Florida State University Libraries
Electronic Theses, Treatises and Dissertations The Graduate School
2006 Design and Analysis of Response Surface Designs with Restricted Randomization Wayne R. Wesley
Follow this and additional works at the FSU Digital Library. For more information, please contact [email protected] THE FLORIDA STATE UNIVERSITY
COLLEGE OF ENGINEERING
DESIGN AND ANALYSIS OF RESPONSE SURFACE DESIGNS WITH
RESTRICTED RANDOMIZATION
By
Wayne R. Wesley
A Dissertation submitted to the Department of Industrial Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Degree Awarded: Summer Semester, 2006
Copyright © 2006 Wayne R. Wesley All Rights Reserved The members of the Committee approve the dissertation of Wayne R. Wesley defended on July 6, 2006.
______James R. Simpson Professor Directing Dissertation
______Anuj Srivastava Outside Committee Member
______Peter A. Parker Outside Committee Member
______Joseph J. Pignatiello, Jr. Committee Member
Approved:
______Chuck Zhang, Chair, Department of Industrial and Manufacturing Engineering
______Ching-Jen Chen, Dean, College of Engineering
The Office of Graduate Studies has verified and approved the above named committee members.
ii
I dedicate this dissertation to my son and daughter Jowayne and Jowaynah Wesley.
iii ACKNOWLEDGEMENTS
I must first give thanks to the almighty God for good health, strength and wisdom to successfully complete my dissertation. It is with a deep sense of gratitude that I express my sincere appreciation to Dr. James Simpson and Dr. Peter Parker for their invaluable advice and contributions in directing the dissertation. Special thanks to Dr. Joseph Pignatiello and Dr. Anuj Srivastava for their meaningful comments and suggestions that helped to improve the quality of the dissertation. Additionally, I would like to give special recognition to my wife Joy Wesley for her unwavering support and commitment throughout my endeavors at The Florida State University. Special thanks to my mother Mary White for her tremendous sacrifice in giving me the opportunity to develop and pursue my career goals. To my fellow students and friends, Michelle, Lisa, Francisco, Rupert and Young-Guo thanks for your support and encouragement. Finally thanks to all those who supported me with their prayers, financially and otherwise.
iv TABLE OF CONTENTS
LIST OF TABLES...... viii
LIST OF FIGURES ...... xi
ABSTRACT...... xii
CHAPTER 1 ...... 1
1.0 INTRODUCTION ...... 1 1.1 Background...... 1 1.2 The Split Plot Model...... 4 1.3 Design Optimality...... 6 1.4 Problem Statement...... 9 1.5 Research Objectives...... 10 CHAPTER 2 ...... 11
2.0 REVIEW OF LITERATURE ...... 11 2.1 Introduction...... 11 2.2 First-Order Split Plot Designs...... 12 2.3 Second-Order Split Plot Designs ...... 14 2.4 Design Optimality Criteria...... 16 2.5 Summary...... 18 CHAPTER 3 ...... 19
3.0 Impact of Restricted Randomization on the Structure of the Information Matrix 19 3.1 Abstract...... 19 3.2 Introduction...... 20 3.3 The Split Plot Model...... 20 3.4 Structure of Second-Order SPD...... 24 3.5 Critical Functions of the Variance Ratio and Whole Plot Sizes ...... 28 3.6 Information Matrix for CCD and BBD Structures ...... 29 3.7 Impact of Design Structure on the Application of the Critical Functions 33 3.8 Outline of Analytical Characterization Scheme ...... 38 −1 3.9 Characterization of (XR' −1 X) for Split Plot CCD...... 40 −1 3.10 Characterization of (XR' −1 X) for Split Plot BBD...... 45 3.11 Generalized Variance of Parameter Estimates...... 46 3.11.1 Determinant for split plot CCD...... 47
v 3.11.2 Determinant for split plot BBD...... 47 3.12 Application...... 48 3.13 Conclusion ...... 54 CHAPTER 4 ...... 56
4.0 Prediction Variance Properties and G-Criterion Location for Second-Order Split Plot Designs ...... 56 4.1 Abstract...... 56 4.2 Introduction...... 57 4.3 The Split Plot Model...... 58 4.4 Prediction Optimality Criteria...... 60 4.5 Second-Order Split Plot Designs ...... 61 4.6 Critical functions of the variance ratio and whole plot sizes...... 62 4.7 Information Matrix for the Split Plot CCD...... 63 4.8 Information Matrix for the Split Plot BBD...... 67 4.9 Minimum and Maximum Prediction Variance ...... 70 4.10 Vminρ and Vmaxρ for Spherical Regions ...... 72 4.11 Vmin and Vmax for Whole Plot Design Space for Spherical Regions ρz ρz ...... 73 4.12 Vmin and Vmax for Subplot Design Space for Spherical Regions .... 74 ρx ρx
4.13 Vminρ and Vmaxρ for Combined Design Space for Spherical Regions.... 74 4.14 Vminρ and Vmaxρ for Cuboidal Regions...... 75 4.15 Vmin and Vmax for Whole Plot Design Space for Cuboidal Regions 76 ρz ρz 4.16 Vmin and Vmax for Subplot Design Space for Cuboidal Regions.... 76 ρx ρx
4.17 Vminρ and Vmaxρ for the Combined Design Space for Cuboidal Regions77 4.18 Delta G-Criterion Measures for Whole Plot and Subplot Design Spaces 79 4.19 Prediction Variance Assessment of Second-Order SPD...... 79 4.19.1 CCD with one whole plot factor and two subplot factors...... 79 4.19.2 BBD with two whole plot factors and two subplot factors...... 85 4.19.3 FCC with two whole plot factors and three subplots factors...... 88 4.20 Conclusion ...... 91 CHAPT ER 5 ...... 93
5.0 Integrated Prediction Variance for Response Surface Designs ...... 93 5.1 Abstract...... 93 5.2 Introduction...... 94 5.3 The Split Plot Model...... 95 5.4 IV – Optimality Criterion...... 97 −1 5.5 Analytical Characterization of (XR' −1 X) ...... 100 5.5.1 Critical functions of the variance ratio and whole plot sizes...... 100 5.5.2 Information Matrix for the Split Plot CCD...... 101 5.5.3 Information Matrix for the Split Plot BBD...... 105
vi 5.6 Matrix of Region Moments...... 108 5.7 Analytical Determination of Integrated Prediction Variance ...... 110 5.8 Integrated Variance for the Split Plot CCD ...... 111 5.9 Integrated Variance for the Split Plot BBD ...... 112 5.10 Integrated Prediction Variance for Split Plot Designs...... 113 5.10.1 SPD with one whole plot factor and three subplot factors...... 114 5.10.2 SPD with two whole plot factors and four subplot factors ...... 117 5.10.3 SPD with three whole plot factors and four subplot factors ..... 122 5.11 Integrated Prediction Variance for Completely Randomized Designs... 126 5.11.1 Hybrid and Small Composite Designs...... 127 5.11.2 Equiradial Designs...... 129 5.12 Comparison of Selected Cuboidal Designs...... 132 5.13 Conclusion ...... 134 CHAPT ER 6 ...... 136
6.0 CONCLUSION...... 136 6.1 Research Objective 1 ...... 137 6.2 Research Objective 2 ...... 137 6.3 Research Objective 3 ...... 138 6.4 Future Work...... 139 APPENDIX A...... 140 A1: Analytical Characterization of (X'R-1X)-1 for Split Plot CCD ...... 140 A2: Analytical Characterization of (X'R-1X)-1 for Split Plot BBD ...... 152 APPENDIX B ...... 163 B1: Determinant of the Information Matrix...... 163 B2: Derivation of the Determinant for Split Plot CCD...... 164 B3: Derivation of the Determinant for Split Plot BBD...... 166 APPENDIX C ...... 168 C1: Analytical Determination of v(z, x) for Split Plot CCD...... 168 C2: Analytical Determination of v(z, x) for Split Plot BBD...... 170 APPENDIX D...... 172 D1: Determination of Critical Points ...... 172 D2: Evaluation of Vminρ and Vmaxρ for Spherical Regions...... 174 D3: Evaluation of Vminρ and Vmax ρ for Cuboidal Regions ...... 177 APPENDIX E ...... 181
E1: Derivation of trace (Q11) and trace (Q22) for Split Plot CCD...... 181 E2: Derivation of trace (Q11) and trace (Q22) for Split Plot BBD...... 185 REFERENCES ...... 188
BIOGRAPHICAL SKETCH ...... 193
vii
LIST OF TABLES
Table 1: The X matrix for a CCD with one whole plot and two subplot factors (VKM D12)...... 26 Table 2: The X matrix for a balanced BBD with one whole plot and two subplot factors (VKM D12) ...... 27 Table 3: The X matrix for an unbalanced BBD with one whole plot and two subplot factors (VKM D12) ...... 27 Table 4: Table of notations for the split plot CCD ...... 30 Table 5: Table of notations for the split plot BBD ...... 31 Table 6: A resolution V CCD with = α = 1, two whole plots and three subplot factors (VKM) ...... 35
Table 7: A CCD with = α = 5 , one whole plot factor and three subplot factors (MWP) ...... 36 Table 8: A CCD with = α = 1, one whole plot factor and three subplot factors (MWP)37 Table 9: Table of scalar quantities for the split plot CCD ...... 44 Table 10: Table of constant quantities for the split plot CCD ...... 44 Table 11: Table of scalar quantities for the split plot BBD ...... 45 Table 12: Table of constant quantities for the split plot BBD ...... 46 Table 13: Modified VKM D23 with = α = 1, two whole plots and three subplot factors – Unbalanced...... 48 Table 14: Design parameters for the split plot CCD...... 49 Table 15: Constants, functions and D-value results for CCD across the different variance ratios ...... 50 Table 16: Relative efficiencies...... 51 Table 17: Design parameters for the BBD...... 52 Table 18: Constants, functions and D-value results for BBD across the different variance ratios ...... 53 Table 19: Table of notations for the split plot CCD ...... 64 Table 20: Table of scalar quantities for the split plot CCD ...... 66 Table 21: Table of constant quantities for the split plot CCD ...... 66 Table 22: Table of notations for the split plot BBD ...... 67
viii Table 23: Table of scalar quantities for the split plot BBD ...... 69 Table 24: Table of constant quantities for the split plot BBD ...... 69 Table 25: Table of coefficients of the variance function v(z, x) for CCD and BBD...... 72 Table 26: Critical Points for the evaluation of v(z, x)...... 73
Table 27: Standard CCD with = α = k and N=17...... 80 Table 28: Prediction variance properties and G-criterion location...... 80 Table 29: Relative measures of whole plot and subplot design spaces ...... 82
Table 30: Modified CCD with = α = k and N=16 ...... 82 Table 31: Prediction variance properties and G-criterion location for the modified CCD83 Table 32: Relative measures of whole plot and subplot design spaces for modified CCD ...... 85 Table 33: MWP_BBD with two whole plot factors and two subplot factors ...... 85 Table 34: Prediction variance properties and G-criterion location for the MWP BBD with two whole plot factors and two subplot factors...... 86 Table 35: Relative measures of whole plot and subplot design spaces MWP BBD with two whole plot factors and two subplot factors...... 88 Table 36: A resolution V VKM FCC with = α = 1, two whole plots and three subplot factors ...... 88 Table 37: Prediction variance properties and G-criterion location for the VKM FCC with two whole plot factors and three subplot factors...... 89 Table 38: Relative measures of whole plot and subplot design spaces for the VKM FCC ...... 91 Table 39: Table of notations for the split plot CCD ...... 102 Table 40: Table of scalar quantities for the split plot CCD ...... 104 Table 41: Table of constant quantities for CCD ...... 104 Table 42: Table of notations for the split plot BBD ...... 105 Table 43: Table of scalar quantities for the split plot BBD ...... 107 Table 44: Table of constant quantities for the split plot BBD ...... 107 Table 45: Spherical second and forth order region moments for 26≤ k ≤ ...... 110 Table 46: Split plot designs selected for evaluation ...... 114 Table 47: Design matrix for a balanced VKM BBD with one whole plot factor and three subplot factors ...... 115 Table 48: Design matrix for a balanced MWP BBD with one whole plot factor and three subplot factors ...... 115
