position x x x 3 y
238 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 4 x x 3 3 x The graphical representation of the design identifies a tetrahedron; therefore, the design is a simplex. Design Expert Output Response y ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model AA BB CC Pure Error Cor Total Std. Dev. RSquared Mean Adj RSquared C.V. % Pred RSquared N/A PRESS N/A Adeq Precision.6446E308 Case(s) with leverage of.0000: Pred RSquared and PRESS statistic not defined Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AA BB CC Final Equation in Terms of Coded Factors: y = * A * B * C Final Equation in Terms of Actual Factors: y = * x * x * x37
239 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The first order model is yˆ x 0.35x. 50x. 3 To find the path of steepest ascent, let the basic step size be x 3. Then using the results obtained in the previous problem, we obtain ˆ 3.50 x3 or.0 = which yields. 50. Then the coordinates of points on the path of steepest ascent are defined by x x ˆ ˆ Therefore, in the coded variables we have: Coded Variables x x x 3 Origin Origin Origin The path of steepest ascent is usually computed assuming that the model is truly firstorder.; that is, there is no interaction. However, even if there is interaction, steepest ascent ignoring the interaction still usually produces good results. To illustrate, suppose that we have fit the model ŷ 0 5x x x x 8 3 using coded variables ( x +) (a) Draw the path of steepest ascent that you would obtain if the interaction were ignored. 8
240 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 0 Path of Steepest Ascent for Main Effects Model   X X (b) Draw the path of steepest ascent that you would obtain with the interaction included in the model. Compare this with the path found in part (a). 0 Path of Steepest Ascent for Full Model   X X.8. The data in Table P. were collected by a chemical engineer. The response y is filtration time, x is temperature, and x is pressure. Fit a secondorder model. Table P. x x y
241 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response Time ANOVA for Response Surface Reduced Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model not significant ATemperature BPressure AB A Residual Lack of Fit significant Pure Error Cor Total The "Model Fvalue" of.80 implies the model is not significant relative to the noise. There is a 0.07 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared C.V. % Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATemperature BPressure AB A Final Equation in Terms of Coded Factors: Time = * A.457 * B * A * B * A Final Equation in Terms of Actual Factors: Time = * Temperature.457 * Pressure * Temperature * Pressure * Temperature The lack of fit test in the above analysis is significant. The residual plots below are not unusual; however, the three largest residuals are axial points. Including cubic terms did not improve the lack of fit estimate and only made the PRESS larger. 0
242 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal Plot of Residuals Residuals vs. Predicted Normal % Probability Internally Studentized Residuals Internally Studentized Residuals Predicted Residuals vs. Temperature Residuals vs. Pressure Internally Studentized Residuals Internally Studentized Residuals A:Temperature B:Pressure (a) What operating conditions would you recommend if the objective is to minimize the filtration time? The contour plot below identifies the minimum filtration rate with the Temperature set at and the Pressure at . By setting the flag at this location in Design Expert, the prediction of filtration rate can be quickly estimated. 
243 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Time Time B: Pressure B: Pressure Prediction Prediction Observed % CI Low % CI High SE Mean SE Pred X X A: Temperature A: Temperature (b) What operating conditions would you recommend if the objective is to operate the process at a mean filtration time very close to 46? The contour plot below identifies filtration time of 46 with the highlighted contours. There are two regions where a filtration rate of 46 is achievable. If higher temperature and pressure require higher operating costs, it would be preferential to choose the region with lower temperature and pressure Time B: Pressure A: Temperature 
244 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.9. The data shown in Table P.3 were collected in an experiment to optimize crystal growth as a function of three variables x, x, and x 3. Large values of y (yield in grams) are desirable. Fit a second order model and analyze the fitted surface. Under what set of conditions is maximum growth achieved? Table P.3 x x x 3 y Design Expert Output Response Growth ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model not significant AA BB CC AB AC BC A B C Residual Lack of Fit not significant Pure Error Cor Total The "Model Fvalue" of.0 implies the model is not significant relative to the noise. There is a.79 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev. 3.4 RSquared Mean 83.5 Adj RSquared 0.36 C.V. % 6.0 Pred RSquared PRESS Adeq Precision
245 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AA BB CC AB AC BC A B C Final Equation in Terms of Actual Factors: Growth = * x * x .08 * x * x * x * x * x * x * x * x * x * x3 There are so many nonsignificant terms in this model that we should consider eliminating some of them. A reasonable reduced model is shown below. Design Expert Output Response Growth ANOVA for Response Surface Reduced Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant BB CC B C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 5. implies the model is significant. There is only a 0.84% chance that a "Model FValue" this large could occur due to noise. Std. Dev..8 RSquared Mean 83.5 Adj RSquared C.V. % 4.76 Pred RSquared PRESS Adeq Precision 5.96 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept BB CC B C
246 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Growth = * B .0 * C .96 * B 9.3 * C Final Equation in Terms of Actual Factors: Growth = * x .08 * x * x * x3 The contour plot identifies a maximum near the center of the design space Growth C: C Prediction % CI Low % CI High SE Mean 4.48 SE Pred.9975 X 0.07 X B: B.0. Consider the threevariable central composite design shown in Table P.4. Analyze the data and draw conclusions, assuming that we wish to maximize conversion (y ) with activity (y ) between 55 and 60. Table P.4 Time Run (min) Activity y Temperature Catalyst Conversion (%) (C) (%) y
247 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Quadratic models are developed for the Conversion and Activity response variables as follows: Design Expert Output Response Conversion ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATime BTemperature CCatalyst AB AC < BC A B C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.0 implies the model is significant. There is only a 0.04% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 4.8 RSquared Mean Adj RSquared C.V. % 6.4 Pred RSquared PRESS Adeq Precision.999 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATime BTemperature CCatalyst AB AC BC A B C Final Equation in Terms of Coded Factors: 6
248 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Conversion = * A * B +5.9 * C +.63 * A * B * A * C * B * C .73 * A * B * C Final Equation in Terms of Actual Factors: Conversion = * Time * Temperature * Catalyst * Time * Temperature * Time * Catalyst * Temperature * Catalyst * Time * Temperature * Catalyst Design Expert Output Response Activity ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATime < BTemperature CCatalyst AB AC BC A B C Residual Lack of Fit significant Pure Error Cor Total The Model Fvalue of 9.0 implies the model is significant. There is only a 0.09% chance that a "Model FValue" this large could occur due to noise. Std. Dev..78 RSquared 0.89 Mean Adj RSquared C.V. %.94 Pred RSquared PRESS 9.35 Adeq Precision
249 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATime BTemperature CCatalyst AB AC BC A B C Final Equation in Terms of Coded Factors: Activity = * A +0.9 * B +.6 * C * A * B 0.5 * A * C +0.4 * B * C * A * B * C Final Equation in Terms of Actual Factors: Activity = * Time * Temperature * Catalyst * Time * Temperature * Time * Catalyst * Temperature * Catalyst * Time * Temperature * Catalyst Because many of the terms are insignificant, the reduced quadratic model is fit as follows: Design Expert Output Response Activity ANOVA for Response Surface Reduced Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ATime < CCatalyst < A Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..50 RSquared
250 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Mean Adj RSquared C.V. %.48 Pred RSquared 0.6 PRESS 3.69 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATime CCatalyst A Final Equation in Terms of Coded Factors: Activity = * A +.6 * C * A Final Equation in Terms of Actual Factors: Activity = * Time * Catalyst * Time.00 Conversion.00 Activity C: Catalyst A: Time A: Time C: Catalyst 9
251 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.00 Overlay Plot 0.50 C: Catalyst 0.00 Activity: 60 Conversion: A: Time The contour plots visually describe the models while the overlay plot identifies the acceptable region for the process... The hexagon design in Table P.5 is used in an experiment that has the objective of fitting a secondorder model. (a) Fit the secondorder model. Table P.5 x x y Design Expert Output Response y ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model not significant AA BB AB A B Residual Lack of Fit not significant 0
252 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Pure Error Cor Total The "Model Fvalue" of.7 implies the model is not significant relative to the noise. There is a 4.8 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared C.V. % Pred RSquared PRESS 57.5 Adeq Precision 4.39 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AA BB AB A B Final Equation in Terms of Coded Factors: y = * A.7304 * B.5469 * A * B * A * B (b) Perform the canonical analysis. What type of surface has been found? The full quadratic model is used in the following analysis because the reduced model is singular. Solution Variable Critical Value X X Predicted Value at Solution These values can be found using Design Expert, or by solving the following equation for xs: x s = /B  b where b= and B= The eigenvalues are found by solving BλI = 0. This equation reduces to λ λ =0. The roots are λ = and λ = Thus, the canonical form of the fitted model is yˆ w w. Since both eigenvalues are positive, the response is a minimum at the stationary point. (c) What operating conditions on x and x lead to the stationary point? The stationary point is (x,x ) = ( , ) (d) Where would you run this process if the objective is to obtain a response that is as close to 65 as possible? 
