Sie sind auf Seite 1von 18

This article was downloaded by: [190.80.134.

42] On: 14 October 2013, At: 17:25 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Quality Engineering
Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/lqen20

A Retrospective View of Mixture Experiments


John A. Cornell
a a

Department of Statistics , University of Florida , Gainesville, Florida Published online: 29 Aug 2011.

To cite this article: John A. Cornell (2011) A Retrospective View of Mixture Experiments, Quality Engineering, 23:4, 315-331, DOI: 10.1080/08982112.2011.602283 To link to this article: http://dx.doi.org/10.1080/08982112.2011.602283

PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the Content) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http:// www.tandfonline.com/page/terms-and-conditions

Quality Engineering, 23:315331, 2011 Copyright # Taylor & Francis Group, LLC ISSN: 0898-2112 print=1532-4222 online DOI: 10.1080/08982112.2011.602283

A Retrospective View of Mixture Experiments


John A. Cornell Department of Statistics, University of Florida, Gainesville, Florida

Downloaded by [190.80.134.42] at 17:25 14 October 2013

ABSTRACT When approached to present A Retrospective View of Mixture Experiments at the 2010 Fall Technical Conference (FTC), I was happy to do so. Why? Because I have been retired as a former member of the faculty of the Department of Statistics for 7 years and I had the opportunity to speak in front of many friends on a subject that has been part of my life for 45 years, since 1966. Following the invitation to speak, I looked up the meaning of retrospective in my copy of Websters Dictionary to nd that retrospective means to reminisce or re-experience past events. In other words, I was asked to look back over the past 55 years or so of published research on mixture experiments and share my thoughts with friends, many of whom were in the audience at the FTC and others who will read this article in Quality Engineering.
KEYWORDS blending ingredients, canonical polynomial, composition space or , John Gorman, experimental region, (q-1)-dimensional simplex, Henry Scheffe linear blending, mixture experiments, nonlinear blending, nonnegative proportions, polyester yarn, fq, mg simplex-lattice, special-cubic model, upper and lower bounds

INTRODUCTION
One spring afternoon in 1966 when I was assigned to the Applied Math Group at Tennessee Eastman Company (TEC) in Kingsport, Tennessee, my supervisor, Bob Brown, approached me and asked whether I had read in 1958 titled Experiments with an article authored by Henry Scheffe Mixtures. Mr. Brown was given a set of data, passed on to him from a colleague working in the Polymer Division of TEC, and asked whether he could determine whether elongation of polyester yarn, made by blending three different polymers, could be optimized. When I said I had not read s article. He the article, Mr. Brown gave me an assignment to read Scheffe wanted to know whether the elongation of spun yarn was a function of the proportions of the three polymers that were mixed together and, if so, did I know how to determine what the optimal blend was? Unfortunately, I had not been taught during my graduate training how to nd the optimal blend, but the challenge somehow seemed exciting. Forty-ve years later, we are able to obtain a chronological listing of authors of selected statistical literature on mixtures from 1953 to 2009 in Cornell (2011, pp. 1415). A fully
315

Address correspondence to John A. Cornell, 2836 SW 92nd Terrace, Gainesville, FL 32608. E-mail: jcornell@stat.u.edu

comprehensive summary of statistical research related to mixture experimental designs, models, and data analysis topics from 1955 to 2004 can be found in Piepel (2006). Drs. Peter W. M. John and John W. Gorman copresented the rst invited talk on mixtures at the 1960 Gordon Research Conference on Statistics in Chemistry and Chemical Engineering, held at New Hampton, New Hampshire. The title of the talk was Experiments with Mixtures and highlighted the in 1958. designs and models introduced by Scheffe

a q component (or q ingredient) mixture in which xi represents the proportion of the ith component present in the mixture, the proportions are nonnegative: 0 xi 1; i 1; 2; . . . ; q 1

and sum to unity or one,


q X i 1

xi 1:

THE MIXTURE PROBLEM


The description of a mixture experiment is as follows. Imagine the sweetening of a cup of coffee by adding sugar or a sweetener or both sugar and the sweetener to the coffee. Blends or mixtures of coffee with sugar or with the sweetener or with both are possible. Another example is the mixing of oil with vinegar and some spices to create and avor a salad dressing. A third example might be the mixing of 93-octane fuel in the family car that contains 87-octane fuel to improve driving performance; for example, improving miles per gallon. In these examples, the adding and=or blending of ingredients in an attempt to try to obtain a more desirable end product is something all of us do in our everyday activities, and these actions are known as mixture experiments; see Figure 1. In mixture experiments, the controllable variables are nonnegative proportionate amounts of the ingredients in a mixture in which the proportions are by volume, by weight, or by mole fraction. In

Downloaded by [190.80.134.42] at 17:25 14 October 2013

The composition space, or experimental region, of the q components, by virtue of restrictions [1] and [2], takes the form of a (q 1)-dimensional simplex. When q 2, the simplex is a straight line; when q 3, the simplex is an equilateral triangle; and when q 4, the simplex is a tetrahedron; see Figure 2. When additional constraints are imposed on the component proportions in the form of lower Li and=or upper Ui bounds, 0 Li xi Ui 1; i 1; 2; . . . ; q 3

or as the linear multicomponent constraints, Cj A1j x1 A2j x2 . . . Aqj xq Dj ; j 1; 2; . . . ; h; 4 where Aij, Cj, and Dj are scalar constants. These additional constraints [3] and [4] can alter the shape of the experimental region from that of a simplex to one of an irregularly shaped convex polyhedron inside the simplex. Thus, when we think about experimental regions for mixture experiments, two main shapes come to mind: a simplex-shaped region

FIGURE 1 Some common examples of mixing ingredients: (a) sweetening coffee by mixing sugar and a sweetener; (b) mixing oil and vinegar to avor a salad; (c) blending 87-octane and 93-octane fuels.
J. A. Cornell

FIGURE 2 Experimental regions for q 3 and q 4 components. For q 3, the simplex is an equilateral triangle. For q 4, the simplex is a tetrahedron.
316