ix Table 49: Design parameters for VKM BBD and MWP BBD with one whole plot factor and three subplot factors...... 116 Table 50: Integrated prediction variance for VKM BBD and MWP BBD with one whole plot factor and three subplot factors ...... 117 Table 51: Design matrix for a balanced VKM CCD with = α = 2.4495, two whole plot factors and four subplot factors ...... 118 Table 52: Design matrix for a balanced MWP CCD with = α = 2.4495, two whole plot factors and four subplot factors ...... 119 Table 53: Design parameters for VKM CCD and MWP CCD with = α = 2.4495, two whole plot factors and four subplot factors ...... 120 Table 54: Integrated prediction variance for VKM CCD and MWP CCD with = α = 2.4495, two whole plot factors and four subplot factors ...... 121 Table 55: Design matrix for a balanced VKM BBD with two whole plot factors and four subplot factors ...... 122 Table 56: Integrated prediction variance for VKM BBD with two whole plot factors and four subplot factors...... 122 Table 57: Design matrix for a balanced VKM CCD with = α = 1, three whole plot factors and four subplot factors ...... 123 Table 58: Design matrix for a balanced MWP CCD with = α = 1, three whole plot factors and four subplot factors ...... 124 Table 59: Design parameters for VKM CCD and MWP CCD with = α = 1, three whole plot factors and four subplot factors...... 125 Table 60: Integrated prediction variance for VKM CCD and MWP CCD with = α = 1, 3 whole plot factors and 4 subplot factors...... 126 Table 61: Integrated prediction variance for standard CCD and BBD spherical designs...... 127 Table 62: Comparison of IV values for near-saturated designs over spherical regions for k = 3 and 4...... 128 Table 63: Comparison of IV values for Hybrid and CCD designs over spherical regions for k = 6 and 7...... 129 Table 64: Integrated variances of selected equiradial designs...... 131 Table 65: Maximum prediction variance properties of selected equiradial designs for k = 2 factors ...... 131 Table 66: Integrated prediction properties of selected cuboidal designs...... 133
x LIST OF FIGURES
Figure 1: Typical assignment of subplot factors to whole plots ...... 3 Figure 2: Analytical Characterization Scheme ...... 40 Figure 3: 3D VDGs for the standard CCD with one whole plot factor and two subplot factors ...... 81 Figure 4: Comparison of whole plot and subplot spaces for the standard CCD with one whole plot factor and two subplot factors ...... 81 Figure 5: 3D VDGs for the modified CCD with one whole plot factor and two subplot factors ...... 83 Figure 6: Comparison of whole plot and subplot spaces for modified CCD with one whole plot factor and two subplot factors ...... 84 Figure 7: 3D VDGs the MWP BBD with two whole plot factors and two subplot factors...... 86 Figure 8: Comparison of whole plot and subplot spaces for the MWP BBD with two whole plot factors and two subplot factors...... 87 Figure 9: 3D VDG the VKM FCC with two whole plot factors and three subplot factors...... 89 Figure 10: Comparison of whole plot and subplot spaces for the MWP BBD with two whole plot factors and two subplot factors...... 90
xi ABSTRACT
Many industrial experiments are conducted under various conditions which do not facilitate complete randomization of all the experimental factors. In response surface methodology whenever there are restrictions on randomization the experimental procedure usually follows the split plot design approach. Split plot designs are used when there are factors which are difficult or costly to change or adjust during an experiment. Split plot designs are currently generating renewed interest because of their usefulness and practical application in industrial settings. Despite the work accomplished through various research efforts, there is still a need to understand the optimality properties of these designs for second-order response surface models. This dissertation provides the development of an analytical approach for the computation of various optimality properties for the assessment of second-order split plot designs. The approach involves a thorough investigation of the impact of restricted randomization on the information matrix, which characterizes much of the relationship between the design points and the proposed response surface model for split plot designs. Several important insights are presented for the construction of second-order split plot designs. In addition, the analytical equations reported compute exact design optimality values and are more efficient than currently available methods. A particular feature of these analytical equations is that they are functions of the design parameters, radius and variance ratio. Further, a significant result is the ability to efficiently compute the exact value of the integrated prediction variance for both split plot designs and completely randomized designs. The functionality of the computational procedures presented provides easy evaluation of the impact of changes in the design structure and variance ratio on the optimality properties of second-order split plot designs.
xii CHAPTER 1
1.0 INTRODUCTION
1.1 Background The three fundamental principles in design and analysis of experiments are randomization, replication and blocking. However, in most industrial experiments it becomes a challenge to adhere to these principles at all times. In situations where it is difficult, expensive or dangerous to adjust certain factors during experimentation it becomes necessary to reduce the frequency of adjusting this factor(s). The reduction in the frequency of adjusting certain factors during experimentation is referred to as restricted randomization. In statistical analysis, designs with restricted randomization are commonly referred to as split plot designs (SPD). Wooding (1973) presented a detailed description on the structure and use of SPD. According to Myers et al (2004), the foundation for response surface methodology was laid by Box and Hunter (1951). Response surface methodology (RSM) is a collection of statistical and mathematical techniques used for the characterization and optimization of systems or processes. In traditional response surface methodology experimentation is generally expected to be completely randomized. Therefore, design optimality criteria were developed based on the assumption that the design will be completely randomized. Unfortunately, there are situations which arise in industrial experimentation that will not facilitate complete randomization of the experimental runs or resetting of the factor levels for each experimental run. For example the magnetic alignment of carbon nanotubes is dependent on several factors. These factors are: 1. Strength of magnetic field 2. Temperature 3. Nanotube dispersion 4. Filtration flow rate 5. Aspect ratio (length/diameter)
1 The strength of the magnetic field can not be adjusted quickly due to the dangers associated with rapidly increasing or decreasing the electrical voltage required. Therefore, the magnetic field will have to follow a different randomization procedure than the remaining four factors. The experiment will have to be conducted such that the levels of the magnetic field are adjusted less frequently than the other four factors. A consequence of restricted randomization is a complicated error structure resulting from a split plot design. Therefore, established design optimality criteria for evaluating response surface designs have to be altered to accommodate the impact of the two error sources on the information matrix. Box and Hunter (1951) defined the information matrix as the sums of squares and products of the independent variables. Deeply embedded in the information matrix is the design of the experiment. The issue of randomization restriction in response surface methods (RSM) was investigated by Letsinger et al (1996). Based on a simulation study Letsinger et al (1996) recommended that the restricted maximum likelihood method (REML) is better for parameter estimation when compared to ordinary least squares (OLS) and iterated reweighed least squares (IRLS). Recently, the split plot design has been given much attention by researchers and continues to generate much interest as it is becoming increasingly popular in industrial experimentation. Most industrial experiments follow the split plot structure but in many cases are incorrectly analyzed as completely randomized experiment. SPDs are used when there are factors which are difficult or costly to change or adjust during an experiment. The experimental approach is to separate the design into two design spaces known as the whole plot, for hard to change factors, and the subplot, for easy to change factors. Applying this experimental approach to the magnetic alignment example, the classification of factors would be as follows: 1. Whole plot factor – Dangerous to change levels z → Magnetic field 2. Subplot factors – Safe and easy to change
x1 → Temperature
x2 → Filtration flow rate
x3 → Suspension concentration
2 x4 → Aspect ratio
Therefore if there are three levels of the whole plot factor (z) then the subplot factors (x1
x2, x3 and x4) would be assigned to each of the factor level settings for the whole plot factor. Each whole plot can only have one whole plot factor level setting. Figure 1 shows a typical assignment of the subplot factors to the whole plots.
Subplot x1 Whole plot 1 z - level 1 x2 x3 x4
Subplot x1 Whole plot 2 z - level 2 x2 x3 x4
Subplot x1 z - level 3 x Whole plot 3 2 x3 x4
Figure 1: Typical assignment of subplot factors to whole plots
The analysis of this experiment would therefore involve two different error terms to reflect the two separate experimental units. The experimental error associated with the whole plot partitioning is referred to as the whole plot error. Similarly, the experimental error associated with the subplot partitioning is referred to as the subplot error. These error terms are used to develop empirical models for characterizing the system or process being studied. The following section will present the details of the split plot model.
3 1.2 The Split Plot Model The general form for split plot models in matrix form is given as
yX= Θ ++ where y is an N x 1 vector of expected responses, N is the total number of observations, Θ is a p x 1 vector of unknown model parameters, p is the number of model parameters, X is the N x p matrix of the levels of the independent variables, and and are N x 1 vectors of random variables for whole plot and subplot errors respectively. N and It is assumed that + ~ N (0, Σ). The structure of the variance-covariance matrix Σ is
⎡⎤Σ1 00...... ⎢⎥ 0 Σ2 0... 0 ⎢⎥ (1.1) Σ = ⎢ 0 ⎥ ⎢⎥ ⎢⎥ 0 ⎢⎥ ⎣⎦00 0Σa
22' where a is for the number of whole plots, Σiwww=+σσδε11×11×× Iw and w is the number of th subplot runs (whole plot size). In matrix form the structure of Σ i for the i whole plot is
222 2 ⎡⎤σσσδεδ+ σ δ ⎢⎥ σ 2 Σ = ⎢δ ⎥ (1.2) i ⎢⎥2 σδ ⎢⎥2222 ⎣⎦⎢⎥σδ σσσδδε+
2 2 2 where σ δ + σ ε represent the variance of individual observations, σ δ is the covariance
2 and whole plot error variance while σ ε is the subplot error variance. Given the structure of the two error terms, estimation of the model parameters is accomplished using the generalized least squares (GLS) method of estimation given by
4 −1 Θˆ = (X''Σ−11XX) Σ− y
−1 ˆ ' −1 2 2 where Var ()Θ = ()X Σ X . The two variance components σ δ and σ ε are needed to obtain parameter estimates and standard errors. Letsinger et al. (1996) point out that design optimality criteria, which are functions of the Var ()Θˆ , depend on the structure of the design particularly the information matrix (A). In the case of completely randomize designs (CRD) the information matrix is given as X' X . However, for SPD alphabetic optimality is now dependent on an information matrix of the form X'Σ−1X , where Σ is the variance- covariance matrix described in Equation (1.1). This situation complicates the process because now the optimal design will depend on the variance ratio or degree of correlation
2 2 η = σ δ σ ε , given by the correlation matrix (R). The correlation matrix (R) of the
22 observations is derived by dividing the variance covariance matrix, Σ, by σ δ +σ ε . In the computation of the information matrix the correlation matrix is used because it provides a unitless measure of the information matrix. An information matrix without units is desirable especially for design comparison purposes. Therefore, the information matrix without units is given as
⎛⎞−1 ''⎛⎞Σ −1 AX==⎜⎟⎜⎟ X XRX ⎜⎟σσ22+ () ⎝⎠⎝⎠δε where, R denotes the correlation matrix of the observations and its structure is of the form
5 ⎡ ηη⎤ 1 ⎢ 11++η η ⎥ ⎡⎤R01 ...... 0 ⎢ ⎥ ⎢⎥⎢ η ⎥ 0R2 0... 0 ⎢⎥⎢1+η ⎥ ⎢ ⎥ R0==⎢⎥ where Ri . ⎢⎥⎢ η ⎥ 0 ⎢⎥⎢ 1+η ⎥ ⎢⎥00 0R ⎢ ⎥ ⎣⎦a ⎢ ηη⎥ ⎢ 1 ⎥ ⎣11++ηη⎦
In terms of design optimality, the effect of the correlation matrix on the information matrix will give different results for SPD when compared to CRD. According to Goos and Vandebroek (2001 and 2004) the difference in design optimality values should not be surpri sing because desig ning a completely randomized experiment consists of determining the design points, while designing a split plot experiment requires simultaneously choosing the number of whole plots and the number of subplots within each whole plot and the design points for each plot type. The next section presents some commonly used design optimality criteria to assess the performance of response surface designs.
