FIVE-FACTOR CENTRAL COMPOSITE DESIGNS ROBUST TO A PAIR OF MISSING OBSERVATIONS. Munir Akhtar Vice-Chanselor, Islamia University Bahawalpur, Pakistan.

Similar documents
8 RESPONSE SURFACE DESIGNS

Analysis of Variance and Design of Experiments-II

On the Compounds of Hat Matrix for Six-Factor Central Composite Design with Fractional Replicates of the Factorial Portion

ON D-OPTIMAL DESIGNS FOR ESTIMATING SLOPE

CONSTRUCTION OF BLOCK DESIGNS WITH NESTED ROWS AND COLUMNS

Chemometrics Unit 4 Response Surface Methodology

Addition of Center Points to a 2 k Designs Section 6-6 page 271

Experimental Design and Optimization

Design of Engineering Experiments Part 5 The 2 k Factorial Design

J. Response Surface Methodology

Weighted Least Squares

Inner Product, Length, and Orthogonality

A General Criterion for Factorial Designs Under Model Uncertainty

Definitive Screening Designs

Useful Numerical Statistics of Some Response Surface Methodology Designs

Efficiency of NNBD and NNBIBD using autoregressive model

APPLICATION OF RIDGE REGRESSION TO MULTICOLLINEAR DATA

Design and Analysis of Experiments

Contents. TAMS38 - Lecture 10 Response surface. Lecturer: Jolanta Pielaszkiewicz. Response surface 3. Response surface, cont. 4

AND JOANNA CHUNG-IN CHENG. University of Michigan. Summary

CONSTRUCTION OF REGULAR GRAPH DESIGNS AND ITS GRAPHICAL REPRESENTATION

MATH 5720: Unconstrained Optimization Hung Phan, UMass Lowell September 13, 2018

Inverse of a Square Matrix. For an N N square matrix A, the inverse of A, 1

Linear Algebra (Review) Volker Tresp 2017

CHAPTER 6 MACHINABILITY MODELS WITH THREE INDEPENDENT VARIABLES

Some optimal criteria of model-robustness for two-level non-regular fractional factorial designs

Optimality of Central Composite Designs Augmented from One-half Fractional Factorial Designs

Linear Algebra (Review) Volker Tresp 2018

MIT Spring 2015

Some Nonregular Designs From the Nordstrom and Robinson Code and Their Statistical Properties

A GENERALIZATION OF BI IDEALS IN SEMIRINGS

A New Generalised Inverse Polynomial Model in the Exploration of Response Surface Methodology

Algebra I Calculator Activities

D-Optimal Designs for Second-Order Response Surface Models with Qualitative Factors

Contents. 2 2 factorial design 4

Math 423/533: The Main Theoretical Topics

Algebra I Chapter 4 Curriculum and IXL

Residuals from regression on original data 1

Math 0031, Final Exam Study Guide December 7, 2015

Designing Two-level Fractional Factorial Experiments in Blocks of Size Two

Response Surface Methodology IV

Design of Screening Experiments with Partial Replication

ON THE DESIGN POINTS FOR A ROTATABLE ORTHOGONAL CENTRAL COMPOSITE DESIGN

VAR Model. (k-variate) VAR(p) model (in the Reduced Form): Y t-2. Y t-1 = A + B 1. Y t + B 2. Y t-p. + ε t. + + B p. where:

Alternative implementations of Monte Carlo EM algorithms for likelihood inferences

Leverage. the response is in line with the other values, or the high leverage has caused the fitted model to be pulled toward the observed response.

X- MATHS IMPORTANT FORMULAS SELF EVALUVATION 1. SETS AND FUNCTIONS. 1. Commutative property i ii. 2. Associative property i ii

Unit 12: Response Surface Methodology and Optimality Criteria

MANOVA is an extension of the univariate ANOVA as it involves more than one Dependent Variable (DV). The following are assumptions for using MANOVA:

Design Moments. Many properties of experimental designs are quantified by design moments. Given a model matrix 1 x 11 x 21 x k1 1 x 12 x 22 x k2 X =

Analysis and Linear Algebra. Lectures 1-3 on the mathematical tools that will be used in C103