(generally the whole simplex) or a non-simplexshaped region inside the simplex. In both cases, the dimensionality of the region is lower than q, the number of components. Mixture model forms most commonly used in tting mixture data, at least until 1965, were the (1958, canonical polynomials introduced by Scheffe 1963). The rst-degree or linear blending model is yu
q X i1

bi xiu eu ;

u 1 ; 2 ; . . . ; N

Downloaded by [190.80.134.42] at 17:25 14 October 2013

where yu represents the uth observed value of the response in N trials and the constant term b0 is not included in the model due to restrictions [1] and [2]. The coefcient bi is the expected response to component i (i.e., the response at xi 1), and eu is the random error in the uth observed response value. We assume that the error eu is sampled from a distribution with mean 0 and variance r2 . In the case of testing the magnitude of the estimate of one or more of the bis in the model during analysis of the data, the errors in the individual yu values are assumed to be sampled from a normal N(0, r2 ) distribution. The second-degree or binary nonlinear blending model is yu
q X i1

bi xiu

q1 X q X i1 j 2

bij xiu xju eu

where bij is a measure of the nonlinear (curvilinear) blending of components i and j, i, j 1, 2, . . ., q and i < j. Higher-degree models contain nonlinear blending terms of the type bijkxixjxk and cijxixj(xi xj), and the canonical forms [5] and [6] will always contain a lower number of terms than their standard polynomial counterparts; see, for example, Cornell (2002, chapter 2).

THE BEGINNING YEARS1955 TO 1965


Many years ago it was rumored that Henry Scheffe was a perfectionist when writing papers for publication. Rarely was he asked to revise a manuscript following a review of the rst draft. Figure 3 displays the rst page of the initial draft of the manuscript that later became the seminal article entitled Experiments with Mixtures that appeared in the Journal of the Royal Statistical Society Series B. It was also
317

rumored that nearly 80% of the completed 34-page manuscript was composed during a cross-country ight from the East Coast of the United States to Berkeley, California, in the spring of 1957. This page, being the rst page of the manuscript, has two corrections on it. The rst misprint is a corrected spelling of the word resistance in line 3 and the second misprint is the crossing out of gasolines, and replaced with of different gasolines in line 5. In the remaining 33 pages of the manuscript, only seven additional corrections were discovered by Henry or the copy editor. For many of us who never had the , a couple of pleasure of meeting Henry Scheffe pictures of Henry can be seen at http://www-history. mcs.st-and.ac.uk/PictDisplay/Scheffe.html. s (1958) seminal paper Three years prior to Scheffe on simplex-lattice designs and the associated canonical form of polynomial model, Claringbold (1955) proposed a three-component simplex (equilateral triangle) along with a second-degree model in two independent variables with which to study the joint action of three separate hormones estrone (x1), estradiol (x2), and estriol (x3) that were injected into groups of mice. The design consisted of two 10-blend sets of points administered at each of three different amounts (low, middle, and high). One 10-blend set is shown by the black dots positioned on the perimeter (vertices and edges) and center of the triangle in Figure 4. The second 10-blend set is shown by the boxes positioned at the vertices, middle of the edges, inside the triangle, and at the center of the triangle. The 20-blend design is illustrated in Figure 4. Although many authors of articles on mixtures in the 1960s and 1970s acknowledged Claringbolds s (1958) (1953) work, most acknowledged Scheffe paper to have a greater impact on published research later than Claringbolds due to the fact that refer s paper was nearly triple the number ences to Scheffe of references to Claringbolds paper. Figure 5 displays the q 3- and q 4-component fq, mg simplex-lattice designs for tting the second- and models, respectively. In the third-degree Scheffe fq, mg notation, q is the number of components and m is the degree of the model to be tted. In the fq, mg lattice, each component is assigned the m 1 equally spaced values ranging from xi 0, 1=m, 2=m, . . . , m=m 1 and because of constraint [2], with each blend or design point, the sum of the xis must equal one or unity.
A Retrospective View of Mixture Experiments

Downloaded by [190.80.134.42] at 17:25 14 October 2013

s handwritten manuscript; (b) Page 1 of initial draft of the manuscript titled Experiments with FIGURE 3 (a) The rst page of Scheffe Mixtures.

(1963) Five years after his initial paper, Scheffe introduced the simplex-centroid design and its associated model. The simplex-centroid design consists of 2q 1 points: q permutations of (1, 0, 0, . . ., 0), q(q 1)=2 permutations of (1=2, 1=2, 0, 0, . . . ,0), q(q2 3q 2)=6 permutations of (1=3, 1=3, 1=3, 0, 0, . . . , 0), . . ., and one permutation of (1=q, 1=q, 1=

q, . . . , 1=q). The 2q 1 points are the vertices, midpoints of the edges, centroids of the faces, etc., of the simplex. The corresponding model to be tted to data at the points of the simplex-centroid design is q q 1 X q X X y bi xi bij xi xj
i1 i1 i<j

q 2 X q 1 X q X i1 i<j j <k

bijk xi xj xk ::: b123:::q x1 x2 x3 :::xq e 7

FIGURE 4 Design points for Claringbolds study of three hormones. The design consists two 10-blend sets administered at each of three amounts. One set is denoted by dots . and the other set is denoted by boxes &. The three vertices (single hormone) and the centroid (all three hormones) are in each set.
J. A. Cornell

The number of terms in model [7] is 2q 1 and there exists a one-to-one correspondence between the number of points (or blends) in the design and the number of terms in the model. Figure 6 shows the simplexcentroid designs for q 3 and q 4. In addition to the simplex-centroid design, Scheffe (1963) introduced the inclusion of process variables in mixture experiments. Process variables are not mixture components but instead are independent
318

FIGURE 5 Some f3, mg and f4, mg simplex-lattice arrangements for m 2 and m 3.