253 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY y B: B A: A Any value of x and x that give a point on the contour with value of 65 would be satisfactory... An experimenter has run a BoxBehnken design and has obtained the results as shown in Table P.6, where the response variable is the viscosity of a polymer. Table P.6 Level Temp. Agitation Rate Pressure x x x 3 High Middle Low (a) Fit the secondorder model. Run x x x 3 y
254 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response Viscosity ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATemp BAgit. Rate CPressure AB AC BC A B C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 9.66 implies the model is significant. There is only a.% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared C.V. % Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATemp BAgit. Rate CPressure AB AC BC A B C Final Equation in Terms of Coded Factors: Viscosity = * A * B * C * A * B * A * C * B * C * A * B * C 3
255 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Viscosity = * Temp * Agit. Rate * Pressure * Temp * Agit. Rate * Temp * Pressure.4000 * Agit. Rate * Pressure * Temp * Agit. Rate * Pressure (b) Perform the canonical analysis. What type of surface has been found? Solution Variable Critical Value X X X Predicted Value at Solution These values are found by solving the following equation for xs: x s= /B  b where b= and B= The eigenvalues are found by solving BλI = 0. This equation reduces to λ λ λ =0. The roots are λ = 9.756, λ =.354 and λ 3 = Thus, the canonical form of the fitted model is yˆ w 4.354w w3. The system is a saddle point. (c) What operating conditions on x, x, and x 3 lead to the stationary point? The stationary point is (x, x, x 3 ) = (3.977, .794, 5.49). This is outside the design region. It would be necessary to either examine contour plots or use numerical optimization methods to find desired operating conditions. (d) What operating conditions would you recommend if it is important to obtain a viscosity that is as close to 600 as possible? 4
256 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Viscosity C: Pressure Prediction A: Temp Any point on either of the contours showing a viscosity of 600 is satisfactory..3. A central composite design is run in a chemical vapor deposition process, resulting in the experimental data shown in Table P.7. Four experimental units were processed simultaneously on each run of the design, and the responses are the mean and variance of thickness, computed across the four units. Table P.7 x x y s (a) Fit a model to the mean response. Analyze the residuals. Design Expert Output Response: Mean Thick ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B A
257 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY B AB Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 6. implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 4.3 RSquared Mean 47.3 Adj RSquared C.V. 5.5 Pred RSquared PRESS Adeq Precision.6 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept Ax Bx A B AB Final Equation in Terms of Coded Factors: Mean Thick = * A * B * A * B +6.5 * A * B Final Equation in Terms of Actual Factors: Mean Thick = * x * x * x * x * x * x 6
258 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted A modest deviation from normality can be observed in the Normal Plot of Residuals; however, not enough to be concerned. (b) Fit a model to the variance response. Analyze the residuals. Design Expert Output Response: Var Thick ANOVA for Response Surface FI Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B AB Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 9.7 Adj RSquared C.V Pred RSquared PRESS 7.64 Adeq Precision 8.57 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept Ax Bx AB Final Equation in Terms of Coded Factors: Var Thick = * A .38 * B .5 * A * B 7
259 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Var Thick = * x * x * x * x Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted The residual plots are not acceptable. A transformation should be considered. If not successful at correcting the residual plots, further investigation into the two apparently unusual points should be made. (c) Fit a model to the ln(s ). Is this model superior to the one you found in part (b)? Design Expert Output Response: Var Thick Transform: Natural log Constant: 0 ANOVA for Response Surface FI Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B AB Residual E003 Lack of Fit E not significant Pure Error E003 Cor Total 0.7 The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean.8 Adj RSquared C.V Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept Ax Bx AB
260 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Ln(Var Thick) = * A 0.3 * B 0.4 * A * B Final Equation in Terms of Actual Factors: Ln(Var Thick) = * x * x * x * x Normal plot of residuals Residuals vs. Predicted Normal % probability Residuals Residual Predicted The residual plots are much improved following the natural log transformation; however, the two runs still appear to be somewhat unusual and should be investigated further. They will be retained in the analysis. (d) Suppose you want the mean thickness to be in the interval 450±5. Find a set of operating conditions that achieve the objective and simultaneously minimize the variance. 9
261 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.00 Mean Thick Ln(Var Thick) B: x B: x A: x A: x.00 Overlay Plot 0.50 B: x 0.00 Ln(Var Thick): Mean Thick: Mean Thick: A: x The contour plots describe the two models while the overlay plot identifies the acceptable region for the process. (e) Discuss the variance minimization aspects of part (d). Have you minimized total process variance? The within run variance has been minimized; however, the runtorun variation has not been minimized in the analysis. This may not be the most robust operating conditions for the process..4. A manufacturer of cutting tools has developed two empirical equations for tool life in hours (y ) and for tool cost in dollars (y ). Both models are linear functions of steel hardness (x ) and manufacturing time (x ). The two equations are ŷ 0 5x ŷ 3 3x x 4x 30
262 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY and both equations are valid over the range .5x i.5. Unit tool cost must be below $7.50 and life must exceed hours for the product to be competitive. Is there a feasible set of operating conditions for this process? Where would you recommend that the process be run? The contour plots below graphically describe the two models. The overlay plot identifies the feasible operating region for the process..50 Life 0.50 Cost B: Time B: Time A: Hardness A: Hardness.50 Overlay Plot 0.75 Cost: 7.5 B: Time 0.00 Lif e: A: Hardness 0 5x 3 3x x 4x Show that augmenting a k design with n c center points does not affect the estimates of the i (i=,,..., k), but that the estimate of the intercept 0 is the average of all k + n c observations. In general, the X matrix for the k design with n c center points and the y vector would be: 3
263 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY k X The upper half of the matrix is the usual notation of the k design The lower half of the matrix represents the center points (n c rows) c k n n n y y y y k +n c observations. k k c k n X X' g k g g g 0 y X' Grand total of all k +n c observations usual contrasts from k Therefore, c k n g ˆ 0 0, which is the average of all c k n observations, while k i i g ˆ, which does not depend on the number of center points, since in computing the contrasts g i, all observations at the center are multiplied by zero..6. Verify that an orthogonal firstorder design is also firstorder rotatable. To show that a first order orthogonal design is also first order rotatable, consider ) ˆ V( x ) ˆ V( ) x ˆ ˆ V( ŷ ) V( k i i i k i i i 0 0 since all covariances between i and j are zero, due to design orthogonality. Furthermore, we have: n ) ˆ V(... ) ˆ V( ) ˆ V( ˆ V k 0, so k i x i n n ŷ ) V( k i x i n y V ˆ) (
264 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY which is a function of distance from the design center (i.e. x=0), and not direction. Thus the design must be rotatable. Note that n is, in general, the number of points in the exterior portion of the design. If there are n c center points, then ˆ V( 0 ). ( n n ) c.7. Verify that the central composite design shown in Table P.8 blocks orthogonally. Table P.8 Block Block Block 3 x x x 3 x x x 3 x x x Note that each block is an orthogonal first order design, since the cross products of elements in different columns add to zero for each block. To verify the second condition, choose a column, say column x. Now For blocks and, So k u xu , and n=0 m m x 4, n m=6 m n u x m x u nm n 0.3 = 0.3 and condition is satisfied by blocks and. For block 3, we have m 6 0 x m , n m = 8, so 33
265 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY m n u x m x u nm n = 0.4 And condition is satisfied by block 3. Similar results hold for the other columns..8. Blocking in the central composite design. Consider a central composite design for k = 4 variables in two blocks. Can a rotatable design always be found that blocks orthogonally? To run a central composite design in two blocks, assign the n f factorial points and the n c center points to block and the k axial points plus n c center points to block. Both blocks will be orthogonal first order designs, so the first condition for orthogonal blocking is satisfied. The second condition implies that m m x x im im block n f n k n block c c However, xim n f in block and x im in block, so m m Which gives: n n n f f c k nc n f k n n f n Since 4 if the design is to be rotatable, then the design must satisfy n f c c n f n f k n n f n c c It is not possible to find rotatable central composite designs which block orthogonally for all k. For example, if k=3, the above condition cannot be satisfied. For k=, there must be an equal number of center points in each block, i.e. n c = n c. For k=4, we must have n c = 4 and n c =. 34
266 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.9. How could a hexagon design be run in two orthogonal blocks? The hexagonal design can be blocked as shown below. There are n c = n c = n c center points with n c even. 6 n Put the points,3,and 5 in block and,4,and 6 in block. Note that each block is a simplex..0. The rotatable central composite design. It can be shown that a secondorder design is rotatable if n x a b u iu x ju 0 if a or b (or both) are odd and if / 4 composite design these conditions lead to the factorial portion. n f n 4 x u iu n 3 x u iux ju. Show that for the central for rotatability, where n f is the number of points in The balance between + and  in the factorial columns and the orthogonality among certain columns in the X matrix for the central composite design will result in all odd moments being zero. To solve for use the following relations: then n n 4 4 x iu n f, u u n f 4 n u 4 4 x 4 iu 4 n n n f f 3 3( n f n u f ) x x iu iu x x ju ju n f.. Yield during the first four cycles of a chemical process is shown in the following table. The variables are percent concentration (x ) at levels 30, 3, and 3 and temperature (x ) at 40, 4, and 44F. Analyze by EVOP methods. 35
267 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Conditions Cycle () () (3) (4) (5) Cycle: n= Phase Calculation of Averages Calculation of Standard Deviation Operating Conditions () () (3) (4) (5) (i) Previous Cycle Sum Previous Sum S= (ii) Previous Cycle Average Previous Average = (iii) New Observation New S=Range x fk,n (iv) Differences Range= (v) New Sums New Sum S= (vi) New Averages New average S = New Sum S/(n)= Calculation of Effects Calculation of Error Limits A y3 y4 y y For New Average: S n B y3 y4 y y53.5 For New Effects: S n.78 AB y3 y4 y y For CIM: S n CIM y3 y4 y y5 4y Cycle: n= Phase Calculation of Averages Calculation of Standard Deviation Operating Conditions () () (3) (4) (5) (i) Previous Cycle Sum Previous Sum S= (ii) Previous Cycle Average Previous Average = (iii) New Observation New S=Range x fk,n=.38 (iv) Differences Range=4.6 (v) New Sums New Sum S=.38 (vi) New Averages New average S = New Sum S/(n)=.38 Calculation of Effects Calculation of Error Limits A y3 y4 y y5 3.8 For New Average: S n B y3 y4 y y53.3 For New Effects: S n.78 AB y3 y4 y y5 0.8 For CIM: S n CIM y3 y4 y y5 4y
268 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Cycle: n=3 Phase Calculation of Averages Calculation of Standard Deviation Operating Conditions () () (3) (4) (5) (i) Previous Cycle Sum Previous Sum S=.38 (ii) Previous Cycle Average Previous Average =.38 (iii) New Observation New S=Range x fk,n=.8 (iv) Differences Range=6.5 (v) New Sums New Sum S=3.66 (vi) New Averages New average S = New Sum S/(n)=.83 Calculation of Effects Calculation of Error Limits A y3 y4 y y5.37 For New Average: S n B y3 y4 y y For New Effects: S n.78 AB y3 y4 y y For CIM: S n CIM y3 y4 y y5 4y Cycle: n=4 Phase Calculation of Averages Calculation of Standard Deviation Operating Conditions () () (3) (4) (5) (i) Previous Cycle Sum Previous Sum S=3.66 (ii) Previous Cycle Average Previous Average =.83 (iii) New Observation New S=Range x fk,n=.45 (iv) Differences Range=6.63 (v) New Sums New Sum S=6. (vi) New Averages New average S = New Sum S/(n)=.04 Calculation of Effects Calculation of Error Limits A y3 y4 y y5.48 For New Average: S n B y3 y4 y y5 .3 For New Effects: S n.78 AB y3 y4 y y50.8 For CIM: S n CIM y3 y4 y y5 4y From studying cycles 3 and 4, it is apparent that A (and possibly B) has a significant effect. A new phase should be started following cycle 3 or 4... Consider a 3 design for fitting a firstorder model. (a) Evaluate the Dcriterion ( X X) for this design. ( X X) =.44E437