Design & Analysis of Experiments 7E 2009 Montgomery

On Sampling Errors in Empirical Orthogonal Functions

Data Preprocessing Tasks

COMPARISON OF FIVE TESTS FOR THE COMMON MEAN OF SEVERAL MULTIVARIATE NORMAL POPULATIONS

One-Way Repeated Measures Contrasts

AN IMPROVEMENT TO THE ALIGNED RANK STATISTIC

Design of Engineering Experiments Chapter 5 Introduction to Factorials

Appendix IV Experimental Design

Topic 7 - Matrix Approach to Simple Linear Regression. Outline. Matrix. Matrix. Review of Matrices. Regression model in matrix form

The 2 k Factorial Design. Dr. Mohammad Abuhaiba 1

Review of Linear Algebra

CHAPTER 6 A STUDY ON DISC BRAKE SQUEAL USING DESIGN OF EXPERIMENTS

Equations in Quadratic Form

Optimal design of experiments

Optimal Two-Level Regular Fractional Factorial Block and. Split-Plot Designs

Optimal Selection of Blocked Two-Level. Fractional Factorial Designs

STATISTICS 174: APPLIED STATISTICS FINAL EXAM DECEMBER 10, 2002

MODEL ANSWERS TO THE FIRST HOMEWORK

20.1. Balanced One-Way Classification Cell means parametrization: ε 1. ε I. + ˆɛ 2 ij =

B. L. Raktoe* and W. T. Federer University of Guelph and Cornell University. Abstract

Formulation of L 1 Norm Minimization in Gauss-Markov Models

Effects of Outliers and Multicollinearity on Some Estimators of Linear Regression Model

Process Robustness Studies

Algebra Mat: Working Towards Year 6

X- MATHS IMPORTANT FORMULAS SELF EVALUVATION

Abstract. Using graphs to find optimal block designs. Problem 1: Factorial design. Conference on Design of Experiments, Tianjin, June 2006

Mohsen Pourahmadi. 1. A sampling theorem for multivariate stationary processes. J. of Multivariate Analysis, Vol. 13, No. 1 (1983),

Analysis of Microtubules using. for Growth Curve modeling.

Classes of Second-Order Split-Plot Designs

Process/product optimization using design of experiments and response surface methodology

Chaos in the Dynamics of the Family of Mappings f c (x) = x 2 x + c

Detecting and Assessing Data Outliers and Leverage Points

TAMS39 Lecture 10 Principal Component Analysis Factor Analysis

Problems. Suppose both models are fitted to the same data. Show that SS Res, A SS Res, B

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at

Workshop 7.4a: Single factor ANOVA

Response surface designs using the generalized variance inflation factors

ON THE USE OF GENERAL POSITIVE QUADRATIC FORMS IN SIMULTANEOUS INFERENCE

Factorial Within Subjects. Psy 420 Ainsworth

COMPREHENSIVE WRITTEN EXAMINATION, PAPER III FRIDAY AUGUST 26, 2005, 9:00 A.M. 1:00 P.M. STATISTICS 174 QUESTION

Worksheets for GCSE Mathematics. Quadratics. mr-mathematics.com Maths Resources for Teachers. Algebra

Model Mis-specification

Chapter 5 Matrix Approach to Simple Linear Regression

Construction and analysis of Es 2 efficient supersaturated designs

Acta Universitatis Palackianae Olomucensis. Facultas Rerum Naturalium. Mathematica

Defense Technical Information Center Compilation Part Notice

Ohio s State Tests ITEM RELEASE SPRING 2018 INTEGRATED MATHEMATICS II

7. Response Surface Methodology (Ch.10. Regression Modeling Ch. 11. Response Surface Methodology)

MA 575 Linear Models: Cedric E. Ginestet, Boston University Midterm Review Week 7

Transcription:

Journal of Research (Science), Bahauddin Zakariya University, Multan, Pakistan. Vol.2, No.2, December 2, pp. 5-5 ISSN 2-2 FIVE-FACTOR CENTRAL COMPOSITE DESIGNS ROBUST TO A PAIR OF MISSING OBSERVATIONS Munir Akhtar Vice-Chanselor, Islamia University Bahawalpur, Pakistan. Abstract: Missing Observations in Central Composite Designs (c.c.d) are well investigated since the development of Minimaxloss criterion presented by Akhtar and Prescott [986]. Here we investigate two values in different configurations of c.c.d with five factors and develop designs robust to a pair of observations under Minimaxloss criterion. Keywords: Missing values, central composite designs, robust designs, minimaxloss criterion. INTRODUCTION Central Composite Design (c.c.d) is a combination of i) a factorial experiment of size 2 k or a fraction of it, ii) 2k axial points (Star points) two on each of k axes a distance ± from the center of the design and iii) one or more points at the centre of the design. Total design points n comprises of n f factorial, n a axial and n c centre point. For five factors the c.c.d consists of 32 factorial, axial and one or more center points. Different configurations of five-factor c.c.d can be obtained by taking different replications of factorial and axial parts. Central Composite Designs with different properties can be developed by taking different values of, i.e. distance of axial points from the centre of the design. Box [954] developed orthogonal c.c.d. Box and Hunter [957] developed rotatable designs. Box and Draper [959] discussed designs robust to inadequate model. Box and Draper [975] studied robust to outliers, which are referred here as outlier robust designs. Designs robust to observations with different probability of for different observations are studied by Herzberg and Andrews [975, 976] and Andrews and Herzberg [979]. Akhtar and Prescott [986] studied the reduction in X X due to observations and developed minimaxloss criterion, which is actually minimizing the maximum loss due to observations. The loss is the relative reduction in X X due to observations. Akhtar and Prescott [987] provide a comprehensive review of robust response surface designs. Here we study two observations in a five-factor c.c.d with three different configurations of factorial (F) and axial (A) parts. These configurations are (i) design with half replicate of factorial part (/2 FA) (ii) design with one replication of factorial and axial part (FA) (iii) design with two replications of axial part (F2A). 5 J. res. Sci., 2, 2(2), 5-5

Munir Akhtar 6 DIFFERENT GROUPS OF PAIRS WITH SIMILAR LOSSES All possible pairs of observations n C 2 can be placed in groups with similar losses. The main groups are ff, fa, aa, fc, ac and cc, where f, a and c are for a factorial, an axial and a centre point respectively. Pairs of two factorial points, ff, can be further divided into groups, ff, ff2, ff3, ff4, ff5, representing pairs of two factorial points having, 2, 3, 4 and 5 signs different. Examples for ff3 are or Pairs of two axial points are of three types aa, aa and aa2. aa are pairs of two axial observations at the same axial point and only occur when axial part is replicated twice. aa and aa2 are two axial points on the same axis and on different axes respectively. The examples of aa, aa and aa2 are, and respectively. Pairs of a factorial and an axial observation fa can be further divided as fa and fa2. fa are pairs of factorial and axial observations on the same side of the cube such as,, ( and corresponding have same signs) and fa2 are pairs in which factorial and axial points are on opposite sides of the cube such as and ( and corresponding have opposite signs). MINIMAXLOSS CRITERION Akhtar and Prescott [986] developed minimaxloss criterion to find designs robust to observations. Let the postulated underlying model is y = X β ε where y is a vector of responses at n design points, X is a matrix of order n p formed from the p input variables in the response model, β is a vector of p coefficients and ε is an error vector of order n.