variables such as temperature, time, spin speed, etc., that could have an effect on the response or affect the blending properties of the mixture components. When considering designs for including process variables in mixture experiments, the most popular strategy is to cross the points of the mixture design with the points of a factorial arrangement in the process variables. Similarly, combined models containing both the mixture components and the process variables consist of combining or crossing the terms in the respective model forms. Gorman and Hinman s designs and (1962) extended the work of Scheffe models by expressing the estimation equations for

the parameters in the full cubic or third-degree and full quartic or fourth-degree models in terms of the average response values at the points of the full cubic fq, 3g and quartic fq, 4g simplex-lattice designs in q components. The formulas for estimat cubic and ing the coefcients in the tted Scheffe quartic models can also be found in Appendix 2B of Cornell (2002). The second half of the 1960s found modelers transforming the q mixture components into q 1 mathematically independent variables. Two reasons cited by those who chose to work with independent variables are (1) familiarity in knowing how to set up designs, such as factorial arrangements, and t standard polynomial models in the independent variables; and (2) belief in knowing how to interpret the unknown parameter estimates in the tted models up to the third degree, as well as familiarity with design optimality criteria. Figure 7 displays a 12-point design in three components, or two independent variables, w1 and w2, suggested by Draper and Lawrence (1965). The design consists of three sets of points (one set being triangular, a second a set of four clear dots, and a third set of four dark dots; see gure). Draper and Lawrence also introduced designs that minimize the bias in the tted model, or designs that minimize the variance of prediction, or designs that minimize the integrated mean square error of the estimate of the response over the simplex region. The differences among the designs were the result of spreading the points away from the center point. Minimizing the integrated mean square error of the predicted response value (w1, w2) was initially suggested by Box and Draper y (1959) when constructing response surface designs in general.

Downloaded by [190.80.134.42] at 17:25 14 October 2013

FIGURE 7 Draper and Lawrences (1965) 12-point design FIGURE 6 Simplex-centroid designs for q 3 and q 4 components. consisting of three sets of points (one triangular and two squares) plus two center-point replicates.

319

A Retrospective View of Mixture Experiments

THE SECOND WAVE OF PROBLEMS TO BE SOLVED, 19661978: ENTER THE COMPUTER


When one or more of the q component proportions are restricted by lower bounds 0 Li xi, i 1, 2, . . . , q, Kurotori (1966) suggested transforming the restricted components to pseudocomponents, which were later called L-pseudocomponents, using xi0 xi Li ; q P 1 Li
i1

on the component proportions, such as 0 Li xi Ui 1; i 1; 2; . . . ; q 10

i 1; 2; . . . ; q

Downloaded by [190.80.134.42] at 17:25 14 October 2013

Pq The quantity 1 i1 Li must be greater than zero Pq (that is, i1 Li < 1) so that the simplex in the x0 i lies entirely inside the simplex in the x0 i. The denomiPq nator 1 i1 Li in [8] is the height of the simplex in the L-pseudocomponents x0 i compared to unity for the simplex in the x0 i. Once the L-pseudocomponent simplex is set up inside the original simplex, s simplex-lattice or simplex-centroid designs Scheffe are easily set up in the L-pseudocomponents and -type models are easily tted in either or the Scheffe both the original or the L-pseudocomponents. Let us illustrate the use of L-pseudocomponents in three components having the lower bound restrictions 0:20 x1 ; 0:10 x2 ; 0:25 x3 9

McLean and Anderson (1966) developed a technique whereby the vertices of the constrained region (convex polyhedron) can be located. The vertices and convex combinations of the vertices can be used as design points from which to collect data for tting -type models. Unfortunately, once the the Scheffe limits or bounds on the component proportions are specied, the possibility of obtaining a uniform distribution of design points over the constrained region or factor space is doubtful. When this nonuniformity of the distribution of design points exists, that is, when there are clusters of points in some areas of the factor space and only a few points in other areas, it is unlikely that the tted model will adequately predict the response over the factor space. This is because the variance of the estimate of the response will be affected by the distribution of design points; poor precision will result in areas of sparse experimentation, though there will be good precision in areas with clusters of points. An example provided by McLean and Anderson (1966) is the development of ares using the chemical constituents magnesium (x1), sodium nitrate (x2), strontium nitrate (x3), and binder (x4) with the following constraints: 0:40 0:10 x1 x3 0:60; 0:10 0:50; 0:03 x2 x4 0:50; 0:08 11

P3 where i1 Li 0.20 0.10 0.25 0.55 and 1 L 0.45. In Figure 8, the L-pseudocomponent simplex inside the simplex in the original components is shown. The orientations of the L-pseudocomponent simplex and the original simplex are the same. When the restrictions on the component proportions are in the form of lower and upper bounds

The constrained region dened by [11] has 8 extreme vertices, 12 edges connecting the vertices, and 6 two-dimensional faces as shown in Figure 9. McLean and Anderson (1966) chose the 8 vertices, the centroids of the 6 faces, and the overall centroid of the region as 15 design points in which to collect data quadratic model for tting the 10-term Scheffe y b1 x1 b2 x2 b3 x3 b4 x4 b12 x1 x2 b13 x1 x3 b14 x1 x4 b23 x2 x3 b24 x2 x4 b34 x3 x4 e 12 Although there are shortcomings to the extreme vertices design when constraints of the type in [11] exist, in particular 0.03 x4 0.08, in 1966 computer software for generating optimal designs as we know
320

FIGURE 8 The L-pseudocomponent (Xi0 ) simplex inside the original component simplex (Xi) i 1,2,3.
J. A. Cornell

FIGURE 9 McLean and Andersons (1966) 15-point constrained region design for tting the quadratic model in four components.

FIGURE 10 Ellipsoidal regions for three components. The inner (solid) ellipse corresponds to the unit spherical region in W1 and W2 and the larger (dashed) ellipse corresponds to the largest spherical region, centered at x0, that will t inside the simplex.

it today did not exist. Therefore, McLean and Andersons paper was an important contribution that soon became the impetus or stimulus for other researchers to develop algorithms for generating optimal designs. Gorman (1966) discussed ways to modify some of the constraints in an effort to remove clusters of points. Thompson and Myers (1968) further expanded the design possibilities by dening an ellipsoidal region of interest centered about a point of maximum interest in the simplex. The point of maximum interest might be the mixture that the current product is made of or just a convenient starting point for the experimentation as well as a base point from which to construct a design. The ellipsoidal region of interest is contained entirely within the simplex, and the shape of the region is determined by the experimenter. See discussion of Cornell and Good (1970). Thompson and Myers (1968) showed how polynomial models of any degree can be used to estimate the response over the region of interest by rst transforming from the set of q mixture components to a set of q 1 linearly independent variables. The transformation to independent variables (Draper and Lawrence [1965a, 1965b]) enables the use of standard methods of design construction (they suggest using rotatable designs) as well as facilitates the use of the criterion the average mean square error of the estimate of the response for determining optimal design congurations. An illustration of a region of interest for three components, centered at the point of main interest x0, is shown in Figure 10.
321