269 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (b) Evaluate the Acriterion tr ( XX) for this design. tr ( XX) = 0.5 (c) Find the maximum scaled prediction variance for this design. Is this design Goptimal? v x NVar ŷx Nx XX x 4. Yes, this is a Goptimal design..3. Repeat Problem. using a first order model with the twofactor interaction. ( X X) = 4.768E7 v x NVar tr ( XX) = ŷx Nx XX x 7. Yes, this is a Goptimal design..4. Suppose that you need to design an experiment to fit a quadratic model over the region x i, i=, subject to the constraint x x. If the constraint is violated, the process will not work properly. You can afford to make no more than n= runs. Set up the following designs: (a) An inscribed CCD with center points at x x 0 x x (a)* An inscribed CCD with center points at x x 0. 5 so that a larger design could be fit within the constrained region 38
270 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY x x (b) An inscribed 3 factorial with center points at x x 0. 5 x x (c) A Doptimal design. x x (d) A modified Doptimal design that is identical to the one in part (c), but with all replicate runs at the design center. 39
271 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY x x (e) Evaluate the ( X X) criteria for each design. X (a) (a)* (b) (c) (d) X (f) Evaluate the Defficiency for each design relative to the Doptimal design in part (c). (a) (a)* (b) (c) (d) Defficiency 4.5%.64% 49.4% 00.00% 87.3% (g) Which design would you prefer? Why? The offset CCD, (a)*, is the preferred design based on the Defficiency. Not only is it better than the D optimal design, (c), but it maintains the desirable design features of the CCD..5. Suppose that we approximate a response surface with a model of order d, such as y=x +, when the true surface is described by a model of order d >d ; that is E(y) = X + X. (a) Show that the regression coefficients are biased, that is, that E( )= +A, where A = (X X )  X X. A is usually called the alias matrix. where A X X XX ' ' β ' ' E XX Xy ' ' XX XEy ' ' XX XXβ Xβ ' ' ' X X X X β X X E ˆ β Aβ ' X X β (b) If d = and d =, and a full k is used to fit the model, use the result in part (a) to determine the alias structure. 40
272 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY In this situation, we have assumed the true surface to be first order, when it is really second order. If a full factorial is used for k=, then 0 X = X = and A = ˆ 0 0 Then, E = E ˆ 0 ˆ 0 The pure quadratic terms bias the intercept (c) If d =, d = and k=3, find the alias structure assuming that a 3 design is used to fit the model. 0 3 X = X = and A = ˆ ˆ Then, E = E ˆ ˆ (d) If d =, d =, k=3, and the simplex design in Problem.3 is used to fit the model, determine the alias structure and compare the results with part (c) X = X = and A =
273 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Then, E = E ˆ ˆ ˆ ˆ Notice that the alias structure is different from that found in the previous part for the 3 design. In general, the A matrix will depend on which simplex design is used..6. A chemical engineer wishes to fit a calibration curve for a new procedure used to measure the concentration of a particular ingredient in a product manufactured in his facility. Twelve samples can be prepared, having known concentration. The engineer wants to build a model for the measured concentrations. He suspects that a linear calibration curve will be adequate to model the measured concentration as a function of the known concentrations; that is, y = β 0 + β 0x +, where x is the actual concentration. Four experimental designs are under consideration. Design consists of 6 runs at known concentration and 6 runs at known concentration 0. Design consists of 4 runs at concentrations, 5.5, and 0. Design 3 consists of 3 runs at concentrations, 4, 7, and 0. Finally, design 4 consists of 3 runs at concentrations and 0 and 6 runs at concentration 5.5. (a) Plot the scaled variance of prediction for all four designs on the same graph over the concentration range. Which design would be preferable, in your opinion? Scaled Variance of Prediction Design 4 Design 3 Design Design Because it has the lowest scaled variance of prediction at all points in the design space with the exception of 5.5, Design is preferred. (b) For each design calculate the determinant of to the D criterion? ( X X). Which design would be preferred according 4
274 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design would be preferred. Design ( XX) (c) Calculate the Defficiency of each design relative to the best design that you found in part b. Design Defficiency 00.00% 8.65% % % (d) For each design, calculate the average variance of prediction over the set of points given by x =,.5,,.5,..., 0. Which design would you prefer according to the Vcriterion? Design is still preferred based on the Vcriterion. Average Variance of Prediction Design Actual Coded (e) Calculate the Vefficiency of each design relative to the best design you found in part (d). (f) What is the Gefficiency of each design? Design Vefficiency 00.00% 88.0% 3 8.4% % Design Gefficiency 00.00% 80.00% % %.7. Rework Problem.6 assuming that the model the engineer wishes to fit is a quadratic. Obviously, only designs, 3, and 4 can now be considered. 43
275 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Scaled Variance of Prediction Design 4 Design 3 Design Based on the plot, the preferred design would depend on the region of interest. Design 4 would be preferred if the center of the region was of interest; otherwise, Design would be preferred. Design ( XX) 4.704E E E07 Design is preferred based on ( X X). Design Defficiency 00.00% % % Average Variance of Prediction Design Actual Coded Design 4 is preferred. Design Vefficiency 9.89% % % 44
276 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Gefficiency 00.00% % %.8. Suppose that you want to fit a secondorder response surface model in a situation where there are k = 4 factors; however, one of the factors is categorical with two levels. What model should you consider for this experiment? Suggest an appropriate design for this situation. The following model is considered for this experiment: y 0 x x 3 x 3 z x z x z 3 x 3 z x x 3 x x 3 3 x x 3 x x 33 x 3 u The following JMP output identifies a computer generated design utilizing the Icriterion. Because this is a response surface model which is generally used for prediction, the Icriterion is an appropriate choice for a computer generated design. JMP Output Custom Design Factors Add N Factors X Continuous X Continuous X3 Continuous X4 Categorical Model Intercept Intercept X X X3 X4 X*X X*X3 X*X4 X*X3 X*X4 X3*X4 X*X X*X X3*X3 Design Run X X X3 X L L 3  L 4 L L 6 0 L 70 L L L L   L   L L L 5 0 L 45
277 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 6  L Prediction Variance Profile Variance L L X X X3 X4.9. Suppose that you want to fit a secondorder model in k = 5 factors. You cannot afford more than 5 runs. Construct both a Doptimal and an Ioptimal design for this situation. Compare the prediction variance properties of the designs. Which design would you prefer? The following JMP outputs identify both Doptimal and Ioptimal designs. The prediction variance profile shown in each JMP output is appreciably less for the Ioptimal design as well as flatter in the middle of the design. JMP Output Doptimal Design Factors Add N Factors X Continuous X Continuous X3 Continuous X4 Continuous X5 Continuous Design Run X X X3 X4 X
278 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Prediction Variance Profile Variance X X X3 X4 X5 Prediction Variance Surface 47
279 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY JMP Output IOptimal Design Factors Add N Factors X Continuous X Continuous X3 Continuous X4 Continuous X5 Continuous Design Run X X X3 X4 X Prediction Variance Profile Variance X X X3 X4 X548
280 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Prediction Variance Surface.30. Myers and Montgomery (00) describe a gasoline blending experiment involving three mixture components. There are no constraints on the mixture proportions, and the following 0 run design is used. Design Point x x x 3 y(mpg) , , , ½ ½ ½ 0 ½ ½ ½ /3 /3 /3 4.8, 4. 8 /3 /6 / /6 /3 / /6 /6 /3 3.7 (a) What type of design did the experimenters use? A simplex centroid design was used. (b) Fit a quadratic mixture model to the data. Is this model adequate? 49
281 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response: y ANOVA for Mixture Quadratic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant Linear Mixture AB AC BC Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 3.90 implies the model is significant. There is only a 4.35% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 4.6 Adj RSquared C.V..93 Pred RSquared 0.44 PRESS 5.7 Adeq Precision Coefficient Standard 95% CI 95% CI Component Estimate DF Error Low High Ax Bx Cx AB AC BC Final Equation in Terms of Pseudo Components: y = * A +4.3 * B +3.8 * C +.5 * A * B +. * A * C .09 * B * C Final Equation in Terms of Real Components: y = * x * x * x * x * x * x * x * x * x3 The quadratic terms appear to be insignificant. The analysis below is for the linear mixture model: Design Expert Output Response: y ANOVA for Mixture Quadratic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant Linear Mixture Residual Lack of Fit not significant Pure Error Cor Total
282 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The Model Fvalue of 0.64 implies the model is significant. There is only a 0.7% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 4.6 Adj RSquared C.V..78 Pred RSquared PRESS 3.6 Adeq Precision 8.75 Coefficient Standard 95% CI 95% CI Component Estimate DF Error Low High Ax Bx Cx Adjusted Adjusted Approx t for H0 Component Effect DF Std Error Effect=0 Prob > t Ax Bx Cx Final Equation in Terms of Pseudo Components: y = * A * B +3.9 * C Final Equation in Terms of Real Components: y = * x * x * x3 (c) Plot the response surface contours. What blend would you recommend to maximize the MPG? A: x B: x 0.00 y.00 C: x3 To maximize the miles per gallon, the recommended blend is x =, x = 0, and x 3 = An experimenter wishes to run a threecomponent mixture experiment. The constraints in the components proportions are as follows: 5
283 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 0. x 0. x 0.4 x (a) Set up an experiment to fit a quadratic mixture model. Use n=4 runs, with 4 replicates. Use the D criteria. (b) Draw the experimental design region. Std x x x A: x B: x C: x3 (c) Set up an experiment to fit a quadratic mixture model with n= runs, assuming that three of these runs are replicated. Use the Dcriterion. Std x x x
284 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A: x B: x C: x3 (d) Comment on the two designs you have found. The design points are the same for both designs except that the edge center on the xx3 edge is not included in the second design. None of the replicates for either design are in the center of the experimental region. The experimental runs are fairly uniformly spaced in the design region. 53
285 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter Robust Parameter Design and Process Robustness Studies Solutions.. Consider the experiment in Problem.0. Suppose that temperature is a noise variable ( z in coded units). Fit response models for both responses. Is there a robust design problem with respect to both responses? Find a set of conditions that maximize conversion with activity between 55 and 60 and that minimize variability transmitted from temperature. The analysis and models as found in problem.0 are shown below for both responses. There is a robust design problem with regards to the conversion response because of the significance of factor B, temperature, and the BC interaction. However, temperature is not significant in the analysis of the second response, activity. Design Expert Output Response Conversion ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATime BTemperature CCatalyst AB AC < BC A B C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.0 implies the model is significant. There is only a 0.04% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 4.8 RSquared Mean Adj RSquared C.V. % 6.4 Pred RSquared PRESS Adeq Precision.999 Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATime BTemperature CCatalyst AB AC BC A B C
286 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Coded Factors: Conversion = * A * B +5.9 * C +.63 * A * B * A * C * B * C .73 * A * B * C Final Equation in Terms of Actual Factors: Conversion = * Time * Temperature * Catalyst * Time * Temperature * Time * Catalyst * Temperature * Catalyst * Time * Temperature * Catalyst Design Expert Output Response Activity ANOVA for Response Surface Reduced Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant ATime < CCatalyst < A Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..50 RSquared Mean Adj RSquared C.V. %.48 Pred RSquared 0.6 PRESS 3.69 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept ATime CCatalyst A Final Equation in Terms of Coded Factors: Activity = * A +.6 * C * A 