1.3 Design Optimality The following discussion presents a brief review of the design optimality criteria used to evaluate response surface designs. For details on these design criteria see Montgomery (2001) and Myers and Montgomery (2002). The main purpose of this section is to illustrate how the impact of the variance ratio, through the correlation matrix, is incorporated into the functions of the design optimality criteria. The A-optimality criterion focus is to minimize the ave rage variance of the parameter estimates based on a pre-specified model. The focus is on individual variances of the coefficients of the model parameters and does not take into account the covariance among the coefficients of the model parameters. The A-optimality criterion deals with minimizing the trace of the inverse of information matrix and is given as
6 Min trace[]A −1 −1 Min trace ⎡XR' −1 X⎤ ⎣ ⎦
−1 where trace is the sum of the diagonal elements of (XR' −1 X) which are the variances of the estimator of the coefficients of the model parameters. D-optimality focuses on maximizing the determinant of the information matrix and is expressed as
1/ p Max XR' −1 X
or equivalently
1/ p −1 Min ()XR' −1 X .
where, p is the number of parameters being estimated in the model. The use of this criterion results in the minimization of the generalized variance of the parameter estimates. The G-optimality criterion is associated with the prediction variance v(z, x) and seeks to minim ize the maximum prediction variance over the region of interest (Ξ). The region of interest characterizes the boundary of the operating conditions for the system (s) under consideration. In response surface methodology the two most commonly encountered regions are the spherical and cuboidal regions. G-optimality emphasizes the use of design points for which the maximum v(z, x) in the region of the design is minimized. Thus a G optimal design (ξ ) is one in which we have
⎡ ⎤ Min Max v()zx, ξ ⎣⎢ x∈Ξ ⎦⎥ where v(z, x) is given as
Var⎡⎤yˆ (zx , ) −1 vfzx,,==⎣⎦ zx' XR' −1 Xf zx, () 22 ()()() σσδε+
7 where f ()zx, is the model vector which gives the location in the design space and reflects the nature of the model under consideration. The general form of the second- order model vector used is given as
f z,x ' = () ⎡ 2222⎤ ⎣1,,z1… zwp x11,,……………… xsp zz21 ,, z wp−− z wp zx11 ,, z wp x sp xx 121 ,, x sp x sp z1 ,, z wp x 1 ,, x sp ⎦ where, z and x are whole plot and subplot factors respectively. The model vector is partitioned into linear, interaction and quadratic terms for the respective whole plot and subplot factors. The integrated variance criterion (IV-optimality) also denoted as either I or Q and is used to compute a single measure of prediction performance through an average process. That is, the average prediction variance v(z, x) is determined over some region of interest, Ξ, by integrating over Ξ. The IV-optimality criterion for a design (ξ ) can be expressed as follows:
−1 ⎧⎫1 ' ' −1 Min⎨⎬ f()zx,, XR X f() zx d zx d ξ ∫ () ⎩⎭Ω Ξ
⎧⎫1 Min⎨⎬ v()zx, d zx d ξ ∫ ⎩⎭Ω Ξ
where Ω= ddzx is the volume of the region Ξ. The IV-optimality can be further ∫Ξ simplified to
−1 ⎧⎫1 ' −1 ' Min⎨⎬ traceXR X f()() z,, x f z x d z d x ξ ()∫ . ⎩⎭Ω Ξ
In contrast to the IV-optimality where the focus is on the entire design region, V- optimality focuses on the average prediction variance associated with a set of points which are of particular interest to the practitioner. This optimality criterion, in many
8 cases, is implemented as the spherical (average) prediction variance of the surface area. The surface area used is based on whether or not the region is spherical or cuboidal.
1.4 Problem Statement The information matrix is the single most useful source of information for evaluating experimental design. All aspects of model building and analysis are based on this powerful source of information. It is therefore logical that if one is performing design analysis or design construction that consideration be given to the nature and structure of the information matrix. Such consideration is particularly important for split plot designs due to the fact that its information matrix takes on an added dimension brought about by restricted randomization. Another issue of concern is that the D-optimality criterion has been more popular than other optimality criteria primarily because of its simplistic computation. Welch (1984) argues that the heavy focus on D-optimality is inappropriate, because in practice many experiments are aimed at predicting the response over the region of interest rather than parameter estimation. Haines (1987 ) suggests that the apparent lack of attention given to the other design optimality criteria, such as G-optimality and IV-optimality in particular, is due to the computation al time required to compute these values. According to the Kiefer-Wolfowicz (1959 and 1961) equivalence theorem D- optimal designs are also G-optimal, in the limit of an infinite number of trials for completely randomized designs. However, this equivalence theorem will not hold for split-plot experiments because th e information matrix is no longer the same due the restriction on randomization. Goos and Vandebroek (2004) stated that there is a need to find good standard designs for split plot experimentation. They suggest that investigations be conducted to establish the relationship between the D-optimality and G- optimality criteria in the context of split plot experimentation. Despite the work accomplished through various research efforts, there are still areas of concern that require further investigation to facilitate a deeper understanding of the impact of restricted randomization on response surface designs. It is anticipated that the results of such an investigation would help to create a general approach that addresses
9 the design and analysis of split plot experiments. It is the intention of this research to advance the knowledge required to achieve such an approach through the characterization of the information matrix and its inverse for split plot designs.
1.5 Research Objectives The purpose of this research is to: 1. Investigate the impact of restricted randomization on the structure of the information matrix. 2. Develop a computationally efficient method to compute maximum and minimum prediction variances and perform an assessment of the impact of changes in the variance ratio on these variance properties for second-order split plot designs. 3. Develop a computationally efficient method to compute integrated prediction variance and perform a comparative assessment of the integrated prediction variance properties of response surface designs.
The dissertation consists of the five remaining chapters. Chapter 2 provides a general review of literature covering pertinent issues relating to split plot designs. Chapters 3 through 5 are stand-alone chapters detailing the new methods developed for evaluating and assessing design optimality properties of second-order response surface split plot designs. Each chapter contains its own literature review, problem statement, research objective, methodology description, conclusion section. The final chapter of this dissertation is Chapter 6 which summarizes the conclusions drawn from the three previous chapters.
10 CHAPTER 2
2.0 REVIEW OF LITERATURE
2.1 Introduction Myers et al. (2004) reviewed the status of response surface methodologies. They summarized work published on split plot designs. They suggested that industrial experimentation in the presence of restrictions on randomization follow the split plot design approach. They concluded that attention must be given to the way in which these experiments are analyzed and that the basic split plot analysis techniques must be extended to different types of industrial experiments. The attention should be focused on providing proper analysis techniques that accounts for the multiple error terms associated with the split plot design. Such analysis techniques are crucial because if split plot designs are incorrectly analyzed as completely randomizes designs then the analysis of the model is subjected to Type 1 and Type 2 errors. Further, incorrect analyses may also result in incorrect variance estimates which will affect the accuracy of the results. Simpson et al. (2004) present a practical approach for the implementation of the split plot design in a factorial setting. They suggest that the combination of carefully structured split-plot designs and mixed model analysis can be a fairly effective scheme for investigations in randomization restricted environment. They further suggest that the development of a general approach to split plot must focus on four important issues when conducting designed experiments particularly for the split plot design. The issues are adequate replication, empirical model building capability, lack of fit and the ability to testing for curvature. The test for lack of fit requires the replication of both whole plot and subplot units.
11 2.2 First-Order Split Plot Designs The design and construction of split plot experiments has been studied extensively in the literature. Bingham and Sitter (2001) present some design issues with fractional factorial split plot experiment. A fractional factorial experiment is a design in which only a subset of the experimental runs are conducted. They discuss the impact of the split plot structure on estimation, precision and the use of resources. They point out that the two main issues that arise when selecting competing designs are identifiability and precision. Identifiability is the ability to estimate as many of the main effects and two factor interactions as possible and precision is the ability to detect significant effects with as much power as possible. Among the issues of concern are costs and run size. They conclude that the decision of where to fractionate comes down to a compromise between minimizing the aberration of the design, maximizing the precision for the effects of interest, and selecting the most economical design. In addition, Huang et al. (1998) and Bingham and Sitter (1999) present methods to construct minimum aberration two-level and two-level fractional factorial split plot designs respectively. Using these methods they have generated an extensive list of minimum aberration designs to assist in comparing these designs. Goos and Vandebroek (2001) also develop an exchange algorithm for constructing D-optimal split plot designs. The resulting split plot designs allow substantial increases in design efficiency. Goos and Vandebroek (2001) argue that a variance ratio of 1 works fairly well over a wide range, however the designs are only D-optimal for the variance ratio specified. Attention has also been given to the analysis techniques for split plot designs. Each split plot design presents its own unique challenges. The fact that split plot designs have two sources of error as a result of the restricted randomization requires the use of the appropriate error term when performing the analysis to detect significant model terms. Bingham and Sitter (2001) summarize the work of Bisgaard (2000) which gave specific guidelines for the use of the respective error terms associated with the split plot design. The guidelines are as follows: 1. Whole plot main effects and interactions involving only whole plot factors are compared to the whole plot error.
12 2. Subplot main effects or interactions that are aliased with whole plot main effects or interactions involving only whole plot factors are compared to the whole plot error. 3. Subplot main effects and interactions involving at least one subplot factor that are not aliased with whole plot main effects or interactions involving only whole plot factors are compared to the subplot error.