FIVE-FACTOR COMPOSITE DESIGN TO PAIR OF MISSING OBSERVATIONS 7 The usual least square estimates are βˆ = ( X X ) - X y Var (βˆ ) = ( X X ) - σ 2 ŷ = X ( X X ) - X. y = R. y and Var ( ŷ ) = [ X ( X X ) - X ] σ 2 = R σ 2 where R is a matrix of order n and is sometimes called hat matrix. Under D-optimality we maximize X X. For c.c.d X X is an increasing function of and is maximum at =. If there are two observations i and j, the X X for the complete design is reduced to ij X X. We want this reduction to be as small as possible. After some algebra it can be shown that ij X X = X X A ij where A ij is the diagonal values of the second compound of (I - R) matrix. For compounds of matrix see Aitken and Rutherford [964]. Akhtar and Prescott [986] defined 'loss' as the relative reduction in X X. Loss pair of observations (i, j) X X ij X X L ij = X X = - A ij = - ( - r ii ) ( - r jj ) - where r ii, r jj and r ij are corresponding elements of R. For particular n and p, ΣA ij is constant which means that ΣL ij is also constant which implies that a particular loss L ij can be reduced only at the cost of increasing other losses. All losses being equal will be ideal. The most useful criterion for reducing the losses will be to minimize the maximum loss pair of observations, which is called minimaxloss criterion. FIVE-FACTOR DEISGNS WITH HALF FACTORIAL REPLICATE Five-factor c.c.d consisting of n f = 6 points from half replicate of factorial part with highest order interaction as defining contrast, n a = points of axial part and three or more centre points. Total design points are 29 or more. The loss due to different pairs of points is studied over a range of from. to 4.. Among the possible groups of pairs, ff, ff3 and ff5 are empty. Only ff2, ff4, aa, aa2, fa, fa2, fc, ac and cc are nonempty. The computations of losses have shown that 2 r ij L ff4 > L ff2 ; L aa > L aa2 and L fa > L fa2.

8 Munir Akhtar The losses due to fc, ac and cc remain less for the whole range of as long as the number of center points is three or more. The losses due to ff4, aa and fa are plotted against in Fig.. L aa L fa L ff4 Fig. : Losses pair of observations for design with k = 5, n f = 6, n a = and n c = 3 plotted against. Table : Loss due to different pairs of observations No. of variables k = 5 Total design points n = 29 No. of parameters p = 2 No. of centre points = 3 {Half replicate of factorial part] Design Alpha =. Orthogonal Rotatable k Alpha..6644 2. X X for complete design.252e23.26e27.3378e28 due to two factorial observation.9992.9942.9876 due to two axial observation.9437.999.93 due to one axial and one factorial observation.9868.9723.9654 Overall maxloss pair of observation.9992.9942.9876 single observation.9648.932.8795 Alpha = 2.236.385E29.985.9243.9622.985.8558 Outlier robust 2.4482.245E3.9769.9437.964.9769.8385 Minimaxloss2 2.7929.654E3.976.976.9625.976**.89 Minimaxloss 3.4972.2684E34.965.9932.9668.9932.7937 * Minimaxloss due to one observation. ** Minimaxloss due to two observations. The loss L ff4 which is near for =, decreases with the increase of, L fa first decreases and then increases slowly with increasing. L aa which, for =, is less than first two losses, decreases up to = 2. and then increases with increasing. Maximumloss which corresponds to L ff4 and then to L aa is minimum when L ff4 = L aa at = 2.7929. Thus the five-

FIVE-FACTOR COMPOSITE DESIGN TO PAIR OF MISSING OBSERVATIONS 9 factor c.c.d of this configuration with n f = 6, n a =, n c = 3, and = 2.7929 is a minimaxloss design robust to a pair of observations. The losses L ff4, L aa, L fa, maximumloss, X X and maximumloss due to single observation for different designs of the same configuration are shown in Table. From now on minimaxloss and minimaxloss2 designs mean minimaxloss designs robust to one and two observations respectively. For robustness to a pair of observations, outlier robust design with = 2.4482 is second best after minimaxloss2 design. When there is a single observation the minimaxloss2 design performs better than all designs other than minimaxloss design. The latter does not perform well when two observations are.. 29.6644 29 2. 29 2.236 29 2.4482 29 2.7929 29 3.4972 29 Table 2: Variances of parameter estimates for complete design and for designs with a pair of observation. No. of variables k = 5 Total design points n = 29 No. of parameter p = 2 No. of centre points = 3 {Half replicate of factorial part] Alpha n Inter- Linear Variances of Parameters Estimates Sum of Linear Quadratic Quadratic Inter- Variance cept.82.232.3333.232.242.2484.3333.237.343.37.3333.3.3333.3333.3333.333.37.3227.3333.362.2356.289.3333.259.349.235.3333.547 (min).556.667.556.667.464.799.464.795.47.625.47.67.538.538.853.529.357.477.357.468.36.398.36.392.247.288.247.285 (max).556.474.625.227.464.42.625.23.47.884.625.92.385.692.625.3.357.477.357.468.36.49.625.94.247.353.625.883 (max).477.436.8333.436.652.776.86.72.48.457.52.43.39.32.333.34.23.23.232.23.33.35.37.34.47.5.56.48 ff- A pair of factorial observations on far corners of cube. aa- A pair of axial observations on the same axis. fa- A factorial observation and a nearest axial observation. (max).477.436 5.625.8578.652.776.3888.394.48.457.875.852.39.32.558.653.23.23.55.53.33.35.745.34.47.5.2268.57 action.625.5876.625.23.625.2358.625.224.625.677.625.75.625.437.625.7.625.343.625.964.625.34.625.98.625.3.625.868 3.492 7.778 5.6254 5.6254.397 3.6 2.297 2.399.345 2.3277.5833.9984.35 2.534.4638.8458.236.8897.46.796.856.757.3767.583.972.532.3688.2598 The variances of the parameter estimates for the designs under discussion and corresponding to pairs from groups ff4, fa and