Downloaded by [190.80.134.42] at 17:25 14 October 2013

s (1958) simplexA generalization of Scheffe lattices and canonical form of polynomials was considered by Lambrakis (1968, 1969) where he dened the following: let the p mixture components be dened as major components or M-components and the proportion of the ith (1 i p) M-component present in the mixture is denoted by ci, so that ci > 0; i 1; 2; . . . ; p;
p X i1

ci 1:

13

Each M-component is a mixture of ni ! 2 minor or m-components. Let us denote by Xij the proportion of the jth minor component in the ith major component as 0 Xij ci ; i 1; 2; . . . ; p; j 1; 2; . . . ; ni ; 14

Pni and j 1 Xij ci in the simplex corresponding to the ith M-component. The proportion of the jth m-component of the ith M-component is then represented by Xij cixij, so that
p X ni X i1 j 1

Xij 1:

15

Because each M-component is a mixture of ni m-components, each major component can be represented geometrically by a regular (ni 1) s (1958) dimensional simplex. If we recall Scheffe fq, mg lattice notation, and realizing that a -type polynomial may be tted to the points Scheffe of the fni, mig lattice associated with the ith Mcomponent, then Lambrakiss (1968, 1969) generalization of the simplex-lattice design is the result of
A Retrospective View of Mixture Experiments

combining the p fni, mig simplex lattices. These combinations are called multiple lattices and consist of Pp i 1  n i mi 1 mi 

expected values of y1 and y2 over the M-components 1 and 2, respectively. The corresponding 12-term model becomes double-Scheffe g12 g1 g2 c13 x1 x3 c14 x1 x4 c23 x2 x3 c24 x2 x4 c123 x1 x2 x3 c124 x1 x2 x4 d1234 x1 x2 x3 x4 c134 x1 x3 x4 c234 x2 x3 x4 c123 x1 x2 x3 x1 x2 c124 x1 x2 x4 x1 x2 d1234 x1 x2 x3 x4 x1 x2 16 A slightly different approach to categorizing the components and blending the categories was considered by Cornell and Good (1970). The q mixture components are assumed to belong to k rather than p (k ! 2) distinct categories where a category is a group of components considered to be similar; for example, a category of acid constituents, a category of bases, etc. (Let us use the notation that Cornell and Good used by switching the total number of M-components from p to k and by letting k be the number of categories and q be the total number of components in the k categories.) The number of categories of mixture components is general (k < 1), and each category is represented in every mixture by one or more of its member components; that is, if ni ! 2 components belong to the Pni Pk ith category, then xij ci , i1 ci 1, and j 1 Pk i1 ni q : Similar to the approach used by Thomson and Myers (1968), Cornell and Good (1970) assumed that the experimenters interest is concentrated in a region of interest centered at a point of main interest denoted by x0. The region of interest in the space of the mixture components is dened analytically to be ellipsoidal in shape and of the form  k  X xi x0i 2
i1

Downloaded by [190.80.134.42] at 17:25 14 October 2013

points or blends. To illustrate the conguration of the double-lattice for p 2 M-components where each M-component contains two M-components, n1 n2 2, suppose that it is desired to t a third-order polynomial to approximate the surface over the lattice corresponding to M-component 1 and to t a second-order polynomial to the surface over the lattice corresponding to M-component 2. Let us set c1 c2 1 2, which forces each M-component to make up 0.50 or 50% of the mixture. Denoting the two m-components in M-component 1 by x1 and x2 and denoting the two m-components in M-component 2 by x3 and x4, respectively, the double-lattice conguration is the result of multiplying or crossing the 4 points of a f2, 3g simplex-lattice and the 3 points of a f2, 2g simplex-lattice, producing the 12 points of the fn1, n2; m1, m2g f2, 2; 3, 2g double-lattice shown in Figure 11. model is constructed The 12-term double-Scheffe by crossing the terms in the two single-lattice models. Let the cubic model in the m-components x1 and x2 be of the form g1 a1x1 a2x2 a12x1x2 c12x1x2(x1 x2) and the quadratic model in the minor components x3 and x4 be of the form g2 b1x3 b2x4 b12x3x4 where g1 and g2 represent the

hi

17

FIGURE 11 A 12-point double-lattice congulation. The blending of M-component 1 is vertical and the blending of M-component 2 is horizontal.
J. A. Cornell

where the x0i and hi (1 i k) are chosen by the experimenter. The x0i denotes the center of the interval of interest for the ith component and hi is a constant that allows for the spread of the symmetric interval of interest for the ith component. This denition of a region of interest isolates the attention to a specic area of interest and therefore enables one to ignore other areas of the simplex factor space
322

Downloaded by [190.80.134.42] at 17:25 14 October 2013

that are not of interest. The factor space is a (q k)dimensional convex polytope. A transformation of the q linearly dependent mixture components is made to q k mathematically independent variables, which are denoted by w1, w2, . . ., wq-k. In addition to removing the physical units associated with the mixture components, the transformation simplies the overall problem in that the region of interest becomes a unit sphere in the metric of the wi centered at wi 0 (1 i q k), which is much easier to work with than an ellipsoidal region is, particularly when constructing designs. To facilitate the understanding of the methodology, an example is presented involving the production of polyethylene teraphthalate (a thin plastic lm for coating) from mixtures of two acids with two glycols. The relation between the mixture components and the independent design variables is discussed and formulas are derived for the relationship. Although the two- and three-category situations were illustrated in detail by Cornell and Good (1970), the methods are completely general for larger numbers of categories. See Figure 12 where two acids x1 and x2 are blended with two glycols x3 and x4. The next 10 to 15 years brought authors from the industrial sector. From DuPont in particular, Drs. Ronald Snee and Don Marquardt were quick to introduce the use of computers for designing and modeling mixture data. The following articles were

a must-read for researchers that posed questions such as:


.

How do I know that I have a mixture experiment rather than the usual independent variable problem? If I have several (q ! 3) mixture components, how do I choose between using a simplex-lattice design or a simplex-centroid design? If there are restrictions such as lower and=or upper bounds on the components proportions, how do I proceed? In addition to mixture components, are there also process variables that might affect the response or affect the blending properties of the mixture components?