The analysis of unreplicated split plot designs has also been studied. Unreplitcated split plot designs do not provide any source of error for the whole plot and therefore testing for significance of the whole plot factors is difficult. Loeppky and Sitter (2002) present two methods for analyzing unreplicated fractional factorial split plot designs. They consider the modification of the Lenth’s Methods by removing the block effect and using a perturbation method which accounts for restriction on randomization within blocks. In recent times attention has been given to applying standard RSM techniques to split plot design and analysis. Attempts have been made to apply method of steepest ascent and second-order models to split plot designs. Kowalski et al. (2005) investigate the issue of applying the method of steepest ascent to the split plot design. They utilize two equations involving Lagrange multipliers to handle the different types of factors, whole plot and subplot. They proposed three methods of performing steepest ascent, two of which involve sequential experimentation by either: 1. Fixing a whole-plot setting first, varying the subplot settings, and then varying the whole-plot settings or, 2. Varying the whole-plot settings with fixed subplot settings and then varying the subplot settings in a fixed whole-plot setting. The third method runs one experiment by varying both the whole-plot settings and subplot settings together. The issue of robust parameter design involving split plot designs has also received attention. Kowalski (2002) applies the principle of semifolding to split-plot experiments and stated that the semifolding on subplot factors must remain inside a given whole plot and explained that this leads to a more restrictive setting where semifolding is not as
13 powerful as in the general case. The work of Kowalski (2002) also extended some of the concepts of partial confounding proposed by Bisgaard (2000). The author suggests that the method works well, in the split-plot case because the aliasing problems are typically with only one or two factors. Kowalski (2002) demonstrates this method using a 24 run design. The author suggests that for all situations involving less than four whole plot factors, a general method can be used to construct 24 run designs. First, one constructs a 16 run design that uses four whole plots with four subplots in each whole plot. Further conditions given are: 1. If there are two whole plot factors, say A and B, then one uses the complete factorial design in the whole plot factors. 2. If there are three whole plot factors, say A, B, and C, then one uses one of the two half fractions found using the defining contrast I = ABC. The designs presented by Kowalski (2002) are based on augmenting the whole plots rather than increasing the number of whole plots. Further, the runs are less expensive than whole plots therefore adding them to the design to break the specific alias chains is cost-effective.
2.3 Second-Order Split Plot Designs Draper and John (1998) introduce the concept of constructing RSM designs within a split plot structure. They point out that second-order composite designs might be regarded as unsuitable for the split plot structure for the following two reasons: 1. Randomization of the run order of the hard to change factor will require it to be changed repeatedly. 2. Also that the central composite design (CCD) would require that the hard to change factor is set at five (5) levels. However, it can be argued that the use of five (5) will provide addition benefits such as degrees of freedom for lack-of-fit test. According to the authors these conditions make the CCD impractical and therefore unsuitable for split plot designs. As a result Draper and John (1998) focused their
14 attention on cube and star designs that will require less experimental levels of the hard to change factors. They illustrate how the designs can be augmented to fit the required quadratic model. They used the rotatability criteria to determine acceptable or optimal designs. Trinca and Gilmour (2001) consider the issue of randomization restrictions and outline the construction of near orthogonal multistratum designs. The term “multistratum” refers to the number of restrictions on randomization within an experimental unit. Each restriction within the experimental unit is referred to as a stratum. Therefore, having more that one stratum is a multistratum design. The two strata designs presented are considered to be a split plot design. Goos and Vandebroek (2003) stated that a major drawback with the method of Trinca and Gilmour (2001) is that the choice of their design points does not take into account the split plot error structure of the experiment. Vining et al. (2005) also address the issue of applying second-order models to split plot designs. They have demonstrated how second-order models such as CCD and Box-Behnken Designs (BBD) can be developed within a split plot structure. Their work provides a detailed proof of the conditions that guarantee the equivalence OLS and GLS coefficient estimates. The conditions outlined are: 1. Balanced design - each whole plot contains the same number of subplots. 2. Orthogonal subplot designs (not necessarily the same design). 3. Axial runs for the subplot factors are run in a single whole plot. Parker et al. (2005, 2006) generalize the conditions given by Vining et al. (2005) for the equivalence of OLS and GLS. Parker et al. (2005, 2006) used the results of their generalization to illustrate the construction of both balanced and unbalanced equivalent estimation designs for second-order split plot designs and focus on the minimum number of whole plots as an additional requirement.
15 2.4 Design Optimality Criteria Box and Draper (1975) outline some properties of “good” response surface designs which were developed by Box (1968). According to the authors all or any combination of these properties might be used for different circumstances. They stress that good designs must consider and satisfy multiple objective criteria. However, satisfying multiple criteria may require trade-offs and comprises among the criteria being considered. According to Myers et al. (1989) much of the work during the 1970s and 1980s resulted in the development of new second-order designs which were based on the concept of D-optimality. The development of optimal design theory in the field of experimental design emerged following World War II. Myers et al. (1989) further explain that optimal design theory has clearly become an important component in the general development of experimental design for the case of regression models. The optimal design theory was developed around variance reduction of the parameter estimates with regards to the fitted model. However, they warn that one major drawback with the optimal design theory is that its application is based on a set of assumptions that may not be very realistic. These assumptions include model assumptions, number of design points for a given set of factors, and estimation is the sole criteria. Atkinson (1988) gave a lengthy discussion on the optimally design theory and the general equivalence theorem. The most prominent design criterion is D-optimality which focuses on minimizing the variance of estimates of the model coefficients. Goos and Vandebroek (2003) provided an efficient algorithm to generate D-optimal designs when the number and sizes of whole plots are predetermined. Goos and Vandebroek (2004) constructed D-optimal first-order and second-order split plot designs and showed that these designs outperform completely randomized designs in terms of both D and G optimality criteria. However, the assumption was made that complete randomization is possible and does not represent the hard-to-change context. Mitchell (1974 and 2000) also presented an algorithm for the construction of D-optimal experimental designs. The decision to use a single criterion for the evaluation of an optimal design is not a good approach because there are other conditions that contribute to the performance of
16 a design. Using only a single criterion such as the D-optimal criterion finds an optimal design for the variance of the estimated coefficients and neglects the issue of variance of prediction which is also an important criterion. Myers et al. (1989) state that many standard designs have prediction variances which increase dramatically as one gets close to the design perimeter. As a result, any conclusions drawn concerning the response near the design boundry are suspect. They lament that very little work has been done that deals with prediction variances in the design assessment or comparison of designs. They explain that too often designs are evaluated on the basis of a single criterion when the important aspects of behavior are multidimensional. However, since 1989 graphical techniques have been developed to include variance dispersion graphs (VDG) by Giovannitti-Jensen and Myers (1989) and fraction of design space (FDS) plots by Zahran et al. (2003). Borkowski and Valeroso (2001) did a comparison of design optimally criteria of reduced models for response surface designs in the hypercube. Four design optimally criteria were evaluated based on their performance with respect to the design optimality criteria. The central composite, computer generated and small central composite response surface designs were used in the study. For each of the designs they considered, robustness against model misspecification was quantified by calculating A, D, G, and IV optimality criteria for the reduced second-order models. They present some interesting conclusion on how the inclusion or non-inclusion of linear, cross-product and quadratic terms affects the behavior of the alphabetic optimality criteria. They conclude that the CCD is robust with respect to the set of reduced models as well as for the four optimality criteria. However, the computer-generated designs did not perform very well across the four criteria and across the set of reduced models. In general, D, A, G, and IV are not remain the same across reduced models. They vary primarily in the following ways:
1. With the removal of pure quadratic terms and, to a lesser extent, the cross product terms. 2. When models contain differing numbers of design variables. 3. For asymmetric designs because of the dependence on the assignment of design factors to the variable labels.
17
2.5 Summary Despite the various research efforts the application and analysis of split plots have several areas for further research. While the value of the contributions made by researchers cannot be belittled, there are too many restrictions and assumptions that limit the general application to industrial experiments. Optimal design theory has been used in the development of split plot experiments but its application is mainly confined to the use of D-optimally and G-optimally to a lesser extent. The point was made that designs that are optimal with respect to one criterion may be poor with respect to the other criteria. It is purpose of this research work to extend the body of knowledge available for the development of split plot designs particularly in the area of second-order split plot designs.
18 CHAPTER 3
3.0 Impact of Restricted Randomization on the Structure of the Information Matrix
3.1 Abstract A thorough investigation is conducted on the impact of restricted randomization on the information matrix. The information matrix is the single most useful source of information for evaluating experimental designs. All aspects of model building and data analysis such as estimation and prediction are based on this one powerful source of information. Therefore, understanding how the information matrix is affected by restricted randomization is very important in developing optimal designs. Several important insights and implications are presented for the construction of second-order split plot designs (SPD) within a response surface structure. Specifically, the study focuses on the effect of the variance ratio of the whole plot error variance to the subplot error variance and the structure of the design, for both balanced and unbalanced SPD. The investigation reveals several critical functions of the variance ratio and whole plot size(s). It was observed that, depending on the assignment of subplot axial points, the pure quadratic terms of the subplot factor will have similar correlation influence as the whole plot pure quadratic terms. This correlation influence has a tendency to negatively affect the optimality of the design. In addition, the study provides the analytical characterization of the inverse of the information matrix for second-order SPD. A particular feature of these explicit expressions is that they are functions of the design parameters. Finally, the application of these analytical expressions is demonstrated using the generalized variance of the parameter estimates.
19 3.2 Introduction The three fundamental principles in design and analysis of experiments are randomization, replication and blocking. However, in most industrial experiments it becomes a challenge to adhere to these principles at all times. In situations where it is difficult, expensive or dangerous to adjust certain factors during experimentation it becomes necessary to reduce the frequency of adjusting this factor(s). The reduction in the frequency of adjusting certain factors during experimentation is referred to as restricted randomization. The issue of restricted randomization in response surface methods was the focus of attention in Letsinger et al. (1996). In statistical analysis, designs with restricted randomization are commonly referred to as split plot designs (SPD). Wooding (1973) presented a detailed description on the structure and use of SPD. The use of SPD is becoming increasingly popular in industrial experimentation. The split plot approach is to separate the experiment into two design spaces known as the whole plot, for hard to change factors, and subplot, for easy to change factors. Therefore, the challenge is to combine the two design spaces which will result in overall efficient experimentation according to some stated design optimality criteria such as D, G and IV optimality criteria.
3.3 The Split Plot Model The general form for split plot models in matrix form is given as
yX= Θ ++ where y is an N x 1 vector of expected responses, N is the total number of observations, Θ is a p x 1 vector of unknown model parameters, p is the number of model parameters, X is the N x p matrix of the levels of the independent variables, and and are N x 1 vectors of random variables for whole plot and subplot errors respectively. N and It is assumed that + ~ N (0, Σ). The structure of the variance-covariance matrix Σ is
20 ⎡⎤Σ1 0...... 0 ⎢⎥0 00... ⎢⎥Σ2 Σ = ⎢⎥ 0 (1.3) ⎢⎥ ⎢⎥ 0 ⎢⎥ ⎣⎦00 0Σa
22' where a is for the number of whole plots, Σiwww=+σσδε11×11×× Iw and w is the number of th subplot runs (w hole plot size). In matrix form the structure of Σ i for th e i whole plot is
222 2 ⎡⎤σσσδεδ+ σ δ ⎢⎥ σ 2 Σ = ⎢⎥δ (1.4) i ⎢⎥2 σδ ⎢⎥2222 ⎣⎦⎢⎥σδ σσσδδε+
2 2 2 where σ δ + σ ε represents the variance of individual observations, σ δ is the covariance
2 and whole plot error variance while σ ε is the subplot error variance. Given the structure of the two error terms estimation of the model parameters is accomplished using the generalized least squares estimation approach given by
−1 Θˆ = (X''Σ−11XX) Σ− y
−1 Var ˆ X' −1X 2 2 where ()Θ = ()Σ . The two variance components, σ δ and σ ε , are needed to obtain parameter estimates and standard errors. Letsinger et al. (1996) point out that design optimality criteria, which are functions of the Var ()Θˆ , depend on the structure of the design, particularly the information matrix (A). In the case of completely randomize designs (CRD) the information matrix is given as X' X . However, for SPD alphabetic optimality is now dependent on an inform ation m atrix of the form X'Σ−1X , where Σ is the variance- covariance matrix described in Equation (1.1). This situation complicates the process
21 because now the optimal design will depend on the variance ratio or degree of correlation
2 2 η = σ δ σ ε , given by the correlation matrix (R). The correlation matrix (R) of the
22 observations is derived by dividing the variance covariance matrix, Σ, by σ δ +σ ε . In the actual computation of the information matrix the correlation matrix is used because it provides a unitless measure of the information matrix. An information matrix without units is desirable especially for design comparison purposes. As a result the information matrix would be without units and is given as
⎛⎞−1 ''⎛⎞Σ −1 AX==⎜⎟⎜⎟ X XRX ⎜⎟σσ22+ () ⎝⎠⎝⎠δε where R denotes the correlation matrix of the observations and its structure is of the form
⎡ ηη⎤ 1 ⎢ 11++η η ⎥ ⎡⎤R01 ...... 0 ⎢ ⎥ ⎢⎥⎢ η ⎥ 0R2 0... 0 ⎢⎥⎢1+η ⎥ . ⎢ ⎥ R0==⎢⎥ where Ri ⎢⎥⎢ η ⎥ 0 ⎢⎥⎢ 1+η ⎥ ⎢⎥00 0R ⎢ ⎥ ⎣a ⎦⎢ ηη⎥ ⎢ 1 ⎥ ⎣11++ηη⎦
Equation (1.5) shows the inverse of the correlation matrix (R-1) for a SPD with four whole plots of size four.