Munir Akhtar aa, are shown in Table 2. The variances for intercept are /3 or less. The increase in the variance of the quadratic parameter estimate corresponding to pairs of two axial points is comparatively large. The increase in the variance of the interaction parameter estimate due to a pair of two factorial points is also large. The variances of the linear, quadratic and interaction parameter estimates for complete or reduced minimaxloss2 designs are small. FIVE-FACTOR DESIGNS WITH SINGLE REPLICATE OF FACTORIAL AND AXIAL PARTS Five-factor c.c.d with one replication of factorial and axial parts have n f = 32, n a = and three or more center points. There are 45 or more design points. From the possible groups of pairs of observations, aa is empty. The losses due to different pairs of observations have been studied for a range of from. to 3.. The computations of losses have shown that L ff5 > L ff > L ff3 > L ff4 > L ff2, L aa > L aa2 and L fa > L fa2. The losses L ff5, L fa, and L aa are plotted against in Fig. 2. It is observed from this figure that L ff5 decreases and L fa increases with increasing. L aa decreases and increases with increase in and is minimum for =2.865. Thus the five-factor design of this configuration with = 2.865 is a minimaxloss2 design robust to two observations. Fig. 2: Losses pair of observations for design with k = 5, n f = 32, n a = and n c = 3 plotted against.

FIVE-FACTOR COMPOSITE DESIGN TO PAIR OF MISSING OBSERVATIONS Table 3: Loss due to different pairs of observations No. of variables k = 5 Total design points n = 45 No. of parameters p = 2 No. of centre points = 3 Design Alpha =. Orthogonal Minimaxloss2 Outlier robust Alpha = k Rotatable Minimaxloss Alpha..7244 2.865 2.265 X X for complete design.236e28.6895e3.62e33.2265e33 due to two fact. obs..7792.7668.7556.7542 due to two axial obs..942.8983.8784.8788 due to one axial and one fact obs..745.7494.756.7527 Overall maxloss pair of obs..942.8983.8784**.8788 single obs..4976.58.5296.5324 2.236 2.3784.745.5795E33.24E34.464E25.756.7465.7822.8839.8986.9477.757.7656.742.8839.8986.9477.5429.562.4953* * Minimaxloss due to one observation. ** Minimaxloss due to two observations The losses L ff5, L fa, L aa and maximum losses due to one or two observations for different c.c.d of the same configuration are shown in Table 3. The maximumloss pair of observations for minimumloss2 design is very near to outlier robust design because the values for both designs are close. The maximum losses for other designs are higher. Minimaxloss design is in the middle of the designs discussed here. The variances of the parameter estimates for the designs under discussion and corresponding to pairs from ff5, fa and aa are shown in the Table 4. The variance of the linear and the interaction parameter estimates are small for all complete and reduced designs. The variances of the quadratic parameter estimates are also small for complete designs with >. The increases pair of observations in these variances are also small except the increase in one quadratic parameter estimate pair of axial observations. FIVE-FACTOR DESIGNS WITH AXIAL PART REPLICATED TWICE Five-factor designs with complete factorial replicate and two replications of the axial part have n f = 32, n a = 2 and n c = 3 or more. There are 55 or more design points. The losses due to different pairs of observations have been studied for three center points and over a range of from. to 4.. The computations of losses have shown that for the whole range of studied L ff5 > L ff > L ff3 > L ff4 > L ff2, L aa > L aa > L aa2 and L fa > L fa2. The maximumloss in minimum for = 2.7547. Thus the five-factor design of this configuration with = 2.7547 is a minimaxloss2 design robust to a pair of observations.