287 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Final Equation in Terms of Actual Factors: Activity = * Time * Catalyst * Time The following contour plots of conversion, activity, and POE and the corresponding optimization plot identify a region where conversion is maximized, activity is between 55 and 60, and the transmitted variability from temperature is minimized. Factor A is set at 0. while C is set at Conversion Activity C: Catalyst C: Catalyst A: Time A: Time POE(Conversion).0 Overlay Plot C: Catalyst Prediction X X C: Catalyst POE(Conversion): 5.0 Conversion: 80. Activity: POE(Conversion): A: Time A: Time.. Reconsider the leaf spring experiment in Table.. Suppose that the objective is to find a set of conditions where the mean free height is as close as possible to 7.6 inches with a variance of free height as small as possible. What conditions would you recommend to achieve these objectives? 3
288 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A B C D E() E(+) y s ,7.78, , 7,5, , 8.8, , 7.88, , 7.56, , 7.56, , 7.56, , 7.75, , 8.00, , 7.44, , 8.09, , 7.69, , 7.5, , 7.8, , 7.8, , 7.50, By overlaying the contour plots for Free Height Mean and the Free Height Variance, optimal solutions can be found. To minimize the variance, factor B must be at the high level while factors A and D are adjusted to assure a mean of 7.6. The two overlay plots below set factor D at both low and high levels. Therefore, a mean as close as possible to 7.6 with minimum variance of can be achieved at A = 0.78, B = +, and D = . This can also be achieved with A = +0.07, B = +, and D = Free Height Mean One Factor Plot B: Heating Time Free Height Variance A: Furnace Temp B: Heating Time.00 Overlay Plot Free Height Variance: Overlay Plot Free Height Variance: Free Height Mean: 7.58 B: Heating Time 0.00 Free Height Mean: 7.6 Free Height Mean: 7.58 B: Heating Time 0.00 Free Height Mean: A: Furnace Temp A: Furnace Temp Factor D =  Factor D = + 4
289 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.3. Consider the bottle filling experiment in Problem 68. Suppose that the percentage of carbonation (A) is a noise variable ( in coded units). z (a) Fit the response model to these data. Is there a robust design problem? The following is the analysis of variance with all terms in the model followed by a reduced model. Because the noise factor A is significant, and the AB interaction is moderately significant, there is a robust design problem. Design Expert Output Response: Fill Deviation ANOVA for Response Surface Reduced Cubic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Cor Total Model significant A < B C AB AC BC ABC Pure Error Cor Total Based on the above analysis, the AC, BC, and ABC interactions are removed from the model and used as error. Design Expert Output Response: Fill Deviation ANOVA for Response Surface Reduced Cubic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B C AB Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of 6.84 implies there is a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev. 0.8 RSquared Mean.00 Adj RSquared C.V. 8.8 Pred RSquared PRESS 5.34 Adeq Precision 5.44 Final Equation in Terms of Coded Factors: Fill Deviation = * A +.3 * B * C * A * B (b) Find the mean model and either the variance model or the POE. 5
290 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY From the final equation shown in the above analysis, the mean model and corresponding contour plot is shown below. Ez y x, z.3x 0.88x Fill Deviation C: Speed B: Pressure Contour and 3D plots of the POE are shown below..000 POE(Fill Deviation) C: Speed POE(Fill Deviation) B: Pressure C: Speed B: Pressure (c) Find a set of conditions that result in mean fill deviation as close to zero as possible with minimum transmitted variance. The overlay plot below identifies a an operating region for pressure and speed that in a mean fill deviation as close to zero as possible with minimum transmitted variance. 6
291 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.00 Overlay Plot 0.50 C: Speed 0.00 POE(Fill Deviation): Fill Deviation: 0. Fill Deviation: B: Pressure.4. Reconsider the leaf spring experiment from Table.. Suppose that factors A, B and C are controllable variables, and that factors D and E are noise factors. Set up a crossed array design to investigate this problem, assuming that all of the twofactor interactions involving the controllable variables are thought to be important. What type of design have you obtained? The following experimental design has a 3 inner array for the controllable variables and a outer array for the noise factors. A total of 3 runs are required. Outer Array Inner Array D   A B C E Consider the experiment in Problem . Suppose that pressure is a noise variable ( z in coded units). Fit the response model for the viscosity response. Find a set of conditions that result in viscosity as close as possible to 600 and that minimize the variability transmitted from the noise variable pressure. 7
292 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Design Expert Output Response Viscosity ANOVA for Response Surface Reduced Quadratic Model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model significant ATemp BAgit. Rate CPressure AC A C Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of.49 implies the model is significant. There is only a 0.% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared 0.83 C.V. % 5.90 Pred RSquared PRESS Adeq Precision.895 Coefficient Standard 95% CI 95% CI FactorEstimate df Error Low High VIF Intercept ATemp BAgit. Rate CPressure AC A C Final Equation in Terms of Coded Factors: Viscosity = * A +7.6 * B * C * A * C * A * C From the final equation shown in the above analysis, the mean model in terms of the coded factors is shown below. E Z y x 7.6x 6x x 38.33x xx3 The corresponding contour and 3D plots for this model are shown below followed by the POE contour and 3D plots. Finally, the stacked contour plot is presented identifying a region with viscosity between 590 and 60 while minimizing the variability transmitted from the noise variable pressure. These conditions are in the region of factor A = 0.5 and factor B = . 8
293 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.000 Viscosity B: Agit. Rate A: Temp B: Agit. Rate A: Temp.000 POE(Viscosity) B: Agit. Rate POE(Viscosity) Prediction A: Temp B: Agit. Rate A: Temp.000 Overlay Plot B: Agit. Rate Viscosity: 590 POE(Viscosity): Viscosity: Viscosity: A: Temp 9
294 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.6. Continuation of Problem.4. Reconsider the leaf spring experiment from Table.. Suppose that A, B and C are controllable factors and that factors D and E are noise factors. Show how a combined array design can be employed to investigate this problem that allows all twofactor interactions to be estimated and only requires 6 runs. Compare this with the crossed array design from Problem.4. Can you see how in general combined array designs have fewer runs than crossed array designs? The following experiment is a 5 fractional factorial experiment where the controllable factors are A, B, and C and the noise factors are D and E. Only 6 runs are required versus the 3 runs required for the crossed array design in problem.4. A B C D E Free Height Consider the connector pulloff force experiment shown in Table.. What main effects and interactions involving the controllable variables can be estimated with this design? Remember that all of the controllable variables are quantitative factors. The design in Table. contains a 3 4 inner array for the controllable variables. This is a resolution III design which aliases the main effects with two factor interactions. The alias table below identifies the alias structure for this design. Because of the partial aliasing in this design, it is difficult to interpret the interactions. Design Expert Output Alias Matrix [Est. Terms] [Intercept] [A] [B] [C] [D] [A] [B] [C] [D] Aliased Terms = Intercept  BC  BD  CD = A * BC * BD * CD = B * AC * AD = C * AB * AD = D * AB * AC = A * BC * BD * CD = B * AC * AD + CD = C * AB * AD + BD = D * AB * AC + BC.8. Consider the connector pulloff force experiment shown in Table.. Show how an experiment can be designed for this problem that will allow a full quadratic model to be fit in the controllable variables 0
295 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY along all main effects of the noise variables and their interactions with the controllable variables. How many runs will be required in this design? How does this compare with the design in Table.? There are several designs that can be employed to achieve the requirements stated above. Below is a small central composite design with the axial points removed for the noise variables. Five center points are also included which brings the total runs to 35. As shown in the alias analysis, the full quadratic model for the controllable variables is achieved. A B C D E F G Design Expert Output Alias Matrix [Est. Terms] Aliased Terms [Intercept] = Intercept [A] = A [B] = B [C] = C [D] = D [E] = E + 0. * EG * FG [F] = F  EF  EG [G] = G  EF * EG * FG [A ] = A [B ] = B [C ] = C [D ] = D 
296 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY [E ] = E + F + G [AB] = AB * EG * FG [AC] = AC * EG * FG [AD] = AD * EG * FG [AE] = AE * EG * FG [AF] = AF + EF +.05 * EG * FG [AG] = AG + EF +.05 * EG * FG [BC] = BC * EG * FG [BD] = BD  EF * EG * FG [BE] = BE * EG * FG [BF] = BF +. * EG * FG [BG] = BG + EF * EG * FG [CD] = CD * EG * FG [CE] = CE  EF * EG * FG [CF] = CF  EF  0. * EG + 0. * FG [CG] = CG . * EG + 0. * FG [DE] = DE * EG * FG [DF] = DF  0. * EG + 0. * FG [DG] = DG  EF * EG * FG.9. A variation of Example.. In example. (which utilized data from Example 6) we found that one of the process variables (B = pressure) was not important. Dropping this variable produced two replicates of a 3 design. The data are shown below. C D A() A(+) y s , 48 7, , 80 60, , 45 00, , 70 86, Assume that C and D are controllable factors and that A is a noise factor. (a) Fit a model to the mean response. The following is the analysis of variance with all terms in the model: Design Expert Output Response: Mean ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model A B AB Pure Error Cor Total Based on the above analysis, the AB interaction is removed from the model and used as error. Design Expert Output Response: Mean ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B Residual Cor Total
297 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The Model Fvalue of implies there is a 5.05% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 70.9 Adj RSquared C.V..5 Pred RSquared PRESS.5 Adeq Precision 3.67 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AConcentration BStir Rate Final Equation in Terms of Coded Factors: Mean = * A +7.9 * B Final Equation in Terms of Actual Factors: Mean = * Concentration * Stir Rate The following is a contour plot of the mean model:.00 Mean B: Stir Rate A: Concentration (b) Fit a model to the ln(s ) response. The following is the analysis of variance with all terms in the model: Design Expert Output Response: Variance Transform: Natural log Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model A
298 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY B AB Pure Error Cor Total Based on the above analysis, the AB interaction is removed from the model and applied to the residual error. Design Expert Output Response: Variance Transform: Natural log Constant: 0 ANOVA for Selected Factorial Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model not significant A B Residual Cor Total The "Model Fvalue" of.94 implies the model is not significant relative to the noise. There is a 38.5 % chance that a "Model Fvalue" this large could occur due to noise. Std. Dev RSquared Mean 5.5 Adj RSquared C.V. 5.6 Pred RSquared PRESS 0.8 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AConcentration BStir Rate Final Equation in Terms of Coded Factors: Ln(Variance) = * A +0.7 * B Final Equation in Terms of Actual Factors: Ln(Variance) = * Concentration * Stir Rate The following is a contour plot of the variance model in the untransformed form: 4
299 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Variance B: Stir Rate A: Concentration (c) Find operating conditions that result in the mean filtration rate response exceeding 75 with minimum variance. The overlay plot shown below identifies the region required by the process:.00 Overlay Plot 0.50 Mean: 75 B: Stir Rate 0.00 Variance: A: Concentration (d) Compare your results with those from Example. which used the transmission of error approach. How similar are the two answers. The results are very similar. Both require the Concentration to be held at the high level while the stirring rate is held near the middle..0. An experiment has been run in a process that applies a coating material to a wafer. Each run in the experiment produced a wafer, and the coating thickness was measured several times at different locations on the wafer. Then the mean y, and standard deviation y of the thickness measurement was obtained. The data [adapted from Box and Draper (987)] are shown in the table below. Run Speed Pressure Distance Mean (y) Std Dev (y) 5
300 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (a) What type of design did the experimenters use? Is this a good choice of design for fitting a quadratic model? The design is a 3 3. A better choice would be a 3 central composite design. The CCD gives more information over the design region with fewer points. (b) Build models of both responses. The model for the mean is developed as follows: Design Expert Output Response: Mean ANOVA for Response Surface Reduced Cubic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model.89E E < significant A 5.640E E < B.55E E < C 3.E E < AB AC BC ABC Residual Cor Total.347E The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean Adj RSquared 0.94 C.V Pred RSquared PRESS.7E+005 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ASpeed