Answers to these and other questions are readily available from Snee (1971, 1973) and Marquart and Snee (1974). In each of the two papers, Snee (1971, 1973) presented a lucid discussion of several tted model forms as well as ways of analyzing mixture data. These papers were a must-read for the new-tomixtures reader because they appeared in print prior to the review article by Cornell (1973). The paper by Marquardt and Snee (1974) titled Test Statistics for Mixture Models was the winner of The Jack Youden Prize for the most outstanding expository paper to appear in Technometrics in 1974. This paper discusses the correct test statistics for testing hypotheses about the parameters (coefcients) in the Scheffe type mixture models. As an example, suppose for q 3 we wish to t the second-degree model of Eq. [6], y b1 x1 b2 x2 b3 x3 b12 x1 x2 b13 x1 x3 b23 x2 x3 e and test the null hypothesis, H0: The response does not depend on the mixture components. or, H0 : b1 b2 b3 b0 and b12 b13 b23 0 18 against the alternative hypothesis,

FIGURE 12 Estimated response contours plotted over the largest spherical region. Two acids x1 and x2 are blended with two glycols x3 and x4 as categorized components.
323

HA: The response does depend on the mixture components.


A Retrospective View of Mixture Experiments

shown in Figure 14a or there is curvature in the surface above the triangle as shown in 14b.

ADDITIONAL TOOLS USED FOR SOLVING MORE DIFFICULT PROBLEMS, 19791989


Snee (1979a) suggested steps for generating the coordinates of the extreme vertices of a highly constrained region dened by the placing of constraints on the component proportions of the form in [4]. Cornell and Khuri (1979) provided a transformation for obtaining contours of constant prediction variance on concentric triangles for ternary mixture systems as shown in Figure 15. Cox (1971) was the rst to suggest measuring component effects along rays inside the simplex by tting the standard polynomial model subject to constraints placed on the coefcient estimates. Piepel (1982) suggested a different direction by utilizing the L-pseudocomponent simplex in dening his direction to measure both partial and total effects of the components; see Figures 16a and 16b. Snee and Rayner (1982) assessed the accuracy of mixture model regression calculations, and Gorman and Cornell (1982) proposed a technique for reducing the form of the combined model containing both mixture components and process variables. Cornell and Gorman (1984) introduced fractional design plans for including process variables in mixture experiments as shown in Figures 17a and 17b.

FIGURE 13 Surface dened by the hypothesis H0: b1 b2 b3 b0 and b12 b13 b23 0.

or, HA : One or more of the in 18 is 6


Downloaded by [190.80.134.42] at 17:25 14 October 2013

19

Then the test statistic is Fcal SSRegression =p 1 SSError =N p 20

where p 6 is the number of terms in the model and N the total number of observations. If in [20] the value of FCal is <F(p-1,N-p,a) from the F-tables, then do not reject H0 in [18], in which case the response surface above the triangle looks approximately like that shown in Figure 13. On the other hand, if the value of FCal in [20] is greater than or equal to the value of F(5,N-6,a), then the shape of the surface above the triangle is not a level plane as shown in Figure 13 but is a plane as

FIGURE 14 (a) Planar surface above the three-component triangle. (b) Curvature of the surface above the triangle along the edge x1x2.
J. A. Cornell 324

FIGURE 15 Obtaining constant prediction variance on concentric triangles for ternary mixture systems. Steps in the transformation of circular variance contours to triangular-shaped contours. Four stages of the transformation are shown in Figures 8.88.11 of Chapter 8 in Cornell (2002, p. 469475).

Downloaded by [190.80.134.42] at 17:25 14 October 2013

Gorman and Cornell (1985) t equations to freezing point data exhibiting eutectics for binary and ternary mixture systems; see Figures 18a18c. Crosier (1986) outlined the geometry of constrained mixture experiments by providing a formula for calculating the number of d-dimensional

boundaries (d 0, 1, 2, . . ., q 2) of a region dened by the set of consistent constraints, x1 x2 . . . xq 1; is 0 Li xi Ui 1;

FIGURE 16 Comparing the directions taken by Cox and by Piepel for measuring the effect of component 1: (a) Coxs (1971) direction and (b) Piepels (1982) direction.

FIGURE 17 Same half fraction of cooking conditions versus mixed half fraction of cooking conditions at the seven blends of a simplex-centroid design consisting of three types of sh. See Cornell (2002, Fig. 7.9, p. 394).
325 A Retrospective View of Mixture Experiments

FIGURE 18 Three-dimensional plots of the freezing temperature surface for the biphenyl-bibenzyl-naphthalene system. Part (c) is part
(a) tilted to show the ternary eutectic T max(TI, TII, TIII) at the bottom of three intersecting surfaces TI, TII and TIII.

Nd C q ; q d 1
Downloaded by [190.80.134.42] at 17:25 14 October 2013
q d 1 X r 1 q X r d 1

Lr C q r ; q r d 1 Lr E r C r ; r d 1

21

where for r 1, 2, . . ., q and d 1, 2, . . ., q 2, RL 1


q X i1

Li

and Rp MinRL ; RU ;

where L(r) is the number of combinations of component ranges that sum to a number that is lower than Rp; E(r) is the number of combinations of component ranges that sum to Rp; and G(r) is the number of combinations of component ranges that sum to a number that is higher than Rp. For d 0 (the vertices), Eq. [21] simplies to N0 q
q X r 1

Lr q 2r E r r 1

22

where C(q,r) q!=[r!(q r)!]. Also, C(q, r) L(r) E(r) G(r); see Cornell (2002, pp. 156157).

Piepel and Cornell (1985) proposed models for mixture experiments when the response also depends on the total amount and, two years later, Piepel and Cornell (1987) suggested designs for mixture amount experiments; see Figure 19. In believing that many mixture surfaces, when viewed over the entire three-component triangle, look and are actually more complicated in shape than can be modeled with a rst- or second-degree -type model, Cornell (1986) compared two Scheffe 10-point designs for studying three-component mixture systems that are capable of supporting a cubic model or allow tting a special quartic model; see Figures 20a and 20b. Sahrmann et al. (1987) searched for the optimum Harvey Wallbanger recipe via mixture experiment techniques. This tongue-in-cheek real-life experiment was performed with 3 M design class employees by showing them how to score avor and how to use a balanced incomplete block design to generate the -type quadratic data followed by the tting of a Scheffe model to approximate the shape of the avor surface above the constrained region in Figure 21a. Figure 21b is a copy of a typical avor rating score sheet.

