One at a time plans for 2p factor sequencing designs by James Leonard Hansen A thesis submitted to the Graduate Faculty in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY in Mathematics Montana State University © Copyright by James Leonard Hansen (1974) Abstract: This thesis examines 2^p factorial experiments where the order of the application of the factors may be significant. In the experiments considered, the low level of a factor is the absence of a factor, while the high level of the factor is the presence of a factor. The effects of the factors are permanent and each unit may be tested at least p+1 times without the test affecting the unit. The assumptions which relate order effects are examined and a system with algebraic properties is proposed to assist the experimenter in estimating and interpreting order effects. A design and analysis are presented which allow for the estimation of order effects in addition to the usual main effects and interactions. The system for order effects is used to construct one at a time plans which allow for the estimation of order effects and factorial effects in experimental situations where the experimenter can get quick results with random error small in comparison to the effects which are to be estimated. ONE AT A-TIME PLANS FOR 2P FACTOR SEQUENCING DESIGNS "by - ' James Leonard Hansen A thesis submitted to the Graduate Faculty in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY in Mathematics Approved:' Heu._, - MONTANA STATE UNIVERSITY Bozeman5 Montana June5 1974 ill ACKNOWLEDGEMENT The author wishes to express his gratitude to the chairman of his graduate committee5 Dr. Kenneth J . Tiahrt5 for suggesting this thesis problem, and for his guidance in the preparation of this thesis. The author is also grateful to Professors Martin A. Hamilton, Richard E. Lund, Byron L. McAllister, -Franklin S . McFeely and William R. Taylor for serving on his graduate committee. iv TABLE OF CONTENTS CHAPTER I. II. III. IV. V. VI. PAGE- INTRODUCTION ..................................... . I A SYSTEM FOR EXAMINING ORDER EFFECTS 5 Preliminary Considerations ...... ....... Assumptions Regarding Binary Operations In Order Effects ...... ............................. 5 12 ESTIMATION AND INTERPRETATION OF ORDER EFFECTS . . 16 Design and Model ....................... . Analysis ................. .......... ....... . Examples ................. ............ ......... 37 Fractional Replications of 2P FSD Designs ..... l6 28 ONE AT A. TIME PLANS FOR THE 23 F S D ............. 46 Preliminary Considerations ...... ........;.... One at a Time Plans for Case I .. . . ........... ■ One at a Time Plans for Case 2 ...... '......... Examples ................... ....... . .........i. 46 50 6l 44 65 ONE AT A TIME PLANS FOR 2P FSD WITH p > 3 ..... Tl I n t r o d u c t i o n ....... ...................... . An Example of a One at a Time Plan for a 2^ FSD Example .......... .......... ................... .. Tl 72 78 SUMMARY AND EXTENSIONS ..... ....... . 86 BIBLIOGRAPHY APPENDIX ___ ....... . ........ .......... .......... 88 ,89 V ABSTRACT This thesis examines 2^ factorial experiments where the order of the application of the factors may be signifi­ cant. In the experiments considered, the low level of a ■ factor is the absence of a factor, while the high level of ■ the factor is the presence of a factor. The effects of the factors are permanent and each unit may be. tested at least p+1 times without the test affecting the unit. The assump­ tions which relate order effects are examined and a system with algebraic properties is proposed to assist the experi­ menter in estimating and interpreting.order effects. A design and analysis are presented which allow for the esti­ mation of order effects in addition to the usual main ef­ fects and interactions. The system for order effects is used to construct one at a time plans which allow for the estimation of order effects and factorial effects in experimental situations where the experimenter can get quick results with random error small in comparison to the effects which are to be estimated. CHAPTER I INTRODUCTION Factorial experiments are useful when the researcher is investigating the effects of each of a number of factors on the response of some variable. Usually all factors are • applied simultaneously to the experimental units and the response is recorded. This is especially true in agri­ cultural experiments where different levels of the treat­ ments may be applied, simultaneously and the response is observed and recorded. This type of experimentation was developed.by'R. A. Fisher [4] in the 1920's and early 1930's. Factorial experimentation is very efficient because every observation supplies information about each factor included in the experiment. In many industrial experiments where the factors are environments, the factors can not be applied simultaneously, but may be applied in any order. For example, in testing electrical switches or relays, one factor may be vibration and another mechanical shock. In this instance, the factors must be applied sequentially and the order of the application of the factors may be important. R. R . -Prarie and W. J . Zimmer treated, this problem in two papers published in 1964 and 1 9 6 8 in the Journal of the American Statistical Society. 2 The type of sequential experimental design referred to in the Praire and Zimmer papers is related to the order of the application of the factors. This is different from the definition of sequential experimental designs, where observations are obtained in sequence in time and it is to be decided at each point in time whether the experiment is to be continued and possibly what treatment combination is to be used. Therefore, Prairie and Zimmer .termed their, designs Factor Sequencing Designs (FSD) to distinguish them from Factorial Designs or from Sequential Designs. In [9] 5 Prairie and Zimmer considered 2^ experiments which apply to the situations where: a) Each unit may be tested p+1 times without the test itself having an effect on the same unit. b) There can be no trend effect with successive tests on the same unit. c) The high level of a factor is the presence of a factor and- the low level is the absence of the factor. d) The effects of the factors are permanent. The experimental designs discussed in this thesis will always be assumed to satisfy this same set of assumptions. 3 The usual 2P factorial designs require -2^'units with one test per unit to estimate all factorial effects. The Factor Sequencing Designs (FSD) developed in [9] require pI units and p+1 tests per unit. In [9], the design and the analysis for full FSD are presented, and in [10], fractions .of -the full FSD experiment are presented. The purpose, of the FSD experiment as presented by Prairie and Zimmer is to determine the importance of the order of application of factors. If order is not important, the factorial effects are estimated in the usual way. But if order is important, they state that .the interpretation of the factorial- effects is. difficult. If the factors cannot be applied simultaneously, they must be applied in sequence one at a time. Many scientists do their experimental work in single steps, and strive to learn something from each trial or^ run. These scientists can react quickly to the results of individual runs ; however, in order to achieve good results, the experimenter should have effects which are at least three or four times as large as his average random error per trial. In [3], C. Daniel proposed, one at a' time plans to produce data of greater value to the experimenter than the sequences of one at a time trials previously used. He 4 indicated that this type of experimentation is economical, but may give biased estimates. In his paper he described many of these biases as two factor interactions, and then provided sequence's of one at a time runs to separate main effects from these two factor