301 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY BPressure CDistance AB AC BC ABC Final Equation in Terms of Coded Factors: Mean = * A * B * C * A * B * A * C * B * C * A * B * C Final Equation in Terms of Actual Factors: Mean = * Speed * Pressure * Distance * Speed * Pressure * Speed * Distance * Pressure * Distance * Speed * Pressure * Distance The model for the Std. Dev. response is as follows. A square root transformation was applied to correct problems with the normality assumption. Design Expert Output Response: Std. Dev. Transform: Square root Constant: 0 ANOVA for Response Surface Linear Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant A B C Residual Cor Total The Model Fvalue of 4.34 implies the model is significant. There is only a.45% chance that a "Model FValue" this large could occur due to noise. Std. Dev..99 RSquared Mean 6.00 Adj RSquared C.V Pred RSquared PRESS Adeq Precision 7.78 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept ASpeed BPressure CDistance Final Equation in Terms of Coded Factors: Sqrt(Std. Dev.) = * A +. * B 7
302 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY +.03 * C Final Equation in Terms of Actual Factors: Sqrt(Std. Dev.) = * Speed * Pressure * Distance Because Factor A is insignificant, it is removed from the model. The reduced linear model analysis is shown below: Design Expert Output Response: Std. Dev. Transform: Square root Constant: 0 ANOVA for Response Surface Reduced Linear Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model significant B C Residual Cor Total The Model Fvalue of 5.40 implies the model is significant. There is only a.6% chance that a "Model FValue" this large could occur due to noise. Std. Dev RSquared Mean 6.00 Adj RSquared 0.59 C.V Pred RSquared PRESS 75.4 Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept BPressure CDistance Final Equation in Terms of Coded Factors: Sqrt(Std. Dev.) = * B +.03 * C Final Equation in Terms of Actual Factors: Sqrt(Std. Dev.) = * Pressure * Distance The following contour plots graphically represent the two models. 8
303 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY DESIGNEXPERT Plot Mean X = B: Pressure Y = C: Distance Design Points Actual Factor A: Speed =.00 C: Distance Mean DESIGNEXPERT Plot Sqrt(Std. Dev.) X = B: Pressure Y = C: Distance Design Points Actual Factor A: Speed =.00 C: Distance Std. Dev B: Pressure B: Pressure (c) Find a set of optimum conditions that result in the mean as large as possible with the standard deviation less than 60. The overlay plot identifies a region that meets the criteria of the mean as large as possible with the standard deviation less than 60. The optimum conditions in coded terms are approximately Speed =.0, Pressure = 0.75 and Distance = 0.5. DESIGNEXPERT Plot Overlay Plot X = B: Pressure Y = C: Distance Design Points Actual Factor A: Speed = Overlay Plot Std. Dev.: 60 Mean: 700 C: Distance B: Pressure.. In an article ( Let s All Beware the Latin Square, Quality Engineering, Vol., 989, pp ) J.S. Hunter illustrates some of the problems associated with 3 kp fractional factorial designs. Factor A is the amount of ethanol added to a standard fuel and factor B represents the air/fuel ratio. The response variable is carbon monoxide (CO) emission in g/m. The design is shown below. Design Observations A B x x y
304 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Notice that we have used the notation system of 0,, and to represent the low, medium, and high levels for the factors. We have also used a geometric notation of , 0, and. Each run in the design is replicated twice. (a) Verify that the secondorder model 7. 0x 4. 5x 4. 0x 9. 0x ŷ x x is a reasonable model for this experiment. Sketch the CO concentration contours in the x, x space. In the computer output that follows, the coded factors model is in the , 0, + scale. Design Expert Output Response: CO Emis ANOVA for Response Surface Quadratic Model Analysis of variance table [Partial sum of squares] Sum of Mean F Source Squares DF Square Value Prob > F Model < significant A < B < A B AB < Residual Lack of Fit not significant Pure Error Cor Total The Model Fvalue of implies the model is significant. There is only a 0.0% chance that a "Model FValue" this large could occur due to noise. Std. Dev..5 RSquared Mean 7.83 Adj RSquared C.V Pred RSquared PRESS 69.7 Adeq Precision.95 Coefficient Standard 95% CI 95% CI Factor Estimate DF Error Low High VIF Intercept AEthanol BAir/Fuel Ratio A B AB Final Equation in Terms of Coded Factors: CO Emis = * A * B * A * B * A * B 0
305 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY CO Emis B: Air/Fuel Ratio A: Ethanol (b) Now suppose that instead of only two factors, we had used four factors in a 3 4 fractional factorial design and obtained exactly the same data in part (a). The design would be as follows: Design Observations A B C D x x x 3 x 4 y y Calculate the marginal averages of the CO response at each level of four factors A, B, C, and D. Construct plots of these marginal averages and interpret the results. Do factors C and D appear to have strong effects? Do these factors really have any effect on CO emission? Why is their appearance effect strong? The marginal averages are shown below. Both Factors C and D appear to have an effect on CO emission. This is probably because both C and D are aliased with components of the interaction involving A and B, and there is a strong AB interaction. Level A B C D
306 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY One Factor Plot One Factor Plot CO Emis 76 CO Emis A: A B: B One Factor Plot One Factor Plot CO Emis 76 CO Emis C: C D: D (c) The design in part (b) allows the model y to be fitted. Suppose that the true model is y i x 4 i i 4 i iixi i xi ii xi i i ij x x Show that if j represents the least squares estimates of the coefficients in the fitted model, then ij i j 
307 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY E E E E E E E E E ˆ ˆ 3 4/ ˆ / ˆ 3 3 4/ ˆ 4 4 3/ ˆ 3 4/ ˆ / ˆ / 4 ˆ / 3 Does this help explain the strong effects for factors C and D observed graphically in part (b)? Let X and X
308 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY A XX X X / 0 / 0 0 / / 0 0 Then, ˆ ˆ ˆ ˆ E ˆ ˆ / ˆ 4 0 / 0 0 / ˆ ˆ 0 / Suppose that there are four controllable variables and two noise variables. It is necessary to fit a complete quadratic model in the controllable variables, the main effects of the noise variables, and the twofactor interactions between all controllable and noise factors. Set up a combined array design for this by modifying a central composite design. The following design is a half fraction central composite design with the axial points removed from the noise factors. The total number of runs is fortyeight which includes eight center points. Std A B C D E F
309 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Reconsider the situation in Problem.. Could a modified small composite design be used for this problem? Are there any disadvantages associated with the use of the small composite design? The axial points for the noise factors were removed in following small central composite design. Five center points are included. The small central composite design does have aliasing with noise factor E aliased with the AD interaction and noise factor F aliased with the BC interaction. These aliases are corrected by leaving the axial points for the noise factors in the design. Std A B C D E F
310 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Reconsider the situation in Problem.. What is the minimum number of runs that can be used to estimate all of the model parameters using a combined array design? Use a Doptimal algorithm to find a reasonable design for this problem. A minimum of 5 runs is required. The following design is a 36 run Doptimal with six additional runs included for lack of fit and five as replicates. It should be noted that Design Expert s Doptimal algorithm might not create the same design if repeated. Std A B C D E F Suppose that there are four controllable variables and two noise variables. It is necessary to estimate the main effects and twofactor interactions of all of the controllable variables, the main effects of the noise variables, and the twofactor interactions between all controllable and noise factors. If all factors are at two levels, what is the minimum number of runs that can be used to estimate all of the model parameters using a combined array design? Use a Doptimal algorithm to find a design. Twentyone runs are required for the model, with five additional runs for lack of fit, and five as replicates for a total of 3 runs as follows. It should be noted that Design Expert s Doptimal algorithm might not create the same design if repeated. Std A B C D E F
311 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Rework Problem.4 using the Icriterion to construct the design. Compare this design to the D optimal design in Problem.4. Which design would you prefer? The JMP Output shown below identifies an Ioptimal design with 5 runs. It should be noted that JMP s I optimal design algorithm might not create the same design if repeated. We would prefer the Ioptimal over the Doptimal because a response surface model is of interest in this problem. JMP Output Custom Design Factors Add N Factors X Continuous X Continuous X3 Continuous X4 Continuous Noise Continuous Noise Continuous Design Run X X X3 X4 Noise Noise
312 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Rework Problem.5 using the Icriterion. Compare this to the Doptimal design in Problem.5. Which design would you prefer? The JMP Output shown below identifies an Ioptimal design with runs. It should be noted that JMP s I optimal design algorithm might not create the same design if repeated. Because the problem requires the experiment be run in at only two levels for each variable, and Ioptimal algorithm in JMP generates center points for continuous variables, categorical variables were used. We would prefer the Doptimal over the I optimal assuming that the interest is to identify the important factors more so than to fit a response surface model. Also, the Doptimal algorithm in Design Expert will generate a design for this model with continuous variables and still maintain the requirement for running the experiment at only two levels for each variable. JMP Output Custom Design Factors Add N Factors X Continuous X Continuous X3 Continuous X4 Continuous Noise Continuous Noise Continuous Design Run X X X3 X4 Noise Noise
313 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY.8. Reconsider the wave soldering experiment in Problem.6. Find a combined array design for this experiment that requires fewer runs. The following experiment is a 84, resolution IV design with the defining relation I = BCDE = ACDF = ABCG = ABDH. Only 6 runs are required. A B C D E F G H Reconsider the wave soldering experiment in Problem.6. Suppose that it was necessary to fit a complete quadratic model in the controllable variables, all main effects of the noise variables, and all controllable variablenoise variable interactions. What design would you recommend? The following experiment is a small central composite design with five center points; the axial points for the noise factors have been removed. A total of 45 runs are required. A B C D E F G H
314 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY An experiment was run in a wave soldering process. There are five controllable variables and three noise variables. The response variable is the number of solder defects per million opportunities. The experimental design employed was the crossed array shown below. Outer Array F   Inner Array G   A B C D E H (a) What types of designs were used for the inner and outer arrays? The inner array is a 5 fractional factorial design with a defining relation of I = ACD = BCE = ABDE. The outer array is a 3 fractional factorial design with a defining relation of I = FGH. (b) Develop models for the mean and variance of solder defects. What set of operating conditions would you recommend? A B C D E y s
315 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY The following analysis identifies factors A, C, and E as being significant for the solder defects mean model. HalfNormal Plot 99 HalfNormal % Probability E A C Standardized Effect Design Expert Output Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model < significant AA CC < EE Residual Cor Total Std. Dev RSquared Mean 97.8 Adj RSquared C.V. % 3.90 Pred RSquared PRESS Adeq Precision Coefficient Standard 95% CI 95% CI Factor Estimate df Error Low High VIF Intercept AA CC EE Final Equation in Terms of Coded Factors: Mean = * A * C +.59 * E 3
316 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Although the natural log transformation is often utilized for variance response, a power transformation actually performed better for this problem per the BoxCox plot below. The analysis for the solder defect variance follows. DesignExpert Software (Variance)^ Lambda Current = Best =.09 Low C.I. =.5 High C.I. = BoxCox Plot for Power Transforms Recommend transform: Power (Lambda =.09) Ln(ResidualSS) Lambda HalfNormal Plot 99 HalfNormal % Probability E A C Standardized Effect Design Expert Output Response Variance Transform: Power Lambda:.09 Constant: 0 ANOVA for selected factorial model Analysis of variance table [Partial sum of squares  Type III] Sum of Mean F pvalue Source Squares df Square Value Prob > F Model 9.303E E significant AA.6E+03.6E BB.758E+0.758E EE 4.73E E AB.69E+03.69E Residual 4.7E E+0 Cor Total 9.344E