-1 ⎡R01 0 0⎤ ⎢ ⎥ 0R00-1 R−1 = ⎢ 2 ⎥ (1.5) ⎢ -1 ⎥ 00R03 ⎢ -1 ⎥ ⎣⎢000R4 ⎦⎥
22 where ⎡⎤ηη(1++ ) ηη (1 ) ()1+−η − ⎢⎥11++wwη η ⎢⎥ ⎢⎥ηη(1+ ) − ⎢⎥1+ wη R−1 = ⎢⎥. i ⎢⎥ηη(1+ ) − ⎢⎥1+ wη ⎢⎥ ⎢⎥ηη(1++ ) ηη (1 ) ⎢⎥−− ()1+η ⎣⎦11++wwηη
where 1+η represents the inverse correlation effect of individual observations within the η(1+η ) whole plots and − represents the inverse correlation co-effect within the whole 1+ wη plots. In terms of design optimality interpretation, the impact of the correlation matrix on the information matrix implies designs which are optimal in a CRD structure might not be optimal for a SPD structure. According to Goos and Vandebroek (2001 and 2004) the difference in optimality properties should not be surprising because designing a completely randomized experiment consists of determining the design points, while designing a split plot experiment requires simultaneously choosing the number of whole plots and the number of subplots within each whole plot and the design points for each plot type. Several research efforts have involved characterizing the information matrix of experimental designs. Srivastava and Chopra (1971) obtained the characteristic roots of the information matrix for balanced 2k-m fractional factorial designs of resolution V. Similarly, Hoke (1975) derived the characteristic polynomial of the information matrix for second-order models. Both works resulted in the formulation of explicit expressions for the determinant of the information matrix and the trace of the inverse of the information matrix. These two expressions are associated with D- and A-optimality criteria. Giovannitti-Jensen and Myers (1989) and Myers et al. (1992) presented some measures of the prediction variance using the eigenvalue decomposition of the inverse of
23 the information matrix. These measures included the maximum and minimum prediction variance for a first order model. Borkowski (1995) considered the case for second-order models and developed closed-form expressions for the prediction variance function for CCD and BBD designs. These closed form expressions were developed by the determination of the analytical form of the information matrix, inverting this analytical form and then pre- and post-multiplying the result by the general form of the function describing the location and the nature of the design model. The similarity with all of these efforts is that they focused on situations where it is assumed that complete randomization of the experimental run order is possible. However, in industrial experimentation there are situations that will lead to restrictions on randomization and therefore affecting the structure of the design. The current research work considers the impact of this restriction on the information matrix and presents analytical expressions for the characterization of the inverse of the information matrix for SPD within a CCD and BBD structure. These expressions will then be used to analytically determine the determinant of the information matrix.
3.4 Structure of Second-Order SPD The general form of the model vector used for the structure of second-order SPD is given as ' f ()z,x = ⎡ 2222⎤ ⎣1,,z111…… zwp x ,, x sp zz21 ,, … z wp−− z wp zx11 ,, … z wp x sp xx 121 ,, … x sp x sp z1 ,, … z wp x 1 ,, … x sp ⎦ where z and x are the whole plot and subplot factors respectively while wp and sp are the number of whole plot and subplot factors respectively. According to the structure of SPD they are four distinct categories of whole plots.
These categories are denoted as a1 to a4 for factorial, whole plot axials, subplot axials and center runs respectively. Therefore each category would have whole plot sizes of w1 to w4 respectively. In general if there are two different whole plot sizes the resulting SPD is unbalanced. Table 1 and Table 2 show typical structures of the design matrix X for a
24 balanced SPD within a CCD and BBD structure respectively. Table 3 shows an example of the design matrix X for an unbalanced SPD within a BBD structure. Note that whole plot 4 in Table 3 is of size two (2) while the other whole plots are of size four (4). For details on the construction of second-order designs within a response surface structure, see Vining et al. (2005) and Parke r ( 20 05 a n d 2 00 6). C onsi deration is given to these designs becau se th ey pro vi de eq uiv al ent estim ation of ordinary least square and generali zed le ast sq uar e. Considera tio n i s also given to the standard second-order designs when they are conducted as SP D. H ow ev er, t he se d e sign s d o not s a tisfy the equivale nt esti mation property. The notation VKM is used to denote that the Vining method of construction is used while the notation MWP is used to denote that the Parker method of construction is used. For labeling designs the first numbe r in the e xtension portion, such as VKM D12, represents the numbe r o f w hol e plo ts while th e second represents the number of subplots in the design.
25
Table 1: The X matrix for a CCD with one whole plot and two subplot factors (VKM D12) 2 2 2 Whole Plot z1 x1 x2 z1x1 z1x2 x1x2 z1 x1 x2 Points 1 -1- 1 -1111111 1 1 -1 1 -1-11-1111 Factorial 1 -1 -111- 1-1111 1 -1 1 1-1-11111 1 1- 1- 1-1-11111 2 1 1 1- 11-1-1111 Factorial 1 1- 1 1-11-1111 1 1 1 1111111 1 - 0 0000 2 00 3 1 - 0 0000 2 00 Whole plot axials 1 - 00000 2 00 1 - 0 0000 2 00 1 0 0000 2 00 4 1 0 0000 2 00 Whole plot axials 1 0 0000 2 00 1 0 0000 2 00 1 0 -α 00000α2 0 5 1 0 α 00000α2 0 Subplot axials 1 0 0 -α 00000α2 1 0 0 α 00000α2 1 0 0 0000000 6 1 0 0 000 0 00 0 Centers 1 00 0000000 1 00 0000000
26
Table 2: The X matrix for a balanced BBD with one whole plot and two subplot factors (VKM D12) 2 2 2 Whole Plot z1 x1 x2 z1x1 z1x2 x1x2 z1 x1 x2 Points 1 -1 -1 0 100110 1 1 -1 1 0 -100110Edge centers 1 -1 0 -1 010101(Axial Structure) 1 -1 0 1 0-10101 1 0 -1 -1 001011 2 1 0 1 -1 00-1011Factorial 1 0 -1 1 00-1011 1 0 1 1 001011 1 1 -1 0 -100110 3 1 1 1 0 100110Edge centers 1 1 0 -1 0-10101(Axial Structure) 1 1 0 1 010101 1 0 0 0 000000 4 1 0 0 0 000000 Centers 1 0 0 0 000000 1 0 0 0 000000
Table 3: The X matrix for an unbalanced BBD with one whole plot and two subplot factors (VKM D12) 2 2 2 Whole Plot z1 x1 x2 z1x1 z1x2 x1x2 z1 x1 x2 Points 1 -1 -1 0 100110 1 1 -1 1 0 -100110Edge centers 1 -1 0 -1 010101(Axial Structure) 1 -1 0 1 0-10101 1 0 -1 -1 001011 2 1 0 1 -1 00-1011Factorial 1 0 -1 1 00-1011 1 0 1 1 001011 1 1 -1 0 -100110 3 1 1 1 0 100110Edge centers 1 1 0 -1 0-10101(Axial Structure) 1 1 0 1 010101 4 1 0 0 0 000000 Centers 1 0 0 0 000000
27 3.5 Critical Functions of the Variance Ratio and Whole Plot Sizes A thorough investigation of the impact of the variance ratio on the information matrix revealed that, for both “balanced” and “unbalanced” symmetric designs, there are several critical functions (Φ) of the variance ratio and whole plot sizes that alter the values of the information matrix for second-order SPD. These critical functions were derived by pre and post multiplying a symbolic representation of the inverse of the correlation matrix (R-1) by the design matrix X. Matlab was used to assist in the determination of the symbolic representation of the information matrix XR' −1 X. The process leads to the critical functions given in Equations (1.6) and (1.7) according to the nature of the design. In the case of unbalanced SPD these critical functions are also applic able but require some modification. This modification is due to the differences in whole plot sizes. However, these modifications are of the same form as Φ1 and Φ3 using the relevant whole sizes. If the unbalanced nature of the SPD is as a result of the different whole plots sizes for the whole plot or subplot axials, then Φ6 or Φ7 is used.
Further, in the case of the subplot axials, Φ3 is modified using the necessary whole plot size associated with the subplot axials. However, if the unbalanced nature of the SPD is as a result of different whole plot size for the whole plot centers then the modification to
Φ1 is reflected using Φ8 and is applicable to the center runs. Φ9 is utilized only when the center runs are grouped with the subplot axials. In the case of balanced SPD there is only one whole plot size (w1) to consider and therefore the critical functions are:
(1+η ) Φ=1 (1 + w1η )
Φ=2 1+η ⎛⎞ηη(1 + ) Φ=−3 2⎜⎟ (1.6) ⎝⎠1 + w1η
Φ=421 or Φ if whole plot axials are grouped with factorial whole plot
Φ=Φ52 or Φ 1 if subplot axials are in sparate whole plots while Φ3 goes to zero
Φ6 = ΦΦ=ΦΦ=781 = and 9 0.
In the case of unbalanced SPD there is more than a single whole plot size to consider and therefore, the critical functions are modified as follows:
28
(1+η ) Φ=1 (1 + w1η )
Φ=2 1+η ⎛⎞ηη(1 + )⎛⎞ ηη (1 + ) Φ=−3 2⎜⎟ or −2⎜⎟ if unbalanced subplot axials ⎝⎠1 + ww13ηη⎝⎠ 1 +
Φ=421 or Φ if whole plot axials are grouped with factorial whole plots (1.7)
Φ=Φ52 or Φ 1 if subplot axials are in sparate whole plots while Φ3 goes to zero (1+η ) ΦΦ67, , and Φ= 8 for i =2,3,4 (1 + wiη ) ⎛⎞(1 + η )2 Φ=9 0 or 2⎜⎟ if the centers are grouped with subplot axials . ⎝⎠1 + w3η
For both the balanced and unbalanced cases the whole plot factor linear (zi) and two- factor (zizj) interaction terms are affected by critical functions of the form Φ1 while the subplot factor linear (xi) terms and its interactions (zixi and xixj) are affected by Φ2. This is an indication that for a first order model only critical functions of the form Φ1 and Φ2 are applicable.
3.6 Information Matrix for CCD and BBD Structures This section outlines the symbolic representation of the information matrix for the split plot CCD and BBD designs. The symbolic representations are done using the design parameters for the characterization of the designs. Table 4 gives the list of notations used throughout the derivation of the analytical expression for the information matrix for a CCD structure.