interactions and gave methods of estimating each two factor interaction separately. Factor Sequencing Designs were designed to estimate the effect of the ,order of application of treatments. But each 2p FSD requires pi units and (p+1) tests per uni t5 there­ fore, one at a time experimentation can be used to determine as quickly as possible if the order of application of the factors makes a difference. The purpose of this thesis is to present a model which will yield the same tests for orders as those presented in [9 ] and. [1 0 ] and in addition, will allow for the estimation of the order effects. When one at a time plans are.used, the proposed model will also enable the experimenter to interpret his results if order effects are present. If the sequence of one at a time runs is incomplete, the model allows for estimation of the re­ sulting biases in the order parameters. The one at a time plans are constructed to yield maximum information about order effects as early as possible in the experiment. CHAPTER II • A.SYSTEM FOR EXAMINING ORDER EFFECTS Preliminary Considerations To facilitate the study of order effects, a model analogous to the usual mixed model will be used with random unit effects and fixed treatment effects. The model used to represent a response for the 2p FSD is Yij(fV * m + ui + rIf1 , ... , f ^ ^ + 6 I. . .j-l + e ij j-l)' where m = the general mean. U^ = the effect of the ith unit. ' • -p = the effect of the factors 1 I " * * * 5 j- 1 Tif. f^,...fj_^, in any order, for j=l the symbol used is Tj « 0^ = the order effect of applying, factors f^,...,f ._^ in the specific order . f ...,f e^. = the random error associated with the jth test on the.ith unit. For example, in the response Y ij(fl"f2 ) m + u. + i + ^ f 1 3 F2 .+ 6 12 + G ij the term 0 ^ 2 indicates the order effect resulting from applying treatments F^ and F2 in that specific order; first 6 If1 and.then fg. Similarly O21 would indicate the order effect resulting from applying treatments in the order f2 and then f^ • The model will be discussed in more detail later. The present section is involved with developing a system to assist is the discussion of order effects. Throughout the discussion,, reference will be made to properties of a binary operation. The properties used are the following: Commutative Property: A binary operation * will be commu­ tative if and only if a * b = b * a. Associative Property: A binary operation * will be assoc­ iative if and only if ( a * b) * c = a * (b * c). To motivate the discussion, consider the following responses hi(i) Y12(A) y IS^fV , m + u ^ + r j ^ + O 11 m + uI n + 1Qf I + 01 + e 12 I f2 ) - m + U1 + rIf^f0 + 612 + 613 1*2 ' f + 6 1 2 3 + 6 14 'l*-2*-L3 The response Y 1 1 (I) is a preliminary test before any Y l4^fl ,f2^f3^ m + U1 + factor has been applied. The response Y 1 2 (T1) is the response after one factor has been applied, and the symbol G 1 is not subject to an order interpretation. In the 7 response Y 13 ( f ) , e 12 could be.denoted by S 1 * e2 where the star would define an operation on the order effects. Hence, and indicates the order effect of applying treatment f^ first and then treatment fg second. Gl * Gg * G 3 Similarly, ^12 * 83 0 -123.' ■ The operation * is well-defined,- but is not commutative, If * were commutative, 0 l2 = * 0g = Gg * G^ G 215 which would imply no two factor order effects were present. Thus .the assumption of commutivity would eliminate precisely the property of the model which the FSD is used to determine. Associativity would imply 6 123 = 9 1 2 * 03 . = Gi * Gg * G3 = 9 I * 9 23* Although this concept does not contradict the basic design assumptions as commutivity does, the expression 8 0 I * e23' 18 difficult to. interpret. However, the assumption -' of associativity simplifies the discussion which follows and for this reason will be included in the definition of *. The following definition is based on the preceding discussion. Definition 2 . 1 : The associative binary operation * relating order effects is defined by 0I * 0 2 = 0 1 2 * The operation is extended to three or more effects by repeated operation on the right. For example, 0 12 * e'3 = 0 123 The following lemma is derived from the definition. Lemma 2 . 1 : Let P and Q, represent two permutations of k factors, then if 6 p = 6 q , then 0 pj = 0 ^. for any factor j not a member of the original k factors. Proof: If 0p = 0 q , then by Definition 2.1 "• - 6P * 8J = 6Q * 6J or 8Pj = 6Qj1 . 9 For example, if the two order effects and ^ 21 are equal, i.e. 6^23 = Sg2 1 , then If factor f^ is applied. Lemma 2.1 with P = 1 2 3 , Q .= 321, and j = 5 implies 6 1235 = 9 3215* The above Definition and Lemma will be used to develop a system which aides in the interpretation of results from FSD experiments where order effects are not negligible and assists in developing one at a time plans for these designs. To motivate the discussion,, consider the following example of a 2 U experiment. Example 2 . 1 : factors f f g The experiment consists of applying three and fg each at two. levels where the low level indicates no treatment is applied and the high level indicates, the treatment is applied. Assume there are six units available and the following experiments are performed. Unit I 2 3 4 Application Sequences •flf 2 • fIf 3 f2 f l f2f3 5 f3f l 6 f3f2 10 In the array given above, the notation f ^f2 represents an application sequence. The symbol T f 2 indicates the unit has been tested three times at this stage of the experiment: first, prior to any treatment ; second, after factor f^ was applied; and third, after the application of factor fg. A similar interpretation follows for application sequences of more than two factors. Then suppose the appropriate contrasts have been tested and the following equivalences are determined from the experimental data. 612 921 = ' ' 923 = ^32 , e31 = e13’ By applying Lemma 2.1, the-following equalities can be derived. 9123 = 0213 923I = 9321 9132 = ' 9312 ' The above relationships are intuitive because if it is assumed that experimental units are not affected by order after the .applications of two factors, .the' two units should be the same except for experimental error. ■Therefore, the 11 same response is expected from the two units after applying the same third treatment to each. H owev e r , t h e r e is no reason to assume a priori that 0 12 3 = e 2 3 1 or 9 1 2 3 = 0-i3 2 * is possible that the three sets could all be non-zero quantities, and in this instance the interpretation would be that three factor order effects occur, but there are no significant two factor order effects Prairie and Zimmer, in their 1 9 6 8 paper, presented some fractions, of the full PSD design. Some of these fractions involved testing only all possible two factor order com - ' binations by an. F test. The above example is intended to illustrate the possibility of a higher order effect, in this case a three factor order effect, even though the test for two factor order effects would not be significant. 12 Assrunptions Regarding Binary Operations In Order Effects Example 2.1 indicates a need to examine the assumptions underlying order effects. The following discussion will consider two situations and each one will he examined by relating the binary operation to the types of properties it satisfies. Case I: The first case considered is the situation where it is known that the set of all order effects of k or fewer factors is insignificant. This prior knowledge gives no information about the set of order effects when more than k factors have been applied. This is equivalent to assuming those properties on the binary operation given in Definition 2.1 and Lemma 2.1 for combining order effects. A situation where'this occurs would be as in Example 2.1, and in this instance there could possibly be three distinct three factor order effects for the 2 , the effects related to (I) O 12^.and 0 2 1 3 ? (2 ) e 2 3 1 and 6321 > and (3) 0 i32 and S3 I2 , even though two factor effects are absent. Lemma 2.1 gives partial commutativity for effects judged insignificant using experimental data. This means .symbols related to two factor order effects which, have been determined to be insignificant can be commuted only when they appear in the left-most positions in the sequence. 