2 Munir Akhtar Table 4: Variances of parameter estimates for complete design and for designs with a pair of observation. No. of variables k = 5 Total design points n = 45 No. of parameters p = 2 No. of centre points = 3 Alpha n Variances of Parameters estimates Sum of Inter- Linear Linear Quadratic Quadratic Interaction Variance cept. (min) (max) (max) (max). 45.7244 45 2.865 45 2.265 45 2.236 45 2.3784 45.745 45.73.75.3333.26.2223.2247.3333.237.39.395.3333.324.3254.3257.3333.3268.3333.3333.3333.3333.322.329.3333.3223.92.92.3333.3.294.39.294.3.264.282.264.277.246.262.246.257.244.259.244.255.238.253.238.249.23.245.23.24.33.329.33.32.294.39.33.33.264.282.33.326.246.262.33.32.244.259.33.32.238.253.33.32.23.245.33.32.33.329.33.33.46.463.8333.433.565.57.942.63.356.359.44.37.34.344.48.353.33.34.333.38.254.254.26.255.6289.6292 3.3829.723.46.463 5.32.76.565.57.2972.877.356.359.39.534.34.344.335.54.33.34.246.467.254.254.223.4.6289.6292 24.766 2.85 ff- A pair of factorial observations on far corners of cube. aa- A pair of axial observations on the same axis. fa- A factorial observation and a nearest axial observation..33.375.33.332.33.374.33.33.33.373.33.33.33.373.33.33.33.372.33.33.33.372.33.33.33.375.33.332 2.5966 2.6732 9.593 3.434.953.253.4566.445.9323.26.92.9888.935..7.9842.963.984.32.9656.8749.942.9959.9244 8.6986 8.7758 39.97.3328 Fig. 3: Losses pair of observations for design with k = 5, n f = 32, n a = 2 and n c = 3 plotted against.

FIVE-FACTOR COMPOSITE DESIGN TO PAIR OF MISSING OBSERVATIONS 3 Table 5: Loss due to different pairs of observations No. of variables k = 5 Total design points n = 55 No. of parameters p = 2 No. of centre points = 3 {Axial observations replicated twice.} Design Alpha X X for complete design due to two fact. obs. due to two axial obs. due to one axial and one fact obs. Overall maxloss pair of obs. Missing single obs. Alpha =. Orthogonal Rotatable k Alpha = Outlier robust Minimaxloss2 Minimaxloss..5774 2. 2.236 2.4978 2.7547 3.5293.4427E29.7495E32.385E34.334E35.4536E36.659E37.5793E4.774.7594.747.7282.723.797.749 * Minimaxloss due to one observation. ** Minimaxloss due to two observations..5367.5834.667.638.677.797.7994.6222.622.634.622.692.6275.643.774.7594.747.7282.723.797**.7994.4976.58.5296.5324.5429.562.4953* Table 6: Variances of parameter estimates for complete design and for designs with a pair of observation. No. of variables k = 5 Total design points n = 55 No. of parameters p = 2 No. of centre points = 3 Alpha n {axial part replicated twice 3} Variances of Parameters estimates Sum of Inter- Linear Linear Quadratic Quadratic Inter- Variances cept. (min) (max) (max) (max) action. 55.5774 55 2. 55 2.236 55 2.4978 55 2.7547 55.5293 55.646.648.74.68.357.378.453.4.28.2824.2844.2844.3333.3333.3333.3333.276.284.2796.2789.924.24.975.972.825.92.846.859.278.299.278.293.238.253.238.249.28.29.28.26.92.2.92.99.76.83.76.8.6.67.6.65.22.26.22.25.278.299.288.34.238.253.258.27.28.29.233.244.92.2.29.229.76.83.25.23.6.65.9.98.22.26.53.57.243.247.24.293.44.4.43.48.262.267.27.268.22.222.222.222.48.48.48.48.89.9.89.89.25.25.25.25.243.247.259.2542.44.4.487.476.263.267.297.294.22.222.25.25.48.48.75.75.89.9...25.25.33.33 ff- A pair of factorial observations on far corners of cube. aa- A pair of axial observations on the same axial. fa- A factorial observation and a nearest axial observation..33.375.33.33`.33.374.33.33.33.372.33.33.33.37.33.33.33.37.33.329.33.37.33.329.33.373.33.329.5376.628.685.6383.7693.8436.874.833.8279.8978.8467.863.8525.955.864.8792.75.86.7649.7785.6296.75.6452.6589.4684.54.4784.4935