29 Table 4: Table of notations for the split plot CCD Notations Meanings
kk−m f # of factorial runs () 2 or 2V fw # of whole plot factor runs th wii whole plot size awii# of whole plots with size rw # of repeated whole plot axials rs # of repeated subplot axials β Whole plot axial setting α Subplot axial setting ζ Factor level setting
The information matrix for the split plot CCD can be partitioned as follows:
XR' −1 X= ⎛ 22'' 22 ⎞ ⎜ ΠΦ0J( 16frwwwζβ+ΦΦ22) p( 17 fr ζα+Φ ss) Jp ⎟ ⎜ ⎟ ⎜ 00Diag() di 0⎟ ⎜ ⎟ ⎜ ' ⎟ Φ+frζβ22Φ22J0 Φ r β 4 I +ΦΦ f ζ 4 JJJ'' f ζ 4⎡⎤J ⎜ 16w w) wp 6 w wp 1 w wp wp 1⎢⎥ sp wp ⎟ ⎜( ⎣⎦ ⎟ ⎜ 22 4'' 4 44⎟ ⎜ Φ+frζαΦ22J0 Φ f ζ JJIJ Φ+ r αΦ +ΦΦ+ fr ζαΦ2J⎟ ⎝ ( 17s ) sp 1 sp wp( 5 s 9) sp( 413 s) sp sp ⎠ (1.8)
where 0’s are zero matrices of appropriate sizes, Jwp and Jsp are unit vectors of wp×1 and sp×1 respectively, I wp and Isp are wp-dimensional and sp-dimensional identity matrices and Diag (di) are diagonal elements which are given as follows:
22 dfiw=Φ16ζβ +Φ2 rf w ori 1 ≤ ≤wp 22 =Φ2 ()frζα +2s for wpi +≤≤ 1 k , where k =+wpsp wp 4 ⎛⎞ =Φ1 ffwζ or k + 1 ≤i ≤k +⎜⎟ ⎝⎠2
4 ⎛⎞wp ⎛⎞ wp ⎛⎞ sp =Φ 2 ffζ or ki +⎜⎟ +≤≤+ 1k ⎜⎟ +×+wpsp ⎜⎟ . ⎝⎠22 ⎝⎠ ⎝⎠2
30 The first entry, Π, in Equation (1.8) is a scalar quantity. For a balanced SPD Π=N Φ1 and for an unbalanced SPD Π=aw11 Φ 1 + aw 2 2 Φ 6 + aw 33 Φ 7 + aw 4 4 Φ 8. The total number
4 of runs is Na= w. ∑i=1 ii Equation (1.8) represents a useful form of the information matrix for the split plot CCD and shows the location of the critical functions. The partitioning of the matrix is done to differentiate between the whole plot and subplot design moments. Because the information matrix is symmetric the components on either side of the main block diagonal are identical and represent the quadratic design moments of order two. The upper entries on the main diagonal represent design moments of order two for the intercept, linear and two-factor interactions. The entries of the lower block diagonal represent the quadratic design moments of order four. The following outlines a similar form for the split plot BBD. Table 5 gives the list of notations used throughout the derivation of the analytical expression for the information matrix for the split plot BBD.
Table 5: Table of notations for the split plot BBD Notations Descriptions f # of factorial runs per block () 2t
fc # of factorial runs at edge centers th wii whole plot size awii# of whole plots with size rw # of blocks within which a whole plot factor appears rs # of blocks within which a subplot factor appears t# of active subplot factors per block (t≥ 1)
λc # of edge centers
λλw # of times a pair of whole plot factors appears in the same block (w ≥ 1)
λs # of times a pair of subplot factors appears in the same block
λint # of times a whole plot factor appears with a subplot factor in the same block β whole plot factor level setting α subplot factor level setting
31 The information matrix for a symmetric SPD, based on a BBD structure, can be partitioned as shown in Equation (1.9). The partitioning of the inf ormation matrix for the split plot BBD is similar to the split plot CCD however the computations for the design moments are different. Note that some of the critical functions are not applicable because of the nature of the split plot BBD.
X'RX−1 =
⎛ ΠΦ0Jfr βα2 '' Φfr 2 J ⎞ ⎜ 1 ()wwp 1 ()ssp ⎟ ⎜ 00Diag d 0⎟ ⎜ ()i ⎟ ⎜ ⎟ ΦΦfrβλ24J0 f r−β I+ f λβ4 JJJ'' Φf λα 4J' ⎜ 11()w wp()() w w wp w wp wp ()1int sp wp ⎟ ⎜ ⎟ ⎜ 244' 444' ⎟ ⎜ ΦΦ11frsαλJ0sp fintαλ JspJIJ wpΦ5 f() r s− sα+Φ1Φ4 fλ sα+Φ3 f () r s−λ sα+ f cλ cα spJ sp ⎟ ⎝ () ( ) sp ()()⎠ (1.9) where
2 dfiw=Φ1 rfβ ori 1 ≤ ≤wp 2 =Φ2 frsα for wp + 1 ≤ i ≤ k
4 ⎛⎞wp =Φ1 ffλβw or k + 1 ≤i ≤k +⎜⎟ ⎝⎠2
4 ⎛⎞wp ⎛⎞ wp =Φ 2intffλα or ki ++≤≤++×⎜⎟ 1 k ⎜⎟wpsp ⎝⎠22 ⎝⎠
4 ⎛⎞wp ⎛⎞ wp ⎛⎞ sp =Φ2 ffλαs or k+⎜⎟ +wp ×+≤≤+ sp1. i k ⎜⎟ + wp ×+ sp ⎜⎟ ⎝⎠22 ⎝⎠ ⎝⎠2
The symbolic representation of the information matrix for the split plot CCD and BBD, given in Equations (1.8) and (1.9) respectively, illustrates how the critical functions and design points are mapped into the information matrix. The last entry of the main block diagonal for both design structures involves several critical functions. Therefore, the assi gnment of subplot axial points plays a crucial role in the application of these critical functions. The next section details the analysis of the impact of design structure on the application of the critical functions.
32 3.7 Impact of Design Structure on the Application of the Critical Functions
In general the information matrix, XR' −1 X, may be partitioned into four (4) sub- matrices according to the parameters needing to be estimated as follows:
' -1 ⎛⎞AA11 12 XR X== A ⎜⎟. ⎝⎠AA21 22
The details of the partitioned matrix A, as shown in Equation (1.8), for the split plot CCD is given as:
⎛⎞Π 0 A11 = ⎜⎟ ⎝⎠0 Diag() di ⎛⎞Φ+frζβ22Φ2222JJ' Φ+ fr ζαΦ ' A = ⎜⎟()16wwp() 17 ssp 12 ⎜⎟ ⎝⎠00 (1.10) ⎛⎞Φ+frζβ2 Φ2 2 J0 ⎜⎟()16wwp A21 = ⎜⎟Φ+frζα22Φ2 J0 ⎝⎠()17ssp ⎛ ' ⎞ Φ+2rfβζ44IJΦΦJJ' f ζ 4⎡⎤J' ⎜ 16wwp wpwp 1⎣⎦spwp ⎟ A22 = ⎜ ⎟. ⎜ ΦΦfrfζαζ444JJ' 22+Φ I+ΦΦ+Φrα4 JJ' ⎟ ⎝ 15sp wp() s9 sp()413 s spsp⎠
Further, it can be observed that A22 is partitioned into four (4) sub-matrices representing the quadratics of the whole plot and subplot terms in the model. As a result, A22 can be partitioned as
⎛⎞CC11 12 A22 = ⎜⎟ ⎝⎠CC21 22 where the components of the C matrix is given as:
33 CI=Φ2rfβζ44 + JJ' 11 1 ()w wp w wp wp CJ=Φ f ζ 4 J'' 12 1 sp wp (1.11) ' CC21= 12
444' CI22=()Φ 52rfsαζ + Φ 9sp +ΦΦ() 4 1 +Φ 3 2r sαJspJsp .
Note that both C11 and C22 repr es ent the whole plot and s ubp lo t qu ad rati c t erm s of order four respectively while C12 and C21 represent the interactio n of w ho le plo t a nd subplot quadratic terms of order four respectively. In situations where the design structure is such that t he wh ole p lot ax ials are placed in separate whole plots and the subplot axials are also placed in s epa rat e whole plots, as shown in Ta bl e 6 , t hen th e imp act o f the varian c e ra tio is th e sa m e fo r both C11 and C22. Therefore, Φ3 = 0, Φ4 = 1 , Φ5 = Φ1 and Φ9 = 0. In other words, the desig n momen ts of o rder fo ur for bo th th e subplot and whole plot axials are affected in a similar manner by Φ1. The r es ulti ng fo rm of C22 is given in Equation (1.12). Observe that Φ1 is the only critical f unction active which is also the case for C11. No te tha t in T ab le 6 the whole plot axials are placed in whole plots 5 through 8 while the subplot axials are placed in whole plots 9 through 11. Further, observe that the subplot axials are repeated in order to preserve the balanced nature of the SPD. Consequently, C22 is given as follows:
CI=Φ2rfαζ44 + JJ' . (1.12) 22 1 ( s sp sp sp )
34
Table 6: A resolution V CCD with = α = 1, two whole plots and three subplot factors (VKM)
Whole Plot z1 z2 x1 x2 x3 Whole Plot z1 z2 x1 x2 x3 1 -1 -1 -1 -1 1 7 0 - 00 0 -1 -1 1 -1 -1 0 - 00 0 -1 -1 -1 1 -1 0 - 00 0 -1 -1 1 1 1 0 - 00 0 2 1 -1 -1 -1 -1 8 0 00 0 1 -1 1 -1 1 0 00 0 1 -1 -1 1 1 0 00 0 1 -1 1 1 -1 0 00 0 3 -1 1 -1 -1 -1 9 0 0 - α 0 0 -1 1 1 -1 1 0 0 α 0 0 -1 1 -1 1 1 0 0 - α 0 0 -1 1 1 1 -1 0 0 α 0 0 4 1 1 -1 -1 1 10 0 0 0 - α 0 1 1 1 -1 -1 0 0 0 α 0 1 1 -1 1 -1 0 0 0 - α 0 1 1 1 1 1 0 0 0 α 0 5 - 0 0 00 11 0000 - α - 00 0 0 0000 α - 00 00 00 00 - α - 00 00 00 00 α 6 00 00 12 00 00 0 00 00 00 00 0 0 0 00 00 00 0 0 0 00 00 00 0
However, if all th e subplot axials ar e placed in the same whole plot, as shown in Table 7, then the sub plo t ax ia ls w o uld b e impacted differently by the variance ratio.
Therefore the critic al fu nc tio ns are ad justed accordingly with Φ4 = 1, Φ5 = Φ 2 and Φ9 = 0.