317 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Std. Dev E+005 RSquared Mean 6.388E+006 Adj RSquared C.V. % 5.80 Pred RSquared PRESS.97E+0 Adeq Precision Final Equation in Terms of Coded Factors: (Variance).09 = E E+006 * A 5.87E+005 * B .430E+006 * E .454E+006 * A * B The contour plots of the mean and variance models are shown below along with the overlay plot. Assuming that we wish to minimize both solder defects mean and variance, a solution is shown in the overlay plot with factors A = +, B = +, C = , D = 0, and E near Mean.000 Variance E: E E: E A: A A: A DesignExpert Software Original Scale Overlay Plot.000 Overlay Plot Mean Variance X = A: A X = E: E Coded Factors B: B =.000 C: C = D: D = E: E Variance: Mean: A: A 33
318 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 34
319 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter 3 Experiments with Random Factors Solutions 3.. An article in the Journal of Quality Technology (Vol. 3, No., 98, pp. 4) describes an experiment that investigates the effects of four bleaching chemicals on pulp brightness. These four chemicals were selected at random from a large population of potential bleaching agents. The data are as follows: Chemical Pulp Brightness (a) Is there a difference in the chemical types? Use = The computer output shows that the null hypothesis cannot be rejected. Therefore, there is no evidence that there is a difference in chemical types. Minitab Output ANOVA: Brightness versus Chemical Factor Type Levels Values Chemical random Analysis of Variance for Brightne Source DF SS MS F P Chemical Error Total Source Variance Error Expected Mean Square for Each Term component term (using restricted model) Chemical .0 () + 5() Error () (b) Estimate the variability due to chemical types. MS Treatment MSE n This agrees with the Minitab output. Because the variance component cannot be negative, this likely means that the variability due to chemical types is zero. (c) Estimate the variability due to random error (d) Analyze the residuals from this experiment and comment on model adequacy. 3
320 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Two data points appear to be outliers in the normal probability plot of effects. These outliers belong to chemical types and 3 and should be investigated. There seems to be much less variability in brightness with chemical type 4. Normal Probability Plot of the Residuals (response is Brightne) Normal Score Residual 5 0 Residuals Versus the Fitted Values (response is Brightne) 0 Residual Fitted Value
321 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals Versus Chemical (response is Brightne) 0 Residual Chemical An article in the Journal of the Electrochemical Society (Vol. 39, No., 99, pp ) describes an experiment to investigate the lowpressure vapor deposition of polysilicon. The experiment was carried out in a largecapacity reactor at Sematech in Austin, Texas. The reactor has several wafer positions, and four of these positions are selected at random. The response variable is film thickness uniformity. Three replicates of the experiments were run, and the data are as follows: Wafer Position Uniformity (a) Is there a difference in the wafer positions? Use = Yes, there is a difference. Minitab Output ANOVA: Uniformity versus Wafer Position Factor Type Levels Values Wafer Po fixed Analysis of Variance for Uniformi Source DF SS MS F P Wafer Po Error Total.4373 Source Variance Error Expected Mean Square for Each Term component term (using restricted model) Wafer Po () + 3Q[] Error 0.65 () 33
322 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (b) Estimate the variability due to wafer positions. (c) Estimate the random error component. MSTreatment MSE ˆ n ˆ (d) Analyze the residuals from this experiment and comment on model adequacy. Variability in film thickness seems to depend on wafer position. These observations also show up as outliers on the normal probability plot. Wafer position number appears to have greater variation in uniformity than the other positions. Normal Probability Plot of the Residuals (response is Uniformi) Normal Score Residual Residuals Versus the Fitted Values (response is Uniformi) Residual 03 Fitted Value 4 34
323 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Residuals Versus Wafer Po (response is Uniformi) Residual 0  Wafer Po A manufacturer suspects that the batches of raw material furnished by her supplier differ significantly in calcium content. There are a large number of batches currently in the warehouse. Five of these are randomly selected for study. A chemist makes five determinations on each batch and obtains the following data: Batch Batch Batch 3 Batch 4 Batch (a) Is there significant variation in calcium content from batch to batch? Use = Yes, as shown in the Minitab Output below, there is a difference. Minitab Output ANOVA: Calcium versus Batch Factor Type Levels Values Batch random 5,, 3, 4, 5 Analysis of Variance for Calcium Source DF SS MS F P Batch Error Total Variance Error Expected Mean Square for Each Source component term Term (using restricted model) Batch () + 5 () Error () 35
324 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY (b) Estimate the components of variance. MS ˆ Model MS n E ˆ MSE (c) Find a 95 percent confidence interval for MS L n MS Treatments E. F MS Treatments U n MSE F, a, N a L U L U, a, N a (d) Analyze the residuals from this experiment. Are the basic analysis of variance assumptions satisfied? There are five residuals that stand out in the normal probability plot. From the Residual vs. Batch plot, we see that one point per batch appears to stand out. A natural log transformation was applied to the data but did not change the results of the residual analysis. Further investigation should probably be performed to determine if these points are outliers. 36
325 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 3.4. A textile mill has a large number of looms. Each loom is supposed to provide the same output of cloth per minute. To investigate this assumption, five looms are chosen at random and their output is noted at different times. The following data are obtained: Loom Output (lb/min) (a) Explain why this is a random effects experiment. Are the looms equal in output? Use = The looms used in the experiment are a random sample of all the looms in the manufacturing area. The following is the analysis of variance for the data: 37
326 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Minitab Output ANOVA: Output versus Loom Factor Type Levels Values Loom random 5,, 3, 4, 5 Analysis of Variance for Output Source DF SS MS F P Loom Error Total Variance Error Expected Mean Square for Each Term Source component term (using restricted model) Loom () + 5 () Error () The looms are not equal in output. (b) Estimate the variability between looms. MS ˆ Treatments MS n E (c) Estimate the experimental error variance. ˆ MSE (d) Find a 95 percent confidence interval for MS L n MS Treatments E. F MS Treatments U n MSE F, a, N a L U L U, a, N a (e) Analyze the residuals from this experiment. Do you think that the analysis of variance assumptions are satisfied? There is nothing unusual about the residual plots; therefore, the analysis of variance assumptions are satisfied. 38
327 Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY 39
Chapter 6 The 2 k Factorial Design Solutions
Solutions from Montgomery, D. C. () Design and Analysis of Experiments, Wiley, NY Chapter 6 The k Factorial Design Solutions 6.. An engineer is interested in the effects of cutting speed (A), tool geometry
More information3.4. A computer ANOVA output is shown below. Fill in the blanks. You may give bounds on the Pvalue.
3.4. A computer ANOVA output is shown below. Fill in the blanks. You may give bounds on the Pvalue. Oneway ANOVA Source DF SS MS F P Factor 3 36.15??? Error??? Total 19 196.04 Completed table is: Oneway
More information20g g g Analyze the residuals from this experiment and comment on the model adequacy.
3.4. A computer ANOVA output is shown below. Fill in the blanks. You may give bounds on the Pvalue. Oneway ANOVA Source DF SS MS F P Factor 3 36.15??? Error??? Total 19 196.04 3.11. A pharmaceutical
More informationChapter 5 Introduction to Factorial Designs Solutions
Solutions from Montgomery, D. C. (1) Design and Analysis of Experiments, Wiley, NY Chapter 5 Introduction to Factorial Designs Solutions 5.1. The following output was obtained from a computer program that
More informationInstitutionen för matematik och matematisk statistik Umeå universitet November 7, Inlämningsuppgift 3. Mariam Shirdel
Institutionen för matematik och matematisk statistik Umeå universitet November 7, 2011 Inlämningsuppgift 3 Mariam Shirdel (mash0007@student.umu.se) Kvalitetsteknik och försöksplanering, 7.5 hp 1 Uppgift
More informationDesign of Engineering Experiments Chapter 5 Introduction to Factorials
Design of Engineering Experiments Chapter 5 Introduction to Factorials Text reference, Chapter 5 page 170 General principles of factorial experiments The twofactor factorial with fixed effects The ANOVA
More informationDesign and Analysis of
Design and Analysis of MultiFactored Experiments Module Engineering 79282 Twolevel Factorial Designs L. M. Lye DOE Course 1 The 2 k Factorial Design Special case of the general factorial design; k factors,
More informationDesign of Engineering Experiments Part 5 The 2 k Factorial Design
Design of Engineering Experiments Part 5 The 2 k Factorial Design Text reference, Special case of the general factorial design; k factors, all at two levels The two levels are usually called low and high
More information3.12 Problems 133 (a) Do the data indicate that there is a difference in results obtained from the three different approaches? Use a = 0.05. (b) Analyze the residuals from this experiment and comment on
More informationDesign & Analysis of Experiments 7E 2009 Montgomery
Chapter 5 1 Introduction to Factorial Design Study the effects of 2 or more factors All possible combinations of factor levels are investigated For example, if there are a levels of factor A and b levels
More informationCSCI 688 Homework 6. Megan Rose Bryant Department of Mathematics William and Mary
CSCI 688 Homework 6 Megan Rose Bryant Department of Mathematics William and Mary November 12, 2014 7.1 Consider the experiment described in Problem 6.1. Analyze this experiment assuming that each replicate
More informationThe 2 k Factorial Design. Dr. Mohammad Abuhaiba 1
The 2 k Factorial Design Dr. Mohammad Abuhaiba 1 HoweWork Assignment Due Tuesday 1/6/2010 6.1, 6.2, 6.17, 6.18, 6.19 Dr. Mohammad Abuhaiba 2 Design of Engineering Experiments The 2 k Factorial Design Special
More informationChapter 13 Experiments with Random Factors Solutions
Solutions from Montgomery, D. C. (01) Design and Analysis of Experiments, Wiley, NY Chapter 13 Experiments with Random Factors Solutions 13.. An article by Hoof and Berman ( Statistical Analysis of Power
More informationAnswer Keys to Homework#10
Answer Keys to Homework#10 Problem 1 Use either restricted or unrestricted mixed models. Problem 2 (a) First, the respective means for the 8 level combinations are listed in the following table A B C Mean
More informationAssignment 9 Answer Keys
Assignment 9 Answer Keys Problem 1 (a) First, the respective means for the 8 level combinations are listed in the following table A B C Mean 26.00 + 34.67 + 39.67 + + 49.33 + 42.33 + + 37.67 + + 54.67
More informationAddition of Center Points to a 2 k Designs Section 66 page 271
to a 2 k Designs Section 66 page 271 Based on the idea of replicating some of the runs in a factorial design 2 level designs assume linearity. If interaction terms are added to model some curvature results
More informationDesign and Analysis of MultiFactored Experiments
Design and Analysis of MultiFactored Experiments Twolevel Factorial Designs L. M. Lye DOE Course 1 The 2 k Factorial Design Special case of the general factorial design; k factors, all at two levels
More informationUnreplicated 2 k Factorial Designs
Unreplicated 2 k Factorial Designs These are 2 k factorial designs with one observation at each corner of the cube An unreplicated 2 k factorial design is also sometimes called a single replicate of the
More informationChapter 4 Randomized Blocks, Latin Squares, and Related Designs Solutions
Solutions from Montgomery, D. C. (008) Design and Analysis of Experiments, Wiley, NY Chapter 4 Randomized Blocks, Latin Squares, and Related Designs Solutions 4.. The ANOVA from a randomized complete block
More informationSTAT451/551 Homework#11 Due: April 22, 2014
STAT451/551 Homework#11 Due: April 22, 2014 1. Read Chapter 8.3 8.9. 2. 8.4. SAS code is provided. 3. 8.18. 4. 8.24. 5. 8.45. 376 Chapter 8 TwoLevel Fractional Factorial Designs more detail. Sequential
More informationChapter 11: Factorial Designs
Chapter : Factorial Designs. Two factor factorial designs ( levels factors ) This situation is similar to the randomized block design from the previous chapter. However, in addition to the effects within
More informationDesign of Experiments SUTD  21/4/2015 1
Design of Experiments SUTD  21/4/2015 1 Outline 1. Introduction 2. 2 k Factorial Design Exercise 3. Choice of Sample Size Exercise 4. 2 k p Fractional Factorial Design Exercise 5. Followup experimentation
More informationReference: Chapter 6 of Montgomery(8e) Maghsoodloo
Reference: Chapter 6 of Montgomery(8e) Maghsoodloo 51 DOE (or DOX) FOR BASE BALANCED FACTORIALS The notation k is used to denote a factorial experiment involving k factors (A, B, C, D,..., K) each at levels.