FIGURE 19 DN-optimal designs for two levels of amount.


Designs 1 and 2 are equivalent for a xed number of points.

FIGURE 20 Three-component 10-point designs: (a) a f3, 3g simplex-lattice; (b) a simplex-centroid design augmented with three interior points.
326

J. A. Cornell

FIGURE 21 (a) The seven Harvey Wallbanger blends and their locations in the triangle relative to the ratio variables constraint region.
(b) Ofcial Judge Harvey Wallbanger avor evaluation scoring sheet.

Downloaded by [190.80.134.42] at 17:25 14 October 2013

Cornell (1988) analyzed data from a real-life mixture experiment containing process variables using a split-plot approach.

THE LAST TWO DECADES, 19902009, FOCUSING ON APPLIED RESEARCH


In the ve sections presented thus far, we introduced the mixture problem and focused mainly on -type the development of designs along with Scheffe models. Time and space limitations have not allowed us to cover and discuss the many different model forms that have been suggested during the past 55 years. For the readers of this article who are interested in learning which empirical model forms have proved or proven to work best, we recommend Cornell (2002, chapters 68). Additional topics of interest that have appeared during the last two decades are listed below with full reference information provided in the References section. 1. Smith and Cornell (1993): Biplot Displays for Looking at Multiple Response Data in Mixture Experiments as shown in Figure 22. 2. Vining et al. (1993): A Graphical Approach for Evaluating Mixture Designs. 3. Montgomery and Voth (1994): Multi-Collinearity and Leverage in Mixture Experiments. 4. Cornell (1995): Fitting Models to Data from Mixture Experiments Containing Other Factors. 5. Heinsman and Montgomery (1995): Optimization of a Household Product Formulation Using a Mixture Experiment.
327

FIGURE 22 Biplot illustrating the effects of glass bers, resin, and microspheres on two strength responses, two modulus responses, and warp of a plastic compound. The projection arrow from the resin component onto the wrap vector illustrates the positive effect of resin on the wrap.

6. Bowles and Montgomery (1997): How to Formulate the Ultimate Margarita: A Tutorial on Experiments with Mixtures. 7. Cornell and Ramsey (1998): A Generalized Mixture Model for Categorized-Components Problems with an Application to a PhotoresistCoating Experiment. 8. Khuri et al. (1999): Using Quantile Plots of the Prediction Variance for Comparing Designs for a Constrained Mixture Region: An Application Involving a Fertilizer Experiment. 9. Anderson and Whitcomb (2002): Designing Experiments That Combine Mixture Components with Process Factors. 10. Draper and Pukelsheim (2002): Generalized Ridge Analysis under Linear Restrictions with
A Retrospective View of Mixture Experiments

11.

12.

13.

14. 15. 16.


Downloaded by [190.80.134.42] at 17:25 14 October 2013

Particular Applications to Mixture Experiment Problems. Kowalski et al. (2002): Split-Plot Designs and Estimation Methods for Mixture Experiments with Process Variables. Prescott et al. (2002): Mixture Experiments: Ill-Conditioning And Quadratic Model Specication. Cornell and Gorman (2003): Two New Mixture Models: Living with Collinearity but Removing Its Inuence. Goldfarb et al. (2003): Mixture-Process Variable Experiments with Noise Variables. Goldfarb et al. (2004): Evaluating MixtureProcess Designs with Control and Noise Variables. Goos and Donev (2006): The D-Optimal Design of Blocked Experiments with Mixture Components.

and to create excellent and sophisticated plots. Very little has been done to explore surface shapes using logistic regression and loglinear models. Why is this? 4. Has todays software been updated to include the modied L-pseudocomponents or the centered and scaled intercept model of Cornell and Gorman (2003)?

REMEMBERING DR. JOHN W. GORMAN; A FRIEND TO EVERYONE WHO WORKED WITH MIXTURES
Dr. John W. Gorman was born in 1925 near Sioux Falls, South Dakota. As a youth he joined the Friendly Indians, a forerunner of the Boy Scouts. Growing up on a farm, John felt that farm life had a way of building up strength, energy, and an interest in life and learning. At the age of 17, John joined the Army Specialized Training Program (ASTP), which offered enlistees an opportunity to attend a college of their choice once they completed basic training. Later the ASTP was called the GI Bill. By April of 1944, John, then 18, was stationed in England, where he volunteered for the U.S. Army Rangers. He landed in Normandy less than a week after D-Day and for the next 18 months served as an infantry scout in the 2nd Ranger Battalion, later to be recognized as one of the most decorated units of the war. John served with distinction during WWII, earning the Combat Infantry Badge, Purple Heart, and Bronze Star Medal while with the 2nd Ranger Battalion in the European Theater. Following the war, John enrolled at the University of Minnesota and earned a B.S. in chemistry and a Ph.D. in chemical engineering. After graduation, he became a process engineer at the Knolls Atomic Power Laboratory in Schenectady, New York. A major portion of his professional life, however, was spent with Amoco Oil Company as a research associate specializing in product and process R&D. Like Jack Youden, John taught himself statistical design and analysis of experiments. John was especially fond of mixture experiments. My initial correspondence with John was in 1972, when he served as an associate editor for Technometrics. An eventual paper, titled Experiments with
328

The following textbooks and a booklet from the American Society for Quality have provided information on experiments with mixtures since 1981: Cornell (1981, 1983, 1990a, 1990b, 2002, 2011) and Smith (2005). Currently a primer is in press with a publication date scheduled for summer 2011.