The resulting form of C22 is given in Equation (1.13). Observe that along with Φ1 there are tw o addition al a cti ve critical f u nc tion s given as Φ2 and Φ 3. Φ2 represent the inverse correlation effect on the individual ob servation of subplot axials while Φ3 represents the inverse correlation co-eff ect within th e whole plot. The overall effect of these critical functions is such that the design moments of order four for the subplot axials are correc ted for th e in flue nce of th e in ver se correlation co-effect within the whole plot. As a result C22 is affected as follows:
35 444' CI22=Φ 222rfrsαζαsp +( Φ 1 +Φ 3 s) J spJ sp . (1.13)
In contrast, if the whole plot axials are placed in the same whole plots as the factorial runs for t hei r respecti ve se ttings (±) then Φ4 = Φ2 and Φ9 = 0 if whole plot centers are used. T her efore C22 will now be affected based on w het her or not the subplot factors are in the same or sep ar ate w hole plots. Consequently, having the subplot axials in the same whole plot, as shown in Table 8, will result in C22 being modified as shown in Equation (1.14). Note that the inverse correlation effect is now affecting the contribution of the factorial points to the design momen ts of o rder four for the subplot as follows:
444' C22=Φ() 2 22rfsα +Φ9IJsp +ΦΦ( 2 1 ζ+ Φ 3 rsα) spJsp . (1.14)
Table 7: A CCD with = α = 5 , one whole plot factor and three subplot factors (MWP)
Whole Plot z1 x 1 x 2 x 3Whole Plot z 1 x 1 x 2 x 3 1 1 -1 -1 -1 4 00 0 1 -1 -1 1 00 0 1 -1 1 -1 00 0 1 -1 1 1 00 0 1 1 -1 -1 00 0 1 1 -1 1 00 0 1 1 1 -1 00 0 1 1 1 1 00 0 2 -1 -1 -1 -1 5 0 - α 0 0 -1 -1 -1 1 0 α 0 0 -1 -1 1 -1 0 0 - α 0 -1 -1 1 1 0 0 α 0 -1 1 -1 -1 0 0 0 - α -1 1 -1 1 0 0 0 α -1 1 1 -1 0 0 0 0 -1 1 1 1 0 0 0 0 3 - 0 0 0 - 0 0 0 - 0 0 0 - 0 0 0 - 0 0 0 - 0 0 0 - 0 0 0 - 0 0 0
36
Table 8: A CCD with = α = 1, one whole plot factor and three subplot factors (MWP)
Whole Plot z1 x1 x2 x3 Whole Plot z1 x1 x2 x3 1 -1 -1 -1 -1 3 1 -1 -1 -1 -1 -1 -1 1 1 -1 -1 1 -1 -1 1 -1 1 -1 1 -1 -1 -1 1 1 1 -1 1 1 -1 1 -1 -1 1 1 -1 -1 -1 1 -1 1 1 1 -1 1 -1 1 1 -1 1 1 1 -1 -1 1 1 1 1111 - 0 0 0 000 2 0 - α 0 0 0 α 0 0 0 0 - α 0 0 0 α 0 0 0 0 - α 0 0 0 α 0 0 0 0 0 0 0 0 0 0 0 0
In situations where we have separate whole plots for the subplot axials C22 is affected as shown in Equation (1.15). Note that Φ3 = 0 which means the inv erse correlation co-effect has no influence on the subplot axials. However, the inverse correlation effect is still affecting the contribution of the factorial points to the design moments of order four for the subplot as follows:
44' CI22=Φ 12rfsαζsp +( Φ 2 Φ 1 ) J spJ sp . (1.15)
The observations above imply that the structure of the design plays a significant role in deciding the impact of the variance ratio on the design moments of order four for the subplot factors. A similar effect exists for the split plot BBD. This type of behavior is not expected for the first order design and model because only linear and interaction terms are involved. The inverse of the information matrix plays a critical role in the determination of optimal design properties pertinent to second-order models such as the
37 maximum and integrated prediction variance. Further analytical characterization of the inverse of the information matrix would provide more insights into the effect of these critical functions on design optimality. The following section outlines the analytical characterization scheme used to determine the i nverse o f the information matrix.
3.8 Outline of Analytical Characterization Scheme This section outlines the procedure followed to derive the analytical expressions for the characterization of the inverse of the information matrix for second-order SPD. In
−1 general the inverse of information matrix (XR'1− X) may be partitioned into four (4) sub-matrices as follows:
−1 ⎛⎞BB11 12 AB= =⎜⎟. ⎝⎠BB21 22
Assuming a symmetric structure of the SPD, then both A11 and A22 are nonsingular. Consequently, it is possible to apply theorem 8.5.11 of Harville (1997) to construct the inverse of the information matrix A. The theorem makes use of a function call the Schur
−1 Complement. It turns out that the Schur Complement of A11 and A22 are equal to B11
−1 and B22 respectively. For the purposes of this work we use the Schur Complement of
A11, which is given as
−−11 BAAAA11=− 11 12 22 21 . (1.16)
In addition, Harville (1997) showed that
AA−−11=− BB 12 22 11 12 (1.17) −−11 ABBBB22=− 22 21 11 12.
38 The two equations given in (1.17) are used to solve for B12 and B22 respectively. Because of the symmetric nature of the information ma trix once B12 is known the components of
B22 are also known. The derived expressions for each component of the information matrix were computed using
1 −1 − BA11 =( 11−A 12AA 22 21 ) BB=− AA−1 12 11 12 22 (1.18) ' BB21= 12 −−11 BABBB22=+ 22 21 11 12.
Using the re sults of the E quation (1.18) the inverse of the information matrix is constructed in terms of the design parameters and variance ratio. The analytical characterization of the inverse of the information matrix was arrived at using the following procedures:
−1 1. Determin e A22 . 2. Form and invert the Schur Complement. 3. Use Equation (1.18) to compute the components of the inverse of the information matrix. Figure 2 summarizes the analytical computation scheme used for the determination of the inverse of the information matrix. Information on the principle of partitioned matrix inversion and the Schur Complement is detailed in Harville (1997) and Searle (1982).
39
' −1 ⎛⎞AA11 12 XR X= ⎜⎟ Partition information matrix ⎝⎠AA21 22
⇓
⎛⎞AA11 12 ⎜⎟ ⎛⎞CC Partition A ⎜⎟A 11 12 22 ⎜⎟21 ⎜⎟ ⎝⎠⎝⎠CC21 22
⇓
⎛⎞AA11 12 ⎜⎟ ⎛⎞DD Determine A-1 ⎜⎟A 11 12 22 ⎜⎟21 ⎜⎟ ⎝⎠⎝⎠DD21 22
⇓ −−11 BAAAA11=− 11 12 22 21 Formulate and invert the Schur Complem ent
⇓
−1 −1 ''−1 ⎛⎞BB11 12 −1 ()XR X = ⎜⎟ Determine the ()XR X ⎝⎠BB21 22 Figure 2: Analytical Characterization Scheme
−1 3.9 Characterization of (XR' −1 X) for Split Plot CCD
−1 In order to begin the process, the determination of the components of A22 is required. The components of A22 are given in Equation (1.11) and reflect the nature of SPD in terms of the two distinct design spaces. Because of this partitioning the principles
−1 outlined in Figure 2 above must also be applied to obtain the components of A22 .
−1 Consequently, A22 is denoted as
−1 ⎛⎞DD11 12 AD22 ==⎜⎟. ⎝⎠DD21 22
40
First, the formulation of the expression of the Schur Complement for C11 is given as
−11− D11 =−CCCC11 12 22 21 (1.19) where the components of the C matrix is given in Equation (1.11). Then it follows that the components of D are given as:
−1 −1 DD11= ( 11 ) DD=− CC−1 12 11 12 22 ' DD21= 12 −−11 DCDDD22=+ 22 21 11 12.
−1 In order to solve the Schur Complement in Equation (1.19), the matrix C22 must be determined. Since C22 is symmetric its inverse is of the form
−1 CIJ−1 =+abJ' 22 ( sp sp sp ) (1.20) 1 ⎛⎞b ' =−⎜⎟IJspspJsp aa⎝+ spb⎠
444 where ar=Φ592ssα +Φ and b =Φ 4 Φ13 fζα +Φ 2 r. Therefore,
1 ⎛⎞ΦΦfrζα44 +Φ2 CI−1 =−⎜⎟41 3 s JJ' . 22 Φ+2r α 4 Φ⎜⎟spsΦ+22rsαζ44Φ+ΦpΦ+frΦα4psp 59s ⎝⎠594ss()13
For details of the principle involved in finding the inverse of matrices of the form given in Equation (1.20) see Graybill (1969). This method is used extensively throughout this derivation process. See Appendices A1 and A2 for details. The result of
41 the computations after some tedious matrix manipulation and linear algebra, the
−1 components of D, and by extension the components of A22 where determined as
1 ⎛⎞c DI=−1 ⎡ JJ' ⎤ 11 44⎜⎟wp⎢ wp wp ⎥ ΦΦ66122rrwwββ⎝⎠+wpc⎣ ⎦ Φ f ζ 4 1 '' (1.21) DJ12 =− spJ wp c2 ' DD21= 12
2 ⎛⎞ΦΦfRrcζα44 +22 − Φ r α 4 +Φ Φ fw ζ 4p 1 ⎜⎟()41 3ss 2(( 5 9)( 1 )) DI=− JJ' . 22 Φ+2r α 4 Φ⎜⎟sp 444sp sp 59s ⎜⎟()Φ+59422rsssαζΦ+Φp()Φ+132frΦαc ⎝⎠ where
2 4 (Φ1 fsζ ) p cf=Φζ 4 − 11w 444 Φ+59422rsssαζΦ+Φp()Φ+13frΦα
4444 cr25=Φ()22ssαζ +Φ+Φ9sp() 4 Φ 1 f +Φ 3 rα() Φ+ 12 rwwβpc 1.
Returning to the information matrix, A, and applying the results obtained in
−1 Equation (1.21) for A22 the Schur Complement for A11 can now be formed and inverted. This then leads to the analytical characterization of the inverse of the information matrix for a CCD structure. Recall that the Schur Complement for A11 is given in Equation (1.16) and that the components of matrix A are defined in Equation (1.10). The components of B can now be determined using Equation (1.18). The results are
42 ⎛⎞γ1 0 ⎜⎟ B11 = ⎛⎞ ⎜⎟0 Diag ⎜⎟1 ⎝⎠⎝⎠di ⎛⎞γγJJ'' B = 23wp sp 12 ⎜⎟ ⎝⎠00 (1.22) ⎛⎞γ 2J0wp B21 = ⎜⎟ ⎝⎠γ 3J0sp ⎛⎞1 ' IJ−γγJJ''⎡⎤J ⎜⎟Φ 2r β 4 ()wp45 wp wp⎣⎦ sp wp B = ⎜⎟6 w . 22 ⎜⎟1 γγJJ'' I− JJ ⎜⎟56sp wp 4 ()sp sp sp ⎝⎠Φ+592rsα Φ
By combining the results of B11, B12, B21 and B22, the components of the inverse of the
−1 information matrix can be obtained. Therefore (XR' −1 X) for CCD is given as
⎛ γγγ0J'' J⎞ ⎜ 123wp sp ⎟ ⎜ ⎛⎞ ⎟ 000Diag ⎜⎟1 ⎜ ⎝⎠di ⎟ ⎜ ⎟ 1 ' . ⎜γγJ0 IJ− JJ''γ⎡⎤J ⎟ ⎜ 24wp Φ 2r β 4 ()wp wp wp5⎣⎦ sp wp ⎟ ⎜ 6 w ⎟ 1 ⎜ γγJ0 JJI''−γJJ⎟ ⎜ 35sp sp wp 4 ()sp6 sp sp ⎟ ⎝ Φ+592rsα Φ ⎠
Table 9 summarizes the scalar quantities represented by γ i in the inverse of the information matrix while Table 10 provides the details of the pertinent constants (c) that
−1 were derived. The inverse of the information, (XR' −1 X) , represents a useful form that will facilitate an understanding of the contribution of each of the distinct design spaces to the optimality of the design.