More informationWhat If There Are More Than. Two Factor Levels?
What If There Are More Than Chapter 3 Two Factor Levels? Comparing more that two factor levels the analysis of variance ANOVA decomposition of total variability Statistical testing & analysis Checking
More informationDesign of Experiments SUTD 06/04/2016 1
Design of Experiments SUTD 06/04/2016 1 Outline 1. Introduction 2. 2 k Factorial Design 3. Choice of Sample Size 4. 2 k p Fractional Factorial Design 5. Followup experimentation (folding over) with factorial
More informationSuppose we needed four batches of formaldehyde, and coulddoonly4runsperbatch. Thisisthena2 4 factorial in 2 2 blocks.
58 2. 2 factorials in 2 blocks Suppose we needed four batches of formaldehyde, and coulddoonly4runsperbatch. Thisisthena2 4 factorial in 2 2 blocks. Some more algebra: If two effects are confounded with
More informationPlasma Etch Tool Gap Distance DOE Final Report
Plasma Etch Tool Gap Distance DOE Final Report IEE 572 Doug Purvis Mei Lee Gallagher 12/4/00 Page 1 of 10 Protocol Purpose: To establish new Power, Pressure, and Gas Ratio setpoints that are acceptable
More informationChapter 30 Design and Analysis of
Chapter 30 Design and Analysis of 2 k DOEs Introduction This chapter describes design alternatives and analysis techniques for conducting a DOE. Tables M1 to M5 in Appendix E can be used to create test
More informationThe Model Building Process Part I: Checking Model Assumptions Best Practice
The Model Building Process Part I: Checking Model Assumptions Best Practice Authored by: Sarah Burke, PhD 31 July 2017 The goal of the STAT T&E COE is to assist in developing rigorous, defensible test
More informationUnit 6: Fractional Factorial Experiments at Three Levels
Unit 6: Fractional Factorial Experiments at Three Levels Largerthebetter and smallerthebetter problems. Basic concepts for 3 k full factorial designs. Analysis of 3 k designs using orthogonal components
More informationThe Model Building Process Part I: Checking Model Assumptions Best Practice (Version 1.1)
The Model Building Process Part I: Checking Model Assumptions Best Practice (Version 1.1) Authored by: Sarah Burke, PhD Version 1: 31 July 2017 Version 1.1: 24 October 2017 The goal of the STAT T&E COE
More informationST3232: Design and Analysis of Experiments
Department of Statistics & Applied Probability 2:004:00 pm, Monday, April 8, 2013 Lecture 21: Fractional 2 p factorial designs The general principles A full 2 p factorial experiment might not be efficient
More informationFractional Factorial Designs
Fractional Factorial Designs ST 516 Each replicate of a 2 k design requires 2 k runs. E.g. 64 runs for k = 6, or 1024 runs for k = 10. When this is infeasible, we use a fraction of the runs. As a result,
More informationExperimental design (DOE)  Design
Experimental design (DOE)  Design Menu: QCExpert Experimental Design Design Full Factorial Fract Factorial This module designs a twolevel multifactorial orthogonal plan 2 n k and perform its analysis.
More information7. Response Surface Methodology (Ch.10. Regression Modeling Ch. 11. Response Surface Methodology)
7. Response Surface Methodology (Ch.10. Regression Modeling Ch. 11. Response Surface Methodology) HaeJin Choi School of Mechanical Engineering, ChungAng University 1 Introduction Response surface methodology,
More informationDesign of Experiments
Design of Experiments D R. S H A S H A N K S H E K H A R M S E, I I T K A N P U R F E B 19 TH 2 0 1 6 T E Q I P ( I I T K A N P U R ) Data Analysis 2 Draw Conclusions Ask a Question Analyze data What to
More information9 Correlation and Regression
9 Correlation and Regression SW, Chapter 12. Suppose we select n = 10 persons from the population of college seniors who plan to take the MCAT exam. Each takes the test, is coached, and then retakes the
More informationVOL. 2, NO. 11, Dec 2012 ISSN ARPN Journal of Science and Technology All rights reserved.
Factorial Design Approach to Investigate the Effect of Different Factors on the Resilient Modulus of Bituminous Paving Mixes 1 Muhammad Babar Khan, Afaq Khattak, 3 Muhammad Irfan, Sarfraz Ahmed 1 Assistant
More informationTaguchi Design of Experiments
Taguchi Design of Experiments Many factors/inputs/variables must be taken into consideration when making a product especially a brand new one The Taguchi method is a structured approach for determining
More informationStudy of the effect of machining parameters on material removal rate and electrode wear during Electric Discharge Machining of mild steel
Journal of Engineering Science and Technology Review 5 (1) (2012) 1418 Research Article JOURNAL OF Engineering Science and Technology Review www.jestr.org Study of the effect of machining parameters on
More informationCHAPTER 5 NONLINEAR SURROGATE MODEL
96 CHAPTER 5 NONLINEAR SURROGATE MODEL 5.1 INTRODUCTION As set out in the research methodology and in sequent with the previous section on development of LSM, construction of the nonlinear surrogate
More information2015 StatEase, Inc. Practical Strategies for Model Verification
Practical Strategies for Model Verification *Presentation is posted at www.statease.com/webinar.html There are many attendees today! To avoid disrupting the Voice over Internet Protocol (VoIP) system,
More information19. Blocking & confounding
146 19. Blocking & confounding Importance of blocking to control nuisance factors  day of week, batch of raw material, etc. Complete Blocks. This is the easy case. Suppose we run a 2 2 factorial experiment,
More informationAn Introduction to Design of Experiments
An Introduction to Design of Experiments Douglas C. Montgomery Regents Professor of Industrial Engineering and Statistics ASU Foundation Professor of Engineering Arizona State University Bradley Jones
More informationChapter 5 Introduction to Factorial Designs
Chapter 5 Introduction to Factorial Designs 5. Basic Definitions and Principles Stud the effects of two or more factors. Factorial designs Crossed: factors are arranged in a factorial design Main effect:
More information23. Fractional factorials  introduction
173 3. Fractional factorials  introduction Consider a 5 factorial. Even without replicates, there are 5 = 3 obs ns required to estimate the effects  5 main effects, 10 two factor interactions, 10 three
More informationThird European DOE User Meeting, Luzern 2010
Correcting for multiple testing Old tricks, stateoftheart solutions and common sense Ivan Langhans CQ Consultancy CQ Consultancy Expertise in Chemometrics; the interaction between Chemistry and Statistics
More informationCHAPTER 6 A STUDY ON DISC BRAKE SQUEAL USING DESIGN OF EXPERIMENTS
134 CHAPTER 6 A STUDY ON DISC BRAKE SQUEAL USING DESIGN OF EXPERIMENTS 6.1 INTRODUCTION In spite of the large amount of research work that has been carried out to solve the squeal problem during the last
More informationCHAPTER EIGHT Linear Regression
7 CHAPTER EIGHT Linear Regression 8. Scatter Diagram Example 8. A chemical engineer is investigating the effect of process operating temperature ( x ) on product yield ( y ). The study results in the following
More informationTwoLevel Fractional Factorial Design
TwoLevel Fractional Factorial Design Reference DeVor, Statistical Quality Design and Control, Ch. 19, 0 1 Andy Guo Types of Experimental Design Paralleltype approach Sequentialtype approach Onefactor
More informationStatistics 512: Solution to Homework#11. Problems 13 refer to the soybean sausage dataset of Problem 20.8 (ch21pr08.dat).
Statistics 512: Solution to Homework#11 Problems 13 refer to the soybean sausage dataset of Problem 20.8 (ch21pr08.dat). 1. Perform the twoway ANOVA without interaction for this model. Use the results
More information2.830J / 6.780J / ESD.63J Control of Manufacturing Processes (SMA 6303) Spring 2008
MIT OpenCourseWare http://ocw.mit.edu 2.830J / 6.780J / ESD.63J Control of Processes (SMA 6303) Spring 2008 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
More informationRegression. Marc H. Mehlman University of New Haven
Regression Marc H. Mehlman marcmehlman@yahoo.com University of New Haven the statistician knows that in nature there never was a normal distribution, there never was a straight line, yet with normal and
More informationSMAM 319 Exam1 Name. a B.The equation of a line is 3x + y =6. The slope is a. 3 b.3 c.6 d.1/3 e.1/3
SMAM 319 Exam1 Name 1. Pick the best choice. (10 points2 each) _c A. A data set consisting of fifteen observations has the five number summary 4 11 12 13 15.5. For this data set it is definitely true
More informationSMA 6304 / MIT / MIT Manufacturing Systems. Lecture 10: Data and Regression Analysis. Lecturer: Prof. Duane S. Boning
SMA 6304 / MIT 2.853 / MIT 2.854 Manufacturing Systems Lecture 10: Data and Regression Analysis Lecturer: Prof. Duane S. Boning 1 Agenda 1. Comparison of Treatments (One Variable) Analysis of Variance
More informationDesign & Analysis of Experiments 7E 2009 Montgomery
1 What If There Are More Than Two Factor Levels? The ttest does not directly apply ppy There are lots of practical situations where there are either more than two levels of interest, or there are several
More informationLecture 10: 2 k Factorial Design Montgomery: Chapter 6
Lecture 10: 2 k Factorial Design Montgomery: Chapter 6 Page 1 2 k Factorial Design Involving k factors Each factor has two levels (often labeled + and ) Factor screening experiment (preliminary study)
More informationAPPENDIX 1. Binodal Curve calculations
APPENDIX 1 Binodal Curve calculations The weight of salt solution necessary for the mixture to cloud and the final concentrations of the phase components were calculated based on the method given by HattiKaul,
More informationUNIVERSITY OF CALIFORNIA College of Engineering Department of Electrical Engineering and Computer Sciences. PROBLEM SET No. 5 Official Solutions
1 UNIVERSITY OF CALIFORNIA College of Engineering Department of Electrical Engineering and Computer Sciences C. SPANOS Special Issues in Semiconductor Manufacturing EECS 290H Fall 0 PROBLEM SET No. Official
More informationUnit 9: Confounding and Fractional Factorial Designs
Unit 9: Confounding and Fractional Factorial Designs STA 643: Advanced Experimental Design Derek S. Young 1 Learning Objectives Understand what it means for a treatment to be confounded with blocks Know
More informationChapter 4: Randomized Blocks and Latin Squares
Chapter 4: Randomized Blocks and Latin Squares 1 Design of Engineering Experiments The Blocking Principle Blocking and nuisance factors The randomized complete block design or the RCBD Extension of the
More information2.830 Homework #6. April 2, 2009
2.830 Homework #6 Dayán Páez April 2, 2009 1 ANOVA The data for four different lithography processes, along with mean and standard deviations are shown in Table 1. Assume a null hypothesis of equality.