TOPICS FOR FUTURE RESEARCH


We come to the point in time now where it seems only tting to ask, Have most of the questions surrounding designs and models for setting up, for carrying out, and used to analyze data from experiments with mixtures raised during the past half century been answered? Or are there still questions that are unanswered? The retrospective journey taken in this article was mainly Whats been done? Certain unanswered questions still remain that create a void in my mixtures toolbox, and some of them are listed here: 1. Are there ways to measure the effects of the components other than using response trace plots when the experimental region is a constrained non-simplex-shaped region inside the original simplex? See Figure 23. 2. Are there model-free techniques for fractionating designs particularly with mixture experiments containing other variables? See Figure 24. 3. Todays software has allowed standard forms of polynomial models to be tted to mixture data
J. A. Cornell

Downloaded by [190.80.134.42] at 17:25 14 October 2013

Mixtures: A Review, was handled by John. I rst met John in person at the 1975 Gordon Research Conference and, as luck would have it, we were assigned as roommates. It was during that week that I was able to learn a years worth of mixtures mainly through questions I asked John as well as from the answers he gave me. He had a keen sense and feel for mixture designs and was always willing to share his wealth of knowledge on the tting and testing of mixture models. Including process variables in mixture experiments was an area that he and I shared a common interest in, and Cuthbert Daniel encouraged us to work together, which we did. As I look back at the many highlights that John and I experienced during a 40-year period of doing research on mixtures, some of my fondest memories were working with John while putting together Cornell and Gorman (1978, 1984, 2003) and Gorman and Cornell (1982, 1985). Johns many contributions to the American Society for Quality Control (ASQC), now (ASQ), and the American Statistical Association (ASA) earned him the honorary rank of Fellow of ASQ and ASA. He was a member of the WWII Rangers Association, a member of the Unity Mens Club, the Skylight Club, and the American Legion. Those who had the good fortune of knowing John have undoubtedly missed his compassionate love of family, gardening, and friends; his thirst for solving research problems; and, most of all, his warm and characteristic smile.

SUMMARY
Fifty-eight years have passed since the rst mention of a mixture experiment appeared in the statistics literature. Very few authors list Quenouilles book, which rst appeared in 1953, in their references, however. I believe the reason is because they either (1961), are not aware of this book or, like Scheffe they believe that Quenouilles approach to modeling s and ingredient blending is different from Scheffe s. And why not? Piepel (2006) they prefer Scheffe compiled a bibliography of mixture experiment publications that numbers over 700 entries, with roughly half of them appearing in non-statistics literature. Piepels (2006) chapter lists 360 references and, yes, he does include Quenouilles book.
329

This retrospective view of the past 57 years has been slightly more extensive than I had planned on reporting when I initially agreed to take this project on. Beginning with the section titled The Mixture Problem we began with a few examples of everyday activities such as sweetening coffee or tea, mixing oil and vinegar to create an Italian dressing for a tossed salad, and blending higher 93-octane fuel with 87-octane fuel presently in the family car to improve miles per gallon or driving performance. These activities fall into the class of mixture experiments. The next ten sections cover The Beginning Years1955 to 1965; The Second Wave of Problems to Be Solved, 19661978: Enter the Computer; Additional Tools Used for Solving More Difcult Problems, 19791989; The Last Two Decades, 19902006; and Topics for Future Research, respectively. In the next to last section, a tribute is paid to Dr. John W. Gorman, a friend to everyone who worked with mixtures. John, age 83, passed away on June 4, 2009, at his home in Plymouth, Minnesota. He was a quiet gentleman with a warm and charismatic smile who loved and inspired others to work on mixture problems. He will be missed. Finally, it is time to thank those individuals who helped me put together this piece of history. I want to thank Professors Geoff Vining and Douglas C. Montgomery for serving as discussants at the Friday morning Invited Session on October 8, 2010, sponsored by the CPI Division at 54th Fall Technical Conference, held in Birmingham, Alabama. It is a nice feeling to know that the topic of mixture experiments is in the good hands of these two pros as well as others currently working with mixtures. One of those is the Editor of Quality Engineering, Dr. Connie Borror, whose name appears with others in the list of references. Another special thanks goes out to Ms. Elizabeth Leis, Production Editor, who graciously pushed me to nish this article by the publication month, and last but certainly not least is Dr. Busaba Laungrungrong, who checked my spelling and helped me put this article in proper order.

ABOUT THE AUTHOR


John A. Cornell is a Emeritus Professor of Statistics who formerly spent 36 years with the Agricultural Experiment Station at the University of Florida, Gainesville. The author of more than 150 technical
A Retrospective View of Mixture Experiments

articles, and three books titled Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data, A Primer on Experiments with Mixtures, and Response Surfaces: Designs and Analyses (coauthored with A. I. Khuri), he is a past Editor of the Journal of Quality Technology. A past recipient of the W. J. Youden Prize, the Shewell Prize, the Brumbaugh Award and The Shewhart Medal from the American Society for Quality (ASQ), he is a Fellow of the American Statistical Association and the ASQ, and a past elected member of the International Statistical Institure. Dr. Cornell received the B.S.E. (1962) and M.Stat. (1965) degrees from the University of Florida, Gainesville, FL, and the Ph.D. (1968) in statistics from Virginia Polytechnic Institute and State University, Blacksburg, Va.
Downloaded by [190.80.134.42] at 17:25 14 October 2013

REFERENCES
Bowles, M. I., Montgomery, D. C. (1997). How to formulate the ultimate margarita: A tutorial on experiments with mixtures. Quality Engineering, 10:239253. Box, G. E. P., Draper, N. R. (1959). A basis for the selection of a response surface design. Journal of the American Statistical Association, 54:622654. Claringbold, P. J. (1953). Use of the simplex design in the study of the joint action of related hormones. Biometrics, 11:174185. Cornell, J. A. (1981). Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data, 1st ed. New York: John Wiley & Sons. Cornell, J. A. (1986). A comparison between two ten-point designs for studying three-component mixture systems. Journal of Quality Technology, 18:115. Cornell, J. A. (1990a). Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data, 2nd ed. New York: John Wiley & Sons. Cornell, J. A. (1990b). How to Run Mixture Experiments for Product Quality, 2nd ed. Milwaukee, WI: American Society for Quality Control. Cornell, J. A. (1995). Fitting models to data from mixture experiments containing other factors. Journal of Quality Technology, 27:1333. Cornell, J. A. (2002). Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data, 3rd ed. New York: John Wiley & Sons. Cornell, J. A. (2010). A retrospective look at experiments with mixtures. Paper presented at the 54th Annual Fall Technical Conference Quality & Statistics The Engines of Success, Birmingham, AL, October 8, 2010. Cornell, J. A. (2011). A Primer on Experiments with Mixtures. New York: John Wiley & Sons. Cornell, J. A., Good, I. J. (1970). The mixture problem for categorized components. Journal of the American Statistical Association, 65:339355. Cornell, J. A., Gorman, J. W. (1978). On the detection of an additive blending in multicomponent mixtures. Biometrics, 34:251263. Cornell, J. A., Gorman, J. W. (1984). Fractional design plans for process variables in mixture experiments. Journal of Quality Technology, 16:2038.