13 . •5 For example, consider a 2 U FSD and assume the appropriate two factor order effects have been tested and. based on these observations the experimenter assumes 6 12 O2 0 1 , O 10 ^ O 01 , and e, 0. 1 5 13 ^ w31’ 23 -.*32' Lemma-2.I indicates the following relationships concerning three factor inter­ actions, ■6123 = 0213 and 0231 = 0321' Based on this information the experimenter can plan the remaining experimentation to get the most information with the least expenditure of resources. Using the assumptions of Case I, the experimenter can use this system to develop relationships.among the order effects. Case I could be referred, to as the case where two factor order effects judged insignificant commute only on the left of the application sequence. Case 2: ..The next situation discussed is one where the symbols related to two factor order effects determined to be insignificant can be commuted anywhere they appear -in the complete order sequence. This is equivalent to assuming the binary operation is commutative for the symbols judged insignificant as well as associative for all symbols. The following definition is based on the preceding discussion. 14 Definition 2 . 2 : if 0 = 6 The binary operation * is commutative then 0 p-j_jQ = 0PjiQ for any two sequences- of factors P and Q. Applying the definition with i = I, j = 2, P = 4 and. Q = 3 5 , and if = 0 12 j then 6 4l235 = 6 42135* This Definition can be used to determine the sequence of factors to be run to yield maximum information. Por example, consider a 2U PSD design where it has been determined that = and e 2 3 = 0 3 2 5 that ^ 0 g^. Since G ^2 ~ @21* Lemma 2.1 implies 6-^23 = 0 213* addition, and- Definition 2.2 implies 0 ^ 2 3 = ®132 ' Hence, 0. '123 6, "213 0n 132* Similarly, Lemma 2.1 and Definition 2.2 imply e312 e321 e231* The two sets can be defined as .equivalence classes of order effects. The result is intuitive in that if there is only an order effect related to factors A and C, i.e. 0 13 ^ ®31J •then the position of B in the treatment sequence has no effect, only the relative positions of A and C. 15 If the experimenter can assume the conditions of Case 2 , he. can find more relationships among the order effects than he could in Case I. Case 2, could be referred to as the case where two factor order effects judged insignificant commute anywhere they appear together in the application sequence. Even if Definition 2.2 can not be assumed, a priori, the experiment can be run as in Case I where the results of Case 2 are possibilities to be tested for experimentally. If.there.is a set of equivalence classes of order effects, the sets can be helpful in interpreting and applying the results of the experiment when order effects are present. For instance>'if an FSD is being used to determine factor effects for some industrial process, the order effect could indicate a type of catalytic effect related to the order of application of factors, and the desired results might be achieved only through a specific order of application. When implementing a production process, the. knowledge that order effects have been separated into equivalence classes allows for.the selection of any particular factor applica­ tion order sequence from an equivalence class on the basis of optimization in terms of criterion such as production time or cost. CHAPTER III ESTIMATION AND INTERPRETATION OF ORDER EFFECTS Design and Model The model used is similar to the one discussed by Prairie and Zimmer [9]j and is the one stated in equation 2.1. The main difference is the addition of specific representations of order parameters and assumptions con­ cerning these parameters to allow for their estimation and an interpretation of these order effects. Each factor f\(i = l,...,p) has a high level (f^ has been applied) and a low level (f^ has not been applied). Because the effects of the factors are assumed to be perma­ nent, a unit which has received factor f^ at some point must be considered as having the high level of that factor from that time on. In addition, in a complete experiment, every unit will receive each of the p factors in some order and will be tested (p+1 ) times with the first test occurring before the application of any factor and each succeeding test occurring after the application of each of the p factors The design considered in [9] was one Where pi groups of r units each were subjected to. exactly one of the possible pi orders. A notation for the model was given in Chapter II ■ and is restated here. 17. (3.1) Y±j f ' ■j-l; - ;m + u i + 1If 1 I 5 *•'5lJ-I + + eIj I — l^eee^^.p* j — lj««°jP+ I where denotes the response from the jth test on the i.th unit resulting from the j-I factors applied in the application sequence f^...f defined as they were in Chapter II. The parameters are In addition, it is also assumed that (I) eIj ~ "ID( (2) U1 ~ NID( 0,a® ) (3) There are no interactions among the 9 1s or 0,a2 e ) between the 6 1s and the -q «s. (4) Let A k be the set of all permutations of k then (3.2) Z IeAk 9L = °9 L for 2 < k < p . Example 3 .I: If the experiment being run is a 2~> F S D 3 then the set of order effects is: {9 1 2 5®2 1 5®1 3 5®3 1 5®2 3 5^1 2 3 3^1 3 2 5®2 1 3 5®2 3 1 5^3 1 2 5^321 ^ 18 and. condition (4) would imply 612 + S21 _ ° 5 9 13 + 031 ■ 0 23 + Sg2 ~ Oj and 9 123 + 9 132 + 9213 + 9231 + 0 312 + 9321 Conditions (I) and (2) are usual assumptions for mixed models where the population of inference is infinite. Con­ dition (3 ) indicates an assumption of independence between the order effects and the factor effects. The order effects are a special type of interaction and to assert they interact with the factor effects, would be a redundancy. To assume they interact with each other would imply that orders inter­ act with orders which is also redundant because the test is for order effects and the 0 *s as defined are order parameters Condition (4) is imposed to provide a full rank model. How­ ever, the assumption does not seem unreasonable because if there is for example an order effect related to f^ and fg, the assumption S 12 + S121' = 0 would indicate one order pro­ vides an increase to the general mean while the other yields a decrease in the response from the general mean*- Similar interpretations hold for order effects for more than two factors, some will decrease the response level and some will , . 19 increase I t 5 but the deviations from the general mean will add to zero.' Using matrix notation, the model stated- in equation 3.1 may be written as (3-3) Y = X*(™») + gu + e where g is the r(p+l)pl x I column vector of observations, X* is the design matrix with r(p+l)p! rows and I + 2^ + P ^ pl/(p-i)l columns, p* is column vector of 1=:2 factor and order parameters with 2 ^ + P ^ p!/(p-i)I. rows, 1=2 u is the rp! x I column vector of unit parameters, ¥ is the r (p+l)pl x rpi matrix, whose cth column is a column of zeros except "for ones in the [(c-1 )(p+l)+l]th' row through the [c(p-hl) ]th row and e is. the r (p+l)pl x I column vector of random errors. To illustrate the model the responses from a 2V are 20 = m + U1 + Ypg(A) - m + H 2 Y 1 ^ f A , B) = m + U1 = m + m + U 2. + Y 2 4 (A,C,B) = m + + m + e 12 + i 1I 1^A, B + tiI U2 + 6 12 9 123 + (Tl H -Fr = ^A C + '9 1 3 2 ^ g 24 n A , B,,C + + OO I—I U) Ygf(I) + G 11 + Y^fA,!)^) ^l + Y 1 1 (I) 6 21 + + !