4 Munir Akhtar The losses L ff5, L aa, L fa are plotted against in Fig. 3. The losses L ff5, L aa, L fa and the maximum losses due to one or two observations for minimaxloss2 and other designs of the same configuration are shown in Table 5. The maximum loss pair of observations in minimaxloss2 design is.797 as compared to.723 for outlier robust design. The maximum losses for other designs are larger. When there is only one observation the minimaxloss2 design performs better than all other designs except minimaxloss design. The variances of the parameter estimates for the designs under discussion and corresponding to a pair of observations from groups ff5, aa and fa are shown in Table 6. Var (β o ) is around.2 for minimaxloss2 design. The variance of the linear parameter estimates for all complete and reduced designs are small. The variance of the quadratic parameter estimates are also small for all complete and reduced designs with >. For designs with = these variances are around.25. It is interesting to note that all the increases in the variances due to different pairs of observations are small. All the variances are very small for complete and reduced minimaxloss and minimaxloss2 designs. CONCLUSIONS Five-factor c.c.d robust to a pair of observations, for different configuration are given in Table 7. The losses due to two and one observation given in the last two columns of Table 7 are minimaxlosses for that configuration. If there is risk of a pair of observations in the experiment then the minimaxloss2 designs developed for each configuration is recommended. The configuration with factorial part replicated twice has not been discussed as the for minimaxloss2 design in that case is less than one, which means the axial points are inside the cube of factorial points. This design also loses its natural balance of factorial and axial points in favour of factorial points. The problem can be investigated further for more factors and larger number of observations. Table 7: Five-Factor c.c.d robust to a pair of observations. n f n a n c N X X Minimaxloss2 Minimaxloss 6 3 29 2.7929.654E3.976.89 32 3 45.865.62E33.8784.5296 32 2 3 55 2.7547.659E37.797.433 References Aitken, A.C. and Rutherford, D.E. (964) Determinants and Matrices, Oliver and Boyd Ltd., U.K. Akhtar, M. and Prescott, P. (986) Response Surface Designs Robust to Missing Observations, Commun. Statist. Simula., 5(2), 345-363.

FIVE-FACTOR COMPOSITE DESIGN TO PAIR OF MISSING OBSERVATIONS 5 Akhtar, M. and Prescott, P. (987) A Review of Robust Response Surface Designs, Pakistan J. Statist., 3B, -26. Andrews, D.F. and Herzberg, A.M. (979) The Robustness and Optimality of Response Surface Designs, J. Statist. Plan. Inf., 3, 249-257. Box, G.E.P. (954) The Exploration and Exploitation of Response Surfaces: Some General Considerations and Examples. Biometrics,, 6-6. Box, G.E.P. and Draper, N.R. (959) A Basis for the Selection of Response Surface Designs, J. Amer. Statist. Assoc., 54, 622-654 Box, G.E.P. and Draper, N.R. (975) Robust Designs, Biometrika, 62, 347-352. Box, G.E.P. and Hunter, J.S. (957) Multifactor Experimental Designs for Exploring Response Surfaces, Ann. Math. Statist., 28, 95-24. Herzberg, A.M. and Andrews, D.F. (975) The Robust Designs of Experiments Bull. Inst. Statist., 45(), 37-374. Herzberg, A.M. and Andrews, D.F. (976) Some consideration in the Optimal Design of Experiment in Non-Optimal Situations, J. Roy. Statistics Soc. B, 38, 284-289.