Cornell, J. A., Gorman, J. W. (2003). Two new mixture models: Living with collinearity but removing its inuence. Journal of Quality Technology, 35:7888. Cornell, J. A., Khuri, A. I. (1979). Obtaining constant prediction variance on concentric triangles for ternary mixture systems. Technometrics, 21:147157. Cornell, J. A., Ramsey, P. J. (1998). A generalized mixture model for categorized-components problems with an application to a photoresist-coating experiment. Technometrics, 40:4861. Cox, D. R. (1971). A note on polynomial response functions for mixtures. Biometrika, 58:155159. Crosier, R. B. (1986). The geometry of constrained mixture experiments. Technometrics, 28:95102. Draper, N. R., Lawrence, W. E. (1965). Mixture designs for three factors. Journal of the Royal Statistical Society B, 27:450465. Draper, N. R., Pukelsheim, F. (2002). Generalized ridge analysis under linear restrictions with particular applications to mixture experiment problems. Technometrics, 44:250259. Goldfarb, H. B., Borror, C. M., Montgomery, D. C. (2003). Mixtureprocess variable experiments with noise variables. Journal of Quality Technology, 36:393405. Goldfarb, H. B., Borror, C. M., Montgomery, D. C., Anderson-Cook, C. M. (2004). Evaluating mixture-process designs with control and noise variables. Journal of Quality Technology, 36:245262. Goos, P., Donev, A. N. (2006). The D-optimal design of blocked experiments with mixture components. Journal of Quality Technology, 38:319332. Gorman, J. W. (1966). Discussion of Extreme Vertices Designs of Mixture Experiments by R. A. McLean and V. L. Anderson. Technometrics, 8:455456. Gorman, J. W., Cornell, J. A. (1982). A note on model reduction for experiments with both mixture components and process variables. Technometrics, 24:243247. Gorman, J. W., Cornell, J. A. (1985). A note on tting equations to freezing point data exhibiting eutectics for binary and ternary mixture systems. Technometrics, 27:229239. Gorman, J. W., Hinman, J. E. (1962). Simplex-lattice designs for multicomponent systems. Technometrics, 4:463487. Heinsman, J. A., Montgomery, D. C. (1995). Optimization of a household product formulation using a mixture experiment. Quality Engineering, 7:583599. . http://www-history.mcs.st-andrews.ac.uk/pictDisplay/ Henry Scheffe Scheffe.html John, P. W. M., Gorman, J. W. (1960). Experiments with mixtures. Paper presented at the Gordon Research Conference on Statistics in Chemistry and Chemical Engineering, New Hampton, NH. Khuri, A. I., Harrison, J. M., Cornell, J. A. (1999). Using quantile plots of the prediction variance for comparing designs for a constrained mixture region: An application involving a fertilizer experiment. Applied Statistics C, 49:521532. Kowalski, S. M., Cornell, J. A., Vining, G. G. (2002). Split-plot designs and estimation methods for mixture experiments with process variables. Technometrics, 44:7279. Kurotori, I. S. (1966). Experiments with mixtures of components having lower bounds. Industrial Quality Control, 22:592596. Lambrakis, D. P. (1968). Experiments with mixtures: A generalization of the simplex-lattice design. Journal of the Royal Statistical Society B, 30:123136. Lambrakis, D. P. (1969). Experiments with mixtures: Estimated regression function of the multiple-lattice design. Journal of the Royal Statistical Society B, 31:276284. Marquardt, D. W., Snee, R. D. (1974). Test statistics for mixture models. Technometrics, 16:533537. McLean, R. A., Anderson, V. L. (1966). Extreme vertices design of mixture experiments. Technometrics, 8:447454. Montgomery, D. C., Voth, S. M. (1994). Multi-collinearity and leverage in mixture experiments. Journal of Quality Technology, 26:96108.

J. A. Cornell

330

Piepel, G. F. (1982). Measuring component effects in constrained mixture experiments. Technometrics, 24:2939. Piepel, G. F. (2006). 50 Years of mixture experiment research: 19552004. In: Khuri, A. I., Ed. Response Surface Methodology and Related Topics. Singapore: World Scientic Press, pp. 281 327. Piepel, G. F., Cornell, J. A. (1985). Models for mixture experiments when the response depends on the total amount. Technometrics, 27:219227. Piepel, G. F., Cornell, J. A. (1987). Designs for mixture-amount experiments. Journal of Quality Technology, 19:1128. Prescott, P., Dean, A. M., Draper, N. R., Lewis, S. M. (2002). Mixture experiments: Ill-conditioning and quadratic model specication. Technometrics, 44:260268. Sahrmann, H. F., Piepel, G. F., Cornell, J. A. (1987). In search of the optimum Harvey Wallbanger recipe via mixture experiment techniques. The American Statistician, 41:190194. , H. (1958). Experiments with mixtures. Journal of the Royal Scheffe Statistical Society B, 20:344360.

, H. (1963). The simplex-centroid design for experiments with Scheffe mixtures. Journal of the Royal Statistical Society B, 25:235263. Smith, W. F. (2005). Experimental Design for Formulation ASA-SIAM Series on Statistics and Applied Probability. Philadelphia, PA: SIAM. Smith, W. F., Cornell, J. A. (1993). Biplot displays for looking at multiple response data in mixture experiments. Technometrics, 35:337350. Snee, R. D. (1971). Design and analysis of mixture data. Journal of Quality Technology, 3:159169. Snee, R. D. (1973). Techniques for the analysis of mixture data. Technometrics, 15:517528. Snee, R. D. (1979). Experimental designs for mixture systems with multi-component constraints. Communications in Statistics, Theory and Methods, A8:303326. Snee, R. D., Rayner, A. A. (1982). Assessing the accuracy of mixture model regression calculations. Journal of Quality Technology, 14:6779. Thompson, W. O., Myers, R. H. (1968). Response surface designs for experiments mixtures. Technometrics, 10:739756. Vining, G. G., Cornell, J. A., Myers, R. H. (1993). A graphical approach for evaluating mixture designs. Applied Statistics, 42:127138.

Downloaded by [190.80.134.42] at 17:25 14 October 2013

331

A Retrospective View of Mixture Experiments

Das könnte Ihnen auch gefallen