^,2 , 0 + + =54 This model is different from the one proposed in [9] in that Prarie and Zimmer did not include explicit parameters for.order effects. However, the model presented in 3.2 has the same number of linearly independent observation vectors/ hence their result concerning the rank of.the design matrix holds for X*. The rank of X* is P (3.5) g PlZ(P-I)I i=0 For example, if unit parameters are ignored (the vector P* does not include unit parameters), then for one repli- . cation of a 2^ PSD, there are 6 units which have T^, two 21 units each for T]^ t)q 5 one unit for each of and its order parameter and one unit for l H 1 Jk with the appropriate order parameters, hence there are 3 Z 3 1 / (3-i)I = ! + 3 + 6 + 6 = 1=0 16 linearly independent column vectors in X* for a 2 . . The argument can he extended and therefore the rank of X* for arbitrary p is as given in Equation 3.5» ■ The rank of X*'X* is the same as the rank of X*, and the size of the X*'X* is equal to the number of columns in X*. The following discussion will show that a series of constraints' on the model parameters will lead to a new model of full rank. A reparameterization of the factor effects formed by subtracting H 1 from each factor parameter and deleting the column of zeros will, reduce the dimension of X* by one without changing the rank. Using condition 4, the order effects can. be -reparameterized to form a full rank model. The following, lemmas show that the reparameterization based on condition 4 is sufficient for a full rank model. Lemma -3.1; In a 2^ FSD experiment,' there are 22 P ■Yi 1=2 p l/(p-i) I order effects. Proof:' For a fixed I. the number of order effects f o r ' I factors is the permutation of p factors taken I. at a time, i.e. pl/(p-i)l. For a 2^ FSD order.effects are possible if i = 2,...,p. So there are P Yj 1=2 pi/(p-i) I possible order effects. This -completes the proof of the lemma. Lemma 3.1 implies there are 3 . Z 1=2 31/(3-i)I = 12 order effects for the 2U FSD. .The twelve effects were enumerated in Example 3.1. The following lemma will show how many constraints are imposed by condition (4). Lemma 3 . 2 : For a 2^ FSD experiment condition (4) implies constraints are imposed on the model defined by Equation 3.1. 23 Proof: •For a fixed, set.of I factors 2 < ± < p, there are possible combinations of factors each of which provides ohe constaint of the form Z eL = 0 . LeA1 L Consequently, the total number of constraints is given by This completes the proof of the lemma. Lemma 3.2 implies that four constraints are imposed on the 2~> F S D e .' The four constraints were given in Example 3.1. If the number of constraints is subtracted from the number of order effects, the number of parameters present after imposing the constraints is given by Lemma 3«I and Lemma 3.2. P I i=2 PV'(p-i) The next lemma indicates the number of order parameters to be estimated after imposing the constraints equals the number of degrees of freedom for order.effects in the 2^ FSD design, ■ 24 In a 2.^ FSD experiment there arq Lgmtna 3 » 3 : L 1=2 degrees of freedom for order effects. ' Proof: 'The number of degrees of freedom is equal to the number of linearly independent comparison's which can be formed between order effects of the same set L of i factors. Hence, each of the combinations can be permuted, in il ways which implies there are (il-l) independent comparisons which can be formed and for fixed i there are ^?J^ii-l degrees of freedom. Since 2 < i < p, the total degrees of freedom for order, effects in a 2p FSD design is P ?YiJ-lJ. This completes the proof of the lemma. i=2 Lemma 3.3 is illustrated using the o3 2 FSD experiment described in Example 3.1. of two factors. There are ('3'' = 3 combinations They are {12, 13, 23} can be permuted in 21 = 2 . Each combination ways, and. only one linearly independent comparison of these two order effects can be obtained. Thus there is (21-1) = I .independent effect for each combination. Similarly, there is (^j = I combinations of the three factors, - and there are 3 ! = 6 order effects. Among these six order effects only five linearly independent 25 comparisons can be made. • 3' Therefore there are . J 2 ( i X 11 -1 ) = 3 - 1 + 1-5 = independent order effects for the "■ 8 -Q FSD. The rank of X* can be decomposed as follows: P . E Z i=0 P Pi/(p-i) I = Z i=0 (f) ii P. Z ■ ^i i=0 + Z i=0 il-l Because the first expression on the right equals 2^ and the first two terms of the second expression are zero, . Z 1=0 PV(p-i) I = 2P + Z (± 1= 2 Xly P I + (2 P -1 ) + Z i=2 (? i )( 11 -I where the expressions on the right correspond to the degrees of freedom for the mean, factors and orders respectively. The above discussion implies that X* can be reparameterized to form a full rank model. The reparam­ eterization can be accomplished by multiplication on the. . 26 right by a matrix M. The design matrix X used will be one which contains the appropriate contrasts for factor effects„ The factorial representation and nomenclature for 2^ experiments is given in many tests, eg., .[8 ] .■ For a 2p factorial experiment, the treatment combination can be represented.- as an n-tuple, X^ = + I. (X^,Xg,...,X^) where each Thus the factorial representation for the result Q of a single run of in a 2^ may be written as jm + 1 / 2 [AXi + BX2 + CX3 + (AB)X1X 2 + (AC)X1X 3 + (BC)X2X 3 + (ABC)X1X 2X 3 ^ where m = the general mean X i = -I at the low level of A, B or C for i = I, 2 , 3 , respectively. = + 1 at the high level of the corresponding factor. The symbols (AB), et cetera are not products, they represent interactions0 The parenthesis will usually be omitted. This representation simplifies the discussion of one at a time plans in Chapters IV and V e A matrix M such that X*M = X exists by the results concerning generalized inverses of Chapter I of Searle [11]. 27 After .the reparameterization the model becomes (3.6) Y-Xf-) \P + Wu + e . / rv/ For an example of X from a 23 FSD with r = I, see the 'Appendix.' mV ■ pj Also5 ■ ' - (IHjA jB5C3AB,ACjBC,ABC,S12^e13,S23,S123^e132,S213. e231’e312) • Examples of responses are: ' Y n (I) = m + l/2{ -A - B - C +■ AB + AC + BC - ABC) .+ U i + G11 Y1 2 (A) = m + l/2{ A - B - C - AB - AC + BC + ABC) + uI + g12 Y13 (A5B) = m + l/2( A + B; - C + AB - AC - B C + 0 1 2 + Uf + e I3 ABC) ■ Y3 2 (B5A) = m + l/2{ A + B - C + AB - AC - BC - ABC) - O 12 + U 3 + G 32 The model used is analogous to the mixed model of ran­ dom unit effects and fixed treatment effects.- The analysis of the design will be explained in the next section. 28. ■ Analysis The analysis of the design given in the previous section is based on the analysis presented in [9]. The procedure used was to transform the model by a,linear transformation and then to analyze the transformed model using the method of least squares. The same technique is used in the section which follows. In the discussion which follows, it will be desirable to use a method of multiplication of two matrices which is different from the usual matrix multiplication. This method called the direct product is very useful when working with blocks of submatrices. The following definition is given by Graybill [5]. Definition —— ... .. 13.1: '' Direct Product: Let ~ P be a m0C X n0C matrix and let Q1 be an m n x nn matrix; then the direct ^ J_ J- product of P and Q written P ® Q is a matrix T of size ralm2 x n in2 - defined by T = [\j] = The symbol I will always denote an identity matrix and J will always denote a matrix with every element equal to one. The model defined, by equation 3.6. is of full rank, but the Gauss-Markov Theorem does.not apply because the Y 1s are 29 The non-independence is a result of the not independent. following theorem. Theorem 3 . 