43 Table 9: Table of scalar quantities for the split plot CCD Sca lars Equations 1 γ1 c11
c5 γ 2 − c11 c γ − 9 3 c11 4 2 c1 Φ652rcwβ () γ 4 4 − Φ+6112rwwβ pc c1 4 Φ1 f ζ cc59 γ 5 −+ cc211 4 2 ()Φ+592rcsα Φ()9 γ 67c − c11
Table 10: Table of constant quantities for the split plot CCD Constants Equations 2 4 ()Φ1 fsζ p cfΦ−ζ 4 11w 444 Φ+59422rsssαζΦ+Φpfr()Φ+13Φα
4444 cr259()Φ+22ssαζΦ+Φsp()4Φ+1fΦ36rα() Φ2rwβ +wpc1 22 Φ+16frwwζβΦ2 c3 4 Φ+612rwwβ pc 42 2 ΦΦ+11ffζζ()Φ 72 rss αp c4 − c2 cc53+ c4 42 2 ΦΦ+11ffζζ()wwΦ 62 rw βp c6 − c2 2 ΦΦf ζ 4 +Φ22rcααζ444 − Φ r +Φ Φ f wp ()41 325ss(( 91)()) c 7 444 ()Φ+5942rsssαζ Φ+Φp()Φ+132frΦ2αc 22 ()Φ+17frcζαΦ21s () − 7sp c 8 4 Φ+592rsα Φ cc96+ c8
2222 cc10 5()Φ+ 1frwwζβΦ 622wp +Φ+c 9() 1fr ζαΦ 7 ssp cc11Π− 10
44 −1 3.10 Characterization of (XR' −1 X) for Split Plot BBD The same analytical computational scheme was followed for the BBD structure.
−1 Therefore, the ()XR' −1 X for BBD is given as
⎛ γγ0J'' γ J⎞ ⎜ 12wp 3sp ⎟ ⎜ ⎛⎞ ⎟ 00Diag ⎜⎟1 0 ⎜ ⎝⎠di ⎟ ⎜ ⎟ . 1 ' ⎜γγJ0 IJ− JJ''γ⎡⎤J ⎟ ⎜ 24wp Φ−fr()λβ4 ()wp wp wp5⎣⎦ sp wp ⎟ ⎜ 1 ww ⎟ 1 ⎜ γγJ0 JJ'' IJ−γJ⎟ ⎜ 35sp sp wp 4 ()sp6 sp sp ⎟ ⎝ Φ−5 fr()ssλα ⎠
Table 11 summarizes the scalar quantities represented by γ i in the inverse of the information matrix while Table 12 provides the details of the pertinent constants (c) that were derived.
Table 11: Table of scalar quantities for the split plot BBD Scalars Equations 1 γ1 c11
c5 γ 2 − c11
c9 γ 3 − c11 4 2 c1 Φ−15f ()rcwwλβ() γ 4 4 − Φ−11fr()wwλβ + wpc c11 4 Φ1intfcλα 59c γ 5 −+ cc211 4 2 Φ−59fr()ssλα( c) γ 67c − c11
45 Table 12: Table of constant quantities for the split plot BBD Constants Equations 2 4 ()Φ1intfsλα p cfΦ−λβ4 11w 44 44 Φ−51fr()ssλα+Φ sp()Φ4 f λα s +Φ3() fr () ss − λα + f cc λα cfΦ−rλα44 +ΦspΦff λα +Φ−+rff λα 4 λα4 Φ−+rw λ β 4pc 25()()ss() 143 s() () ss cc() 1() ww 1 2 Φ1 ()frwβ c3 4 Φ−11fr()wwλβ + wpc
2 24 ()Φ1ifrsαλαnt sp c4 − c2 c53c+ c4 2 2 4 ()Φ1 frwβ λαint wp c6 − c2 2 ΦΦfλα44 +Φ fr− λα + fλα4c− Φ fr − λ α 4Φ fw λ α4p ()14ssc 3()()sc2 ( 5() s s(1 int )) c7 ΦΦfr−λα4+ spΦfλ α4 +Φ fr − λα 4 +fλ α4c ()51()ss (4s 3() ()ss cc) 2 Φ−rα 2 1 csp 17()s () c8 4 Φ−5 ()rsλsα cc96+ c8 22 cc10 5ΦΦ 1frwβwpf+c9 1 rsαsp cc11Π− 10
3.11 Generaliz ed Va ria nc e of Pa ra met er Est ima tes To demonstrate th e u se of t he se ana lytic al expressions we obtain t he generalize d varianc e of para me ter estim ates which is measured by the D-optim a lity cr iteri on . Th e D- optima lity crite rion is the de term inant of t he info rmation m a trix. In o rde r to d eter m ine the determinant of the infor mation m atrix (A), theorem 13.3.8 of Harvil le (1 9 97) was applied. See App end ix B1 fo r details. According to the theo rem , if a m atrix A is partitioned such that A11 is nonsingular and the inverse of A is given as B, then the determinant of A is com p uted a s fo llo ws:
A A = 11 . (1.23) B22
46
3.11.1 Determinant for split plot CCD Using Equation (1.23) the analytical expressions for the determinant of second- order split plot CCD is given as follows:
wp sp cw c cs 22 224 4+ ΠΦ()16frζβ +Φ22ws( Φ 2( frf ζ +αζζ)) () Φ 1() Φ 2 f XR' −1X = . wp−1 ⎛⎞4 2 sp−1 ⎛⎞1−γ 4wp ⎛11⎞1−γ 6sp Φ652rwwβγ() psp⎛ ⎞ ⎜⎟44⎜⎟⎜⎟ 4− ⎜4 ⎟ Φ22rββ Φr⎜⎟ Φ+ 2 r αΦ 1 − γ wp Φ+ 2 r αΦ ⎝⎠61w⎝w⎠⎝⎠ 59 s 4⎝ 59s ⎠
3.11.2 Determinant for split plot BBD Using Equation (1.23) the analytical expressions for the determinant of second- order split plot BBD is given as follows:
wp sp cw c cs ΠΦfrβαλ22 Φ fr Φ fβλ 4 Φ fαλ 44 Φ f α ()1212( ) ()ws()int()2 XR' −1 X = ws . wp−1 2 sp−1 ⎛⎞11−γ wp ⎛⎞⎛⎞1−γ sp Φ−fr()λβγ4 () wpsp ⎛⎞1 ⎜⎟4 ⎜⎟⎜⎟6 − 15ww ⎜⎟ ⎜⎟Φ−frλβ44⎜⎟ Φ− fr λβ⎜⎟ Φ− fr λα 41 − γ wp ⎜⎟ Φ− fr λα4 ⎝⎠11()ww⎝⎠() ww⎝⎠ 5() ss 45 ⎝⎠()ss
⎛⎞wp where cw = ⎜⎟ is the number of cross whole plots, c = wp * sp is the number of cross ⎝⎠2 ⎛⎞sp products and cs = ⎜⎟ is the number of cross subplots. ⎝⎠2
See Appendices B2 and B3 for details on the derivation of these equations. The actual
1/ p computation used for the D-criterion is XR' −1 X where p is the number of parameters to be estimated.
47 3.12 Application To demonstrate the application of these equa tions, the D-op timality criterion was computed for the BBD g iven in T ab le 2 and the CCD given in Table 6. In addition, a modified design of th e CCD i n Table 6 was also studied . This m odified design is given in Table 13. The modified de s ign ha s two who le plots l ess tha n the VKM D23 due to the fact that the all the subplot axials are pl aced in o ne wh ol e plot . T his m eans the modified design also requires few er run s.
Table 13: M odified VK M D2 3 with = α = 1, tw o who l e plots a nd thre e subplot factors – Unbalanced
Whole Plot z1 z2 x1 x2 x3 Whol e Plot z1 z2 x1 x2 x3 1 -1 -1 -1 -1 1 7 0 - 0 0 0 -1 -1 1 -1 -1 0 - 0 0 0 -1 -1 -1 1 -1 0 - 0 0 0 -1 -1 1 1 1 0 - 0 0 0 2 1 -1 -1 -1 -1 8 0 0 0 0 1 -1 1 -1 1 0 0 0 0 1 -1 -1 1 1 0 0 0 0 1 -1 1 1 -1 0 0 0 0 3 -1 1 -1 -1 -1 9 0 0 0 0 0 -1 1 1 -1 1 0 0 0 0 0 -1 1 -1 11 00 0 0 0 -1 1 1 1 -1 0 0 0 0 0 4 1 1 -1 -1 1 10 0 0 - α 0 0 1 1 1 -1 -1 0 0 α 0 0 1 1 -1 1 -1 0 0 0 - α 0 1 1 1 1 1 0 0 0 α 0 5 - 0 0 0 0 0 0 0 0 - α - 0 0 0 0 0 0 0 0 α - 0 0 0 0 - 0 0 0 0 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
The parameters for the two CCD designs are given in Table 14. The results are presented in Table 15 and are computed for three (3) commonly used variance ratios of η = 0.5, 1 and 10. Both the standardized and scaled D-criterion values are given.
48 Table 14: Design parameters for the split plot CCD Parameters VKM D23 Modified VKM D23 f 16 16
a1 12 9
a3 0 1
w1 4 4
w3 0 6 wp 2 2 sp 3 3 N 48 42
rw 4 4
rs 2 1 1 1 α 1 1 1 1 p 21 21
An examination of the critical functions conveys the influence of the design structures on the application of these functions. Note that for the VKM D23 design given in Table 6 the subplot axial points are in separate whole plots. Therefore Φ3 = 0 indicates that the design moments for the subplot axial runs are not influenced by the inverse correlation co-effect within the whole plot. The resulting impact is that both the whole plot axial points and the subplot axial points are affected by the same critical function Φ1.
In contrast, the critical functions Φ3 is active for the modified design and therefore influence the subplot design moments. Note that value of Φ3 as shown in Table 15 for the modified design is negative across the variance ratios. Therefore, the design moments for the subplot axial runs are corrected for the influence of the inverse correlation co-effect within the whole plot. For both designs Φ4 = 1 which indicates that the factorial points are not affected by the inverse correlation effect within the whole plot.
On the other hand, Φ5 = Φ1 for the VKM D23 design which also indicates that the subplot axial points are not affected by the inverse correlation effect within the whole plot.
However, for the modified design Φ5 = Φ2 meaning that the subplot ax ial points are affected by the invers e corr elation e ffect within the whole plot and is a direct consequence of placing al l the subp lot axials in the sam e wh ole p lot.
49 Table 15: Constants, functions and D-value results for CCD across the different variance ratios VKM D23 Modified VKM D23 Variance Ratio (η) 0.5 1 10 0.5 1 10 Constants
c1 0.615 0.492 0.330 0.242 0.185 0.117
c2 136.000 87.040 39.158 111.000 70.583 31.511
c3 2.294 2.294 2.294 2.676 2.689 2.705
c4 -1.765 -1.765 -1.765 -1.892 -1.896 -1.902
c5 0.52 9 0.529 0.529 0.784 0.793 0.803
c6 -1.412 -1.412 -1.412 -1.730 -1.741 -1.754
c7 0.23 5 0.2 35 0.235 0.153 0.031 -2.164
c8 1.471 1.4 71 1.471 1.577 1.580 1.585
c9 0.05 9 0.0 59 0.059 -0.153 -0.161 -0.170
c10 14.4 71 11.5 76 7.765 14.791 11.861 7.980
c11 9.52 9 7.62 4 5.113 5.460 4.253 2.760 Scalars