More informationData Set 1A: Algal Photosynthesis vs. Salinity and Temperature
Data Set A: Algal Photosynthesis vs. Salinity and Temperature Statistical setting These data are from a controlled experiment in which two quantitative variables were manipulated, to determine their effects
More information2010 StatEase, Inc. Dual Response Surface Methods (RSM) to Make Processes More Robust* Presented by Mark J. Anderson (
Dual Response Surface Methods (RSM) to Make Processes More Robust* *Posted at www.statease.com/webinar.html Presented by Mark J. Anderson (Email: Mark@StatEase.com ) Timer by Hank Anderson July 2008 Webinar
More informationAnswer all questions from part I. Answer two question from part II.a, and one question from part II.b.
B203: Quantitative Methods Answer all questions from part I. Answer two question from part II.a, and one question from part II.b. Part I: Compulsory Questions. Answer all questions. Each question carries
More information3. Factorial Experiments (Ch.5. Factorial Experiments)
3. Factorial Experiments (Ch.5. Factorial Experiments) HaeJin Choi School of Mechanical Engineering, ChungAng University DOE and Optimization 1 Introduction to Factorials Most experiments for process
More informationMEMORIAL UNIVERSITY OF NEWFOUNDLAND DEPARTMENT OF MATHEMATICS AND STATISTICS FINAL EXAM  STATISTICS FALL 1999
MEMORIAL UNIVERSITY OF NEWFOUNDLAND DEPARTMENT OF MATHEMATICS AND STATISTICS FINAL EXAM  STATISTICS 350  FALL 1999 Instructor: A. Oyet Date: December 16, 1999 Name(Surname First): Student Number INSTRUCTIONS
More information2 k, 2 k r and 2 kp Factorial Designs
2 k, 2 k r and 2 kp Factorial Designs 1 Types of Experimental Designs! Full Factorial Design: " Uses all possible combinations of all levels of all factors. n=3*2*2=12 Too costly! 2 Types of Experimental
More informationSTA 302 H1F / 1001 HF Fall 2007 Test 1 October 24, 2007
STA 302 H1F / 1001 HF Fall 2007 Test 1 October 24, 2007 LAST NAME: SOLUTIONS FIRST NAME: STUDENT NUMBER: ENROLLED IN: (circle one) STA 302 STA 1001 INSTRUCTIONS: Time: 90 minutes Aids allowed: calculator.
More information2.830J / 6.780J / ESD.63J Control of Manufacturing Processes (SMA 6303) Spring 2008
MIT OpenCourseWare http://ocw.mit.edu 2.830J / 6.780J / ESD.63J Control of Processes (SMA 6303) Spring 2008 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
More informationOPTIMIZATION OF FIRST ORDER MODELS
Chapter 2 OPTIMIZATION OF FIRST ORDER MODELS One should not multiply explanations and causes unless it is strictly necessary William of Bakersville in Umberto Eco s In the Name of the Rose 1 In Response
More informationStatistics GIDP Ph.D. Qualifying Exam Methodology May 26 9:00am1:00pm
Statistics GIDP Ph.D. Qualifying Exam Methodology May 26 9:00am1:00pm Instructions: Put your ID (not name) on each sheet. Complete exactly 5 of 6 problems; turn in only those sheets you wish to have graded.
More informationLecture 11: Simple Linear Regression
Lecture 11: Simple Linear Regression Readings: Sections 3.13.3, 11.111.3 Apr 17, 2009 In linear regression, we examine the association between two quantitative variables. Number of beers that you drink
More informationReference: CHAPTER 7 of Montgomery(8e)
Reference: CHAPTER 7 of Montgomery(8e) 60 Maghsoodloo BLOCK CONFOUNDING IN 2 k FACTORIALS (k factors each at 2 levels) It is often impossible to run all the 2 k observations in a 2 k factorial design (or
More informationConfidence Interval for the mean response
Week 3: Prediction and Confidence Intervals at specified x. Testing lack of fit with replicates at some x's. Inference for the correlation. Introduction to regression with several explanatory variables.
More informationLINEAR REGRESSION ANALYSIS. MODULE XVI Lecture Exercises
LINEAR REGRESSION ANALYSIS MODULE XVI Lecture  44 Exercises Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur Exercise 1 The following data has been obtained on
More informationStat664 Homework #3 due April 21 Turn in problems marked with
Stat664 Homework #3 due April 2 Turn in problems marked with Note: Whenever appropriate/possible, inspect the possibilities for need of transformation, determine suitable power (considering only power
More informationTable 1: Fish Biomass data set on 26 streams
Math 221: Multiple Regression S. K. Hyde Chapter 27 (Moore, 5th Ed.) The following data set contains observations on the fish biomass of 26 streams. The potential regressors from which we wish to explain
More informationExperimental design (KKR031, KBT120) Tuesday 11/ :3013:30 V
Experimental design (KKR031, KBT120) Tuesday 11/1 20118:3013:30 V Jan Rodmar will be available at ext 3024 and will visit the examination room ca 10:30. The examination results will be available for
More informationIE 361 Exam 1 October 2004 Prof. Vardeman
October 5, 004 IE 6 Exam Prof. Vardeman. IE 6 students Demerath, Gottschalk, Rodgers and Watson worked with a manufacturer on improving the consistency of several critical dimensions of a part. One of
More informationConfidence Intervals, Testing and ANOVA Summary
Confidence Intervals, Testing and ANOVA Summary 1 One Sample Tests 1.1 One Sample z test: Mean (σ known) Let X 1,, X n a r.s. from N(µ, σ) or n > 30. Let The test statistic is H 0 : µ = µ 0. z = x µ 0
More informationDensity Temp vs Ratio. temp
Temp Ratio Density 0.00 0.02 0.04 0.06 0.08 0.10 0.12 Density 0.0 0.2 0.4 0.6 0.8 1.0 1. (a) 170 175 180 185 temp 1.0 1.5 2.0 2.5 3.0 ratio The histogram shows that the temperature measures have two peaks,
More informationStrategy of Experimentation II
LECTURE 2 Strategy of Experimentation II Comments Computer Code. Last week s homework Interaction plots Helicopter project +1 1 1 +1 [4I 2A 2B 2AB] = [µ 1) µ A µ B µ AB ] +1 +1 1 1 +1 1 +1 1 +1 +1 +1 +1
More informationMultiple Regression Examples
Multiple Regression Examples Example: Tree data. we have seen that a simple linear regression of usable volume on diameter at chest height is not suitable, but that a quadratic model y = β 0 + β 1 x +
More informationSizing Mixture (RSM) Designs for Adequate Precision via Fraction of Design Space (FDS)
for Adequate Precision via Fraction of Design Space (FDS) Pat Whitcomb StatEase, Inc. 61.746.036 036 fax 61.746.056 pat@statease.com Gary W. Oehlert School of Statistics University it of Minnesota 61.65.1557
More informationA discussion on multiple regression models
A discussion on multiple regression models In our previous discussion of simple linear regression, we focused on a model in which one independent or explanatory variable X was used to predict the value
More informationLec 5: Factorial Experiment
November 21, 2011 Example Study of the battery life vs the factors temperatures and types of material. A: Types of material, 3 levels. B: Temperatures, 3 levels. Example Study of the battery life vs the
More informationResponse Surface Methodology
Response Surface Methodology Process and Product Optimization Using Designed Experiments Second Edition RAYMOND H. MYERS Virginia Polytechnic Institute and State University DOUGLAS C. MONTGOMERY Arizona
More informationStatistics GIDP Ph.D. Qualifying Exam Methodology May 26 9:00am1:00pm
Statistics GIDP Ph.D. Qualifying Exam Methodology May 26 9:00am1:00pm Instructions: Put your ID (not name) on each sheet. Complete exactly 5 of 6 problems; turn in only those sheets you wish to have graded.
More informationSession 3 Fractional Factorial Designs 4
Session 3 Fractional Factorial Designs 3 a Modification of a Bearing Example 3. Fractional Factorial Designs Twolevel fractional factorial designs Confounding Blocking TwoLevel Eight Run Orthogonal Array
More informationHowever, this definition can be expanded to include: biology (biometrics), environmental science (environmetrics), economics (econometrics).
Chemometrics Application of mathematical, statistical, graphical or symbolic methods to maximize chemical information. However, this definition can be expanded to include: biology (biometrics), environmental
More informationApplication of Full Factorial Design for Optimization of Feed Rate of Stationary Hook Hopper
Application of Full Factorial Design for Optimization of Feed Rate of Stationary Hook Hopper Astha Kukreja, Pankaj Chopra, Akshay Aggarwal, and Pradeep Khanna Abstract the work aims at the optimization
More informationCh 2: Simple Linear Regression
Ch 2: Simple Linear Regression 1. Simple Linear Regression Model A simple regression model with a single regressor x is y = β 0 + β 1 x + ɛ, where we assume that the error ɛ is independent random component
More informationCHAPTER 6 MACHINABILITY MODELS WITH THREE INDEPENDENT VARIABLES
CHAPTER 6 MACHINABILITY MODELS WITH THREE INDEPENDENT VARIABLES 6.1 Introduction It has been found from the literature review that not much research has taken place in the area of machining of carbon silicon
More informationAssociate Professor, Department of Mechanical Engineering, Birla Institute of Technology, Mesra, Ranchi, India
Optimization of machine process parameters on material removal rate in EDM for EN19 material using RSM Shashikant 1, Apurba Kumar Roy 2, Kaushik Kumar 2 1 Research Scholar, Department of Mechanical Engineering,
More informationInference for Regression Inference about the Regression Model and Using the Regression Line, with Details. Section 10.1, 2, 3
Inference for Regression Inference about the Regression Model and Using the Regression Line, with Details Section 10.1, 2, 3 Basic components of regression setup Target of inference: linear dependency
More informationDESAIN EKSPERIMEN BLOCKING FACTORS. Semester Genap 2017/2018 Jurusan Teknik Industri Universitas Brawijaya
DESAIN EKSPERIMEN BLOCKING FACTORS Semester Genap Jurusan Teknik Industri Universitas Brawijaya Outline The Randomized Complete Block Design The Latin Square Design The GraecoLatin Square Design Balanced
More informationA Sample Durability Study of a Circuit Board under Random Vibration and Design Optimization
A Sample Durability Study of a Circuit Board under Random Vibration and Design Optimization By: MS.ME Ahmad A. Abbas Ahmad.Abbas@AdvancedCAE.com www.advancedcae.com Sunday, March 07, 2010 Advanced CAE
More informationStatistical Modelling in Stata 5: Linear Models
Statistical Modelling in Stata 5: Linear Models Mark Lunt Arthritis Research UK Epidemiology Unit University of Manchester 07/11/2017 Structure This Week What is a linear model? How good is my model? Does
More informationLecture 1 Linear Regression with One Predictor Variable.p2
Lecture Linear Regression with One Predictor Variablep  Basics  Meaning of regression parameters p  β  the slope of the regression line it indicates the change in mean of the probability distn of
More information