1 : For the model given by Equation 3,6, (!) e (S) Tar(Y) = c h + ^ [ J ® I] Proof.: (1) U ) = x(g) . Applying conditions I and 2 of expression 3.2, B(Y) = E(x(™) + W u + e) = B x(“) + E(Wu) + E(e) = 2(g) + Fte) - - zlg) 2 (2) Var(Y) = . E(Y - X 0 ) ■ = EfWu + e)2 = E(e2 ) + E(Wue) + E(eWu) + W' E(uu' )W' 2 2 = CT^I + a, WW' VW TJ1 VWVW 2 2 = »e I + = u e ® where the I associated with a 2 I) is the identity of order r(p+l)pi, and [J (g I] is the direct product of a J matrix of size (p+1) X Hence, (p+1) and I is the identity of order rpl. [J ® I] is a square matrix of order r(p+l)p! which is 30 p the same dimension as the identity associated with a The Gauss-Markov Theorem for least squares estimation can be applied to a transformed vector Z = TY if Var(Z) = a®I. T h e r e f o r e t h e transformation must satisfy ■' Var(Z) = Var(TY) = T Var(Y)T = T [ct2I + C t 2W }T' = CT2TT + tC 1 ^TW(TW) * - aL 1 Hence the matrix T must satisfy the following two conditions (1) TI" = I (2) (TW)(TW)1 = 0 or equivalently TW = 0. If r = I, a matrix which satisfies the above conditions is (3.7) T = [H® I]- where T is a rectangular matrix of dimension p p I x (p+l)p I and. H is a.p x P+1 matrix which is a Helmert matrix with the 31 first row deleted (see Searle [11], page 33). 1/V2 -IXzrP 0 1/V6 1X/T5 -2X/6 ..• 0 -i 0 H ■ I I I ; -P VxP(P+ I) y p (p+ i) V F T p + I) ; ... V p (P+ I) •' For r replications of the FSD experiment, the desired matrix is. ■ T= [T ® j 1 ]■ where T is defined as above and j ' is a r x I column vector of ones. 2 3 A n example of a matrix T for one replication of a ' is the Appendix. The following theorem follows from the definition of T. Theorem 3 . 2 : For Y defined by Equation 3.6, and T defined Equation 3.7. (I) If Z = TY, then Z is normal (£) >(!> = e (p ) (3) Var(Z) = Proof: The proof follows from Theorem 3.1 and a well known result from multivariate analysis (see Anderson Theorem 32 2.4.5) which states if X is N(jj.,y) and. if Z = DX then X is N ( D ^ D W ) ' ). Hence, Y normal implies Z = TY is normal and B(Z) = E(IT) = IB(Y) = IK(™) and Var(Z) = T Var(Y)T.' = fT[ o-2I +ct 2T aTW1]T' v .£ U.~ «"W 2 2 = c £ TT' + a"jj_,TW(TW)' fw zvrrv» 'rv<x»^ = V 1 6~ . where I is of order rp.pl x rp*pl. This completes the proof. The first row of TX is all zeros because the trans­ formation removes the mean effect as well as the unit effects. A full rank model for Z will result if the first column of TX is deleted, and the parameter m deleted from the parameter vector. Set R equal to the matrix TX with the first column deleted, then (3.8) where R is a Z = Rp + e 33 p rppi x 2 i=l Ip -TJT matrix of full column rank, p is a I ■ill X p W x I x 1 column vector of parameters, and e .is a rppi % I vector of random errors. Because the mean and unit effects have been removed, all that remains for degrees of freedom are those for treatments,.orders and error. O distributed with Var(Z) = By Theorem 3.2, Z is normally therefore, by the Gauss- Markov Theorem the best linear unbiased estimate of {3 is (3.9) P = (R1R)-1R 1Z If order is neglected, a new parameter vector is formed, composed only of factor effectss P — (A,B,C,AB, .. .) Corresponding to this parameter vector, a design matrix R^ of order rppi (2^-1) can be found by (3 .10) where is an augmented matrix with I the (2^-1) x identi' by and 0 is a null matrix of order (2^-1) 34 ■P 2 ■1=1 - (2p -l))x (Sp-I). (P -i) The vector of factor" effects P 1 is estimated "by (3-11) Il = (KlRi)-1JiZ For a FSD with r = I, the matrices R, (R1R ) R #<# /X/ #\# ( I and (R!^R^) ^ are given in the Appendix. A partition of the total sum of squares Z 1Z is Z'Z = Z 1 [I - R ( R 1R)-1R 1 ]Z '+ Z' [R(R1R)-1R' - R 1YR' R 1 )"1R' ]Z /n/ #X/ #X# rv ^ *V> rxz yx# y\yJ. 'yx>Jp^* J. «%#J_ + 5' [ K i ( K i K f 1Ri)S where all of the matrices in the brackets are idempotent. Using Equation 3.10, it can be shown that all cross products are zero. The first term of the partition is the error sum of squares associated with fitting the full model. The second term is the sum of squares associated with fitting' order effects only and the third term is the sum of squares for factor effects. . The degrees of freedom for factor effects are 2^-I5 and by Lemma 3.3, the degree of freedom for order effects are P 2 i=2 The total degrees of freedom are 'rppi, and by subtraction. those for error are a. 35 rpp! - (gP-l) The preceding discussion can he summarized in the following Analysis of Variance table. Table 3.1 Source (SV) Degrees of Freedom (DF) Total rpp! Z'Z r—I % Factors Z 1B 1 (EiS1 ) "1Bliz Orders Error * Sum. of Squares (SS) . Z t [R(R1R)-1R - R 1 (R'Rn )"1R* ]Z rpp! - (2^-1) - :J 2 ( a — Z ' [I - R ( R 1R)-1RjZ ) The expected mean squares corresponding to the sums of squares of Table 3•I are CTe + E f E S f e P / D F ' “ e + [6' (s'5)g - ei(Ei5i)6iJ/EI? ... for factors, orders and error, respectively. 36 Because the matrices in the quadratic forms for the sums of squares are idempotent with cross product zero, under the null hypothesis of no factor and. no. order effects P the sums of squares are independently distributed as a times a % 2 variable. Thus the F-test can be used to test the hypothesis that order effects are all neglible, also the factorial effects can be tested by the appropriate F-test. In the situation where the order effects are negligible,, the factorial effects would be estimated by If the A order effects are significant, p provides an estimate of the factorial and order effects e ■ In this situation, the discussion of Chapter Two applies and aides in the interpre­ tation of the. order effects. Two examples are presented to illustrate the analysis and to illustrate the use of this design in interpreting factor and order effects when order effects- are significant. 37 Examples In [9]j examples are provided to illustrate the model presented in that paper. The same data is used in the following two examples to show that the model and design presented in this thesis leads- to the same F tests for order effects and f o r .factorial effects. Example 3 . 2 : The following set of observations is ob­ tained from the data of Table II of [9]• Y' = (56.258, 56.579, 52.661, 51.315, 55.500, 57.461, .57.475, 50.396, 58.515, 56.323, 62.023, 61.673, 56.583, 56.924, 56.111, 62.085, 54.217, 55.914, 53.974, 49.754, 56.034, 57.895, 55.440, 62.863) The model used is the one given by (3.3)• Now using the transformation Z = TY , the vector of transformed observations Z is z ! = (-0.227, 3. 068, 3.335, -1.387, .-0 .812, 5.556, 1.550, -3. 759, -2.355, -0.241 , 0.524, -4.803, ■-1.200, 0. 891, 4.285, -1.316, 1.245, ■-5.548). The best estimate of g given Ny (3. 9) is A P1 = (.62, -I. 19, .27 , .60, - .65, -.39,' .47, -3.66, .88, -.23,-4 •7, -5.157, 3.62 , 4.71, -4.59). 38 By neglecting- order the best estimate of f3^- given by (3 .1 0 ) is Bi = (3.06, -3 .3 9 , .0 1 9 , .58, -.76, -.26, .47). In Table 3*2, the analysis o f variance corresponding to. ' Table 3.1 is given. Table 3.2 I 88 DF SV ITotal 18 MS F .P-value 154.18 ' Factors 7 5 6 .9 9 8.14 10.71 .0 3 8 7 Orders 8 94.91 11.86 15.61 .0229 Error 3 2.28 .76 S The test of significance indicates that the order effects are significant. used. Table 3 . 3 Therefore the model given in (3.9) is lists the parameters, estimates, standard errors and tests of significance for this model. The t entry in the table is the usual t-test of significance where .t - = ' estimated effect n standard error of estimate The estimates of the variances are given by VarIci1B) = k (R1R) For this example, (R1R) I 'ov . is given in the Appendix. 39 . Table 3.3 Parameter A ■ t S P-value .570 1 .0 8 .36 -1.19 .‘570 -2.08 .13 C .27 .570 • .47 .67 AB . .6 0 .445 1.35 .27 AC -.65 - .445 1.46 .24 B C -.39 .445 .88 .44 .47 .399 1 .1 8 .32 .8 1 5 ’ 4.49 .02 1 .0 8 .36 ABC ' a ■ -3.66 CO H CD .88 00 H Ul h Standard Error .62 B- * . Estimate CO CXJ CD -.23 ■ .815 . .28 .80 -4.70 1.04 -4.52 .02 -5.67 i.o4 ' -5.45 .01 6 123 8132 e2l3 CD UO H PO e231 3.62 1.04 3.48 .04 4.71 i.o4 4.52 .02 -4.59 i.o4 -4.4i .02 Using Condition (4) of the model, ^he.additional estimates are obtained. 02i = 3.66 ' 9 31 “ '•88 ' 4o Ggg - '23 /N G3 2 I “ ^ '^ 2 This example is given to illustrate possible interpre­ tations when order effects are determined, to be significant. If the standard factorial analysis had been performed, the A and the B main effects.would have been judged significant at the 10$ level on the basis of the.results of the 24 tests. However, the FSD indicates the observed effects are partly due to the order of the application of the factors. Significant effects include the B main effect (p-value = .13), the two factor order effect 8 and. the three factor order effects. The estimates of the three factor order effects form two equivalence classes {0123 = -4..70, = "5.67, O ^ 12 = -4.59} and {e2i3 = 3.62, P231 = 4.Tl5 Gggl = 6'°62} . Since 0^2 was the only significant two factor order effect, the system developed in Chapter II could have been used to predict the possible existence of these classes. 4l AsSLime this experiment was designed to test a sequential- process for cleaning grease and particles from transparent circuit boards. ments There are three possible treat­ (cleaning techniques)5 a chemical treatment (A)j a vacuum cleaning (B), and a second chemical treatment (C), The board is tested for cleanliness by measuring the diffusion of light passed through the b o a r d . , Lower response values (little light diffusion) indicate cleaner circuit boards„ The previous analysis indicates there are no significant factor effects; however5 vacuum treatment B seems to be the only one significantly reducing diffusion. There also seems to be a catalytic effect by doing the vacuum treatment B . after treatment A. The experimenter probably would recom­ mend one of the three treatment.sequences related to the set of effects,' ' {©ABC* 6A C B 5 6 CAB^ or he could conduct further experiments specifically designed to see whether treatment C was really necessary. 42 Example 3 « 2 r The following set of observations is obtained, from Example 2 of [9]* Y' = (9:219, 1 5 .0 6 1 , 1 4 .9 5 0 , 1 7 .5 5 0 , 1 1 .0 9 5 , 15.964, 15.732, 14.033, 9.541, 8.517, 15.213, 13.922, 9.104,-11.484, 9.109, 14.153, 9.045, 9.173, 1 4 .7 0 1 , 12.596, 9 .2 2 6 , 10.591, 15.806). The transformation Z = TY yields, -1.798, Z 1 = (-4 .1 3 1 , -2 .2 9 4 , -3 .8 7 4 ,. -3.443, 0.200, 0.724, -5 .0 4 9 , -2 .4 5 2 , -1 .6 8 3 , 0.968,. -3 .6 8 4 , -0.091, -4:676, -3.190, 2.383, 0.261, -4.332). The best estimate of P given by (3.9) is P ' = (5.78, .5 0 , -1.21, -.16, .02, .14, 3.16, -1.75, .44, -.17, -.14, .3 0 , -.73, -.52, -.1 5 ). By neglecting order the best estimate of P , given by (3.10) is p = (5.54, .56, -.12, .2 9 , -.75, -.33, -.14). A The first seven entries of p 1 estimate the same param~ A eters as the corresponding entries in p . By inspection the estimates appear to be approximately the same. Intuitively, it appears that order has little or no effect for this experiment. This conjecture is verified by constructing the analysis of variance table corresponding to Table 3.1. ■ 43. Table 3.4 DF ' Total Factors Orders Error - . SS MS F 9.769 -=t O SV .666 .7 1 18 158.395 ' 7 141.202 20.172 8 10.999 1.375 3 6.194 2.065 P-value The F test for order effects indicates order of application is insignificant. Therefore, the model given in (3.10) is used. Table 3•5 lists the parameters, estimates, standard errors and.' tests of significance for this model. Table 3.5 Parameter. Standard Error tS P-value .006 . .80 6.93 .29 .80 .36 C -.75 .80 -.94 AB -.33 .69 . -.47 .56 .69 .80 BC -.12 .69 ■ -.17 .88 ABC -.14 .65 -.22 .84 A . 5.54 B .74 .42 • .6 7 CO , '' -d* ' AC I Estimate The only significant effect is the main effect A and its significance level is less than .01. 44 Fractional Replications of 2^ FSD Designs The purpose of. Prairie and Zimmer's 1 9 6 8 paper [10] ■Q U c was to present some fractional designs of 2 , 2 , 2 ^ and 2 6 FSD designs. The designs were constructed to satisfy the following characteristics. (a ) . All main effects and two factor interactions are estimable. (b) For a given design the variances of all main effects are equal and the variance of all two factor interactions are equal. Also, the variances of all k-factor interactions are equal for fixed k. ( c) The importance of some order effects can be tested by a test of significance. The purpose of this section is to introduce the notation and terminology invented by Prairie and Zimmer. The designs are of two basic types, ,( a) l/r x 2 ^ designs ( b) s/p x (l/r x 2 P ) designs„ The l/r x 2^ designs are fractions of the 2^ FSD requiring (l/r)pI units where- each unit is tes.ted p+1 times. 45 All factorial effects are estimable and information is sacrificed only on order effects„ In the s/p x (l/r x 2^) designs5 each of (l/r)pl units is subjected to s factorS5 s < p 5 and s+1 tests. . The analysis of these designs is similar, to the analysis presented, in this chapter. referred to [10]. For more detail the reader is The terminology presented in this section will be utilized in the development of one at a time plans. CHAPTER IV ■ ONE AT A TIME- PLANS FOR THE 2 3 FSD Preliminary Considerations The plans proposed in this section are designed to provide estimates of order parameters as soon as possible in the experimental sequence. The main purpose of the FSD is to decide whether order of application of factors is important. ■. By the time enough trials are run to provide estimates of order effects, main effects and two factor interactions are estimable. Usually the three factor inter­ action is estimable or is estimable .with the addition of one more run. The estimates used for the one at a time plans will d e ­ constructed to remove the unit effect, but will not neces­ sarily be of minimum variance. The estimates used are linear combinations, of the observations and therefore, they are ■ simple and.easy to calculate. The ease of calculation and simplicity will be useful for the one at a time experimenter. The variances of the estimates given are compared to t h e ' variances for the best linear unbiased estimators for the full 24 runs as found in Chapter III. The variances of the estimating contrasts follow readily from Theorem 3.1. 4? Lemma 4 . 1 : If . are are two responses from the ■ same unit i , then Var(Yl j - Y lfc) = S a e 2. Proof: By Theorem 3.1, Var(Y^)-= -Ij ^ ^cl CovfY^j, Y^) = Therefore, Var(Tlj - Ylk) = Var(Yl j ) + Var(Ylk) - 2 CovfYl j , Ylfc) = (°e + °u) + (°s + au) " 2au . ■. Lemma 4 . 2 : = 20EIf Y^^, Y ^ are two responses from unit i, and Ygm , Y^n are two responses from unit £> then Var[(Y1 j - Ylfc) - ( Y ^ - Yj n )] = 4a2 . Proof: Again applying Theorem 3.1, V a r f t Y y - Ylfc) - (Y^ - Yin)] = Var(Yy ) + Var(Yy ) ' + Var (Yim) + Yar(Yin) - 2 CovfYlj, Yy ) - 2 CovfYy ., Yim) + 2 CovtYlj, Yjn) + 2 CovfYy , Yim) - 2 Cov (Yy , Yjn) ' ' - 2 ^ v ( Y i m , Yi n ) 48 = 4(0^ + a 2^ _ ga^ - 0 + 0 + 0- 0- 2a^ = 4a J . This completes the proof of the lemma. The notation used in the one at a time plans will be as follows. T h e .individual trials will be called runs. For each run the unit and the number of the test the unit has been subjected to will be given. This labeling will make it easier to relate the one at a time plans to the model presented in Chapter III. The plans presented are runs from Q one replication of a 2 FSD. Therefore, there must be six units available, one for each specific order of application. The six .possible application sequences will be assigned to the units randomly; however, the units will be labeled to correspond to the alphabetical order of the application sequence as follows: Unit Application Sequence I abc 2 acb 3 bac 4 bca 5 cab 6 cba 49 This is the same convention used in constructing the matrices in the Appendix. The sequence coding specification will refer to treat­ ments that have been applied before a specific test. The specification will always be denoted by lower case letters with the notation (1)^ for no treatments on unit i. Thus (1 )2| will mean that no treatment has been applied on the fourth unit3 and (ba)^ will indicate that the third unit has been subjected to the two treatments b a n d a in that specific order. It is not necessary to subscript ba with the unit number since the factors are applied in this se­ quence only on unit 3• The notation for this observation for the model given in (3.1) is Y ^ ( B 3A). The specification ba is another simpler symbol for this same response. One at a time plans are developed for both of the cases considered in Chapter II. The systems developed for order . effects for each case are used to augment an initial set of runs to achieve unbiased estimation of order parameters as soon as possible in the sequence. According to Daniel [3] •> the one at a time experimenter who achieves good results looks for effects three or four 50 times the magnitude of the experimental error. If the experimenter has some prior knowledge of the magnitude of the random error in his experiment, he will he able to look for these large effects without first estimating cr^. One At a Time Plans for Case I This case is the one where symbols related to insignifi­ cant order effects commute only on the left of the appli­ cation sequence. The first nine runs will be on units designated .1 , 4, and 5 with application sequences abc, bca, and cab respectively.. This will enable the experimenter to get biased estimates of all main effects, two factor inter­ actions and two factor order effects. The experimenter may get some information from these estimates for use in plan­ ning the next sequence of runs. The first nine runs are: 51 RUN UNIT TEST S E Q . NO . NO. NO, SPEC. I I I 2 I 2 3 I •' 3 4 4 I ’ 5 4 2 . ESTIMABLE FUNCTIONS (at run indicated) 1I a 4 3 7 5 ■I 8 5 2 9 5 a-1! A-AB-AC+ABC ab 14 .b B-AB-BC+ABC 2 (AB-ABC)+S12 '6 ' ESTIMATORS ab—a—b+lj^ be 1S C-AC-BC+ABC C ca 3 C_15 2(B C -ABC)+S2 3 bc-b-c.+lr5 2 (AC-ABC)-S13 ca—c—a+l1 Using the model. given in Equation (3.6), E (a — 1I J •= {m + 1/2(A - B - C - AB - AC + BC + ABC)} — (m + 1/2(-A - B - C + AB + AC + B C - ABC} . ■'= A - AB - AC + ABC. Similarly, ■ E(ca - c) = A - AB + AC + ABC and by subtraction, ,E(ca c - a + I1) = 2 (AC + ABC) - S13. 52 By applying 'Lemmas 4.1 and 4 . 2 5 the variances of the estimates are Var(a - I1)'= 2a^ 2 V a r (ca, - c - a + I1 ) = 4c^. The other estimates are estimated by the appropriate contrasts and can be.derived similarly using (3.6). The variances of these estimates can also be calculated using the. appropriate result, from Lemmas 4.1 and 4.2. At this stage of the experiment, main effects and order effects are confounded with two and three factor inter­ actions. .Many one at a time experimenters assume that three factor interactions are insignificant, and under this assumption,.estimates of two factor order effects confounded ■ with a corresponding two factor interaction can be found. The probability that these two effects are offsetting is very small so if the estimate of 2AB + 0 ^ is "small”, the experimenter may conclude after five runs that there is no significant two factor order effect related to treatments . A and B (and also ho AB interaction). Based on the assumptions of negligible three factor interactions and of the small probability of offsetting 53 effects the experimenter can conclude that all or some of the two factor interactions are insignificant. There are four possible situations which can result depending on how many■ of the effects can be assumed insignificant. They are: (1) all three' two factor order-effects insignificant/ (2) two of the three two factor order effects insignificant, (3) one of the three two factor order effects insignificant, (4) none of the three two factor order effects insignificant. If a scientist was unwilling to make the above assumptions concerning three factor interactions and off­ setting effects, he would need nine additional runs to get unbiased estimates of the two factor order effects. This situation is experimentally identical to the fourth possi­ bility just listed. Each possibility will now be examined and sequences of runs will be given for each situation. I) Suppose all of these biased estimates of two factor order effects are small, leading to the conclusion that two factor interactions and two factor order effects are assumed to be insignificant. Thus, since ' 54 912 e21f 9 13 9S l 5 9 2 3 93 2 . by Lemma 2.1 the system for order effects Implies 9123 = 9 213^ 9 132 = 9 31 2 5 9231 = 9 321' Therefore, 'three more runs are needed, to test for equality of these classes of three factor order effects.. RUN UNIT TEST SEQ'. NO. SPEC. NO. NO. 10 I 4 abc 11 4 4 bca 12 5 ■ 4' ■ cab ESTIMATORS ESTIMABLE FUNCTIONS (at run indicated) abc-l^-bac+1^ 9 123 “ 6231 abc-l^-cab+lp. 9 123 " 0 312 . e231 " 0 312 • ■ bac-l^-cab+1 ^ Each of the differences is estimated by the indicated. contrast, and the variances of the estimates follow from Lemma 4.2 and. are equal to 4a£ . ■ The estimators of these differences based on twelve runs compared favorably to the least square's estimators from the full F S D . Using the matrix (R1R ) - 1 found in the Appendix, the variance of the least.squares estimator for one replication requiring 24 2 runs is 3 .5 o"e-• ■ If the estimated contrasts from runs .10,'11 and 12 are small, the one at a time experimenter w o u l d .assume there 55 were no significant three factor order effects because he is looking for effects approximately three or four times the magnitude of his experimental error. O This sequence of 12 runs is the (1/2 x 2J) fraction of the FSD given in [10]. For this fraction there are no degrees of freedom for error; if the experimenter wanted an estimate of error, an additional four tests, on another unit could be run to yield a (2/3 x 2^) FSD with one degree of freedom for error. 2) Suppose two of the three estimates for two factor order effects are small compared to the assumed magnitude of the random error. Then two of the three two factor order effects are judged insignificant and assume 6 effect which may be significant. is the Because the estimate was biased, the significant response may have been a result of the interactions which biased the estimate. assumed that ■G ’ ^31 0 23 6-32 ^ ^ the system developed for Case I implies G — ^312 ^231 ^ 321' Since it is .56' . In this situation the. following additional runs are suggested.. The first three to provide an unbiased estimate of 0^25 and the last four to estimate the three factor order effects. RlM UNIT TEST SEQ. NO. NO. NO. SPEC. 10 3 •I 11 3 ,2 12 3 3 ■13 3 ■4 b:ac 14 I ' 4 abc 15 4 4' bca 16 5 4 cab ESTIMABLE FUNCTIONS (at run indicated) ESTIMATORS ■ 1S b ba ab-I1-LaRl^ 2012 abc-l^bac+lg 0 123 " 0 213 0 123 ■“ 0231 ■ 0123 " 0 312 abc—11-bca+l^i abc-l1-cab+l^ The unbiased estimate of 012 is S12 = 1/2(ab - 1^ - ba + l^). Applying Lemma 4.2, Var(G12) = l/4(4o2) = ( A This compares favorably to the minimum variance, estimate 2 based on '24 runs which has variance equal to .875crG After 16 runs have been completed, two of the two factor order effects have been judged insignificant, an 57 unbiased, estimate of exists, and estimates of the three factor order effects have been obtained. Hence, this one at a time plan of 16 runs allows the experimenter to determine if order effects are significnat. The 16 runs constitute O a (2/3 X 2J ) PSD with one degree of freedom for error. 3) The third possibility occurs when only one of the three two factor order effects is judged insignificant on the basis of the biased estimates from the first nine runs. Assume the insignificant effect is G2^, than Lemma 2.1 implies S23I = 0 3 2 1 - The following six runs will produce unbiased estimates of 0 no and 0 § RUH UHIT TEST S E Q . HO. HO. SPEC. 10 3 I '■ 11 3 2 b 12 3 3 ba 13 2 I 4 14 2 2 a 15 2 3 ■ac 23' ESTIMABLE FUHCTIOHS (at run indicated) ESTIMATORS 1B •ab—I^-ba+!^ 2612 ac~l2 ~ca+l^ 2 6 I3 A A The two. unbiased estimates 0 ^ 2 and 0 ^ 2 equal to ae. as indicated, earlier. have variance If one or both of the two ■factor order effects related, to these estimates are 58 judged insignificant, the estimates of three factor order effects can be found by augmenting with the proper three factor sequences previously discussed in possibility one or two. However, if both are judged significant, runs 16-20, given below, are necessary for the estimation of three factor order effects. RIM UNIT TEST S E Q . NO. NO. SPEC. NO. 16 I .4 17 2 ' 4. acb 18 3 4 bac 19 4 4v bca 20 5 '4 ESTIMABLE FUNCTIONS (at run indicated) ESTIMATORS ■ abc cab abc-l^-acb+lg 8l23 ” 6 132 - abc-l^-bac+lg 6 123 ™ e2l3 8123 0231 abc-l-^-bca+1 ^ ■ abc-l^-cab+1 ^ 8 1 2 3 " 6 312 As before , the contrasts, constraints and system yielded estimates of two and three factor order effects as well as the usual factor effects while allowing two degrees of freedom for error. PSD. . 4) O This design of 20 runs is a (5/6 x 2'5) I . The fourth possible situation occurs when the experimenter does not want to base his judgements on biased estimates "or if all of the biased estimates indicate that all two factor order effects may be significant. The 59 following nine additional runs provide unbiased estimates of two factor order effects.. RUN UNIT TEST SEQ. NO. NO. NO. SPEC. 10 2 'I 11 2 '2 a 12 2 3 ab 13 3 I 14 3 2 b 15 3 3 ba 16 6 ■I ■ 17 6 '2 18 6 ESTIMABLE FUNCTIONS ..(.at run indicated) .ESTIMATORS * 26I3 ac-Ig-c'a+/ • /3 : C 3 .'cb ab-l-^-ba+1 26IS be—lj|-cb+l| ^23 These 18 runs form a 2/3 x ( l x 2^) FSD. If any of the above estimates are judged insignificant, these tests can be augmented, b y the appropriate runs where three factors have been applied by using the appropriate situation from the first three possibilities presented. If all two factor order effects appear to be significant, each unit must be tested again.after application of the last treatment. completes a full FSD which can be analyzed using the techniques developed in Chapter III. This 6o These four situations consider all possibilities for examining one at a time plans' assuming only the basic ordering operation (Definition 2.1) and commutativity of insignificant order effects only on the left (Lemma 2.1). If in addition one assumes commutativity of any factors previously judged insignificant (Definition 2.2), then the second case of Chapter II is encountered. 6l ■One At A Time Plans For Case 2 This situation is where the symbols related to insig­ nificant effects commute anywhere they appear in the sequence. This additional assumption reduces the number of runs required to estimate the order effects. The first nine runs are the same as those for Case I 3 and if the experimenter is willing to make the same assump­ tions for determining if two factor order effects are insignificant3 then the same four situations that occurred in the plans for Case I will occur for Case 2. I) All three biased estimates are small in magnitude compared to the size of the a priori random error. In this Situation3.the three two factor, order effects are determined to be insignificant. The system for order.effects developed for Case 2 implies ='0213 = 0231 = 8321 = 0 312 = S132 by Lemma 2.1 ■ by Definition 2.2 by Lemma 2.1 by Definition 2.2 by Lemma 2.1 Therefore3 all of the three factor order effects are equal. The fourth model assumption (3*2) states that the 62 sum of all of the three factor order effects is equal to zero. Therefore, the system for order effects implies there are no significant three factor order effects. One more run is required to have unbiased estimation of all factorial effects. RIM UNIT TEST SE Q . NO. NO. NO. SPEC. 10 2) .I 4 abc ESTIMABLE FUNCTIONS (at run indicated) All factorial effects The. second possibility,occurs if one order effect is possibly significant. Assume 9 the order effect in question. ■ Then the system implies there are possibly two equivalence classes of order effects. {©123^ 6^32^ ®312^ ® 213 3 e2315 ® 321^ The following five runs will provide estimates of these order effects. RUN UNIT TEST S E Q . NO. NO. SPEC. NO. 10 3 "I 11 3 '2 12 3 13 I ■’ 4 '• abc ' l4 3- ; 4 '3: ESTIMABLE FUNCTIONS (at run indicated) ' ESTIMATORS 1S b ba bac 2612 6 123 " e2l3 ab-l^-ba+l^ abc-l-^-bac+lg 63 At this stage of the experiment, estimates of the factor and. order parameters can be found using the con­ straints, the estimable functions, and the equivalence classes. 3) The third possibility is when two of the two factor order effects may he significant. the order effects in question. Assmue 0^2 anc^ 0 i3 are In this instance, the system implies there may be four equivalence classes of three factor order effects. {0123, 0132^ 5 te23l> 9321$ 5 132} 5 ^ 231} The first six additional runs will provide unbiased estimates of the pair of two factor order effects under investigation. The next four runs will provide contrasts which will enable the experimenter to estimate the three factor order effects. 64 RUU UNIT TEST S E Q . NO. NO. NO. SPEC. 10 3 I 11 3 2 12 ‘3 3 13 2 14 2 '2 15 2 3 16 I -4 abc 17 2 -4 acb 18 4 19 5 ESTIMABLE FUNCTIONS (at run indicated) ESTIMATORS • 1S 'b ba 2 0 12 ab-l^-ba+l^ I ... I2 4' • a ac bca . 4 ; cab 2013 0 123 “ 0132 . 0 123 " 0231 0 123 " 0312 ac-lg-ca+l^. abc-l^-acb+lg abc-l^-bca+1 ^ abc-In -cab+lr1 5 Again, the contrasts, the conditions of the model, and the. equivalence classes provide estimates of the order effects." Using these the experimenter can get estimates of factorial effects. 4) Possibility four is identical to the fourth situ­ ation discussed in the plans for Case I. The experimenter does l8 runs to estimate two factor order effects and then one to six more runs are necessary to estimate the three factor order effects. 65 Examples -.1 : If the data from Example time analysis, the first nine RUN S E Q . SPEC. I R 56.258 2 a 56.579 3' ab 52.661 14 56.583 5 b 56.924 6 be 56.111 "4 7 ■ OBSERVATION 54.217 1S 8 C 55.914 9 ca 53.974 The following "biased estimates of two factor order effects are obtained. After run f i v e : 2 (AB - ^BE) + G = 52.661 - 56.579 - 56.924 + 56.583 -4.259 66 After run eight: 2(BC Sc) + S2 3 = 56.111 - 56.924 - 55.914 + 54.217 = -2.510 After run n i n e : 2(£c - ABC) + G13 = 53.974 - 55.914 - 56.579 + 5 6 . 2 5 8 = 2 .2 6 1 If the experimenter has prior knowledge of his experi­ mental error, he would be able to make decisions concerning the possible two factor order effects. In particular, for A this experiment, it is known from Example 3.2 that = .8 7 . Consider the. following two instances; one where the expert.‘ 1 A menter assumes a a G ■ * is about . 5 and the other where he assumes is about .1. A I) For. cre = .5 5 the experimenter would decide after nine runs that all of the effects were possibly significant and would do the following additional- runs.. 67 RlM 10 SEQ. SPEC. OBSERVATION V 58.515 .11 b 56.323 . 12 ba 62.023 13 1E 55.500 a 57.461 ac 57.475 ' 14 15 16 '1S 56.034 17 C 57.895 . 18 cb 55.440. The following unbiased estimates are obtained. After run twelve: 8l2 = 1/2(56.661 - 56.258 - 62.023 4- 58.515) =-3.553. After run fifteen: = 1/2(57.475 - 55.500 - 53.900 + 5.4,217) = ’1.146.• After run eighteen: S23 = 1/2(56.111 - 56.583 - 55-440 + 56.034) = .0 6 l. 68 The experimenter would decide that only was significant, and then do two or four more runs depending on whether he was using Case I or Case 2 of the system for order effects. Suppose he is operating under Case I requiring four more runs. RUN S E Q . SPEC. OBSERVATION 19 ah c .51.315 20 bca .62.085 21 cab 49.754 22 bac 61.673 The following contrasts among three factor order effects can now he formed. After run twenty: (S1 2 3 r S231) = 51.315 - 5 6 . 2 5 8 - 6 2 . 0 8 5 + 56.111 = -10.917 After run twenty-one: "8312) = 51.315 - 56.258 - 49.754 + 54.217 . = — .48 After run twenty-two: (9213 " ^ 3Ig) = 61.673 - 58.515 - 49.754 + 54.217 = 7.621 69 Also after run twenty-two: 62.085 (^231 ~ ^ 2 1 3 ) 56.111 - 61.673 + 58.515 = 2.8l6 The above differences imply the existence of two equivalence classes of three factor order effects 61235 9132’ e312^ and { S2 13, d 2 3 1 ’ 9 32 J ' If the- assumption of the system for Case 2 (symbols commute anywhere) had been made, the above classes would have been verified by runs' 1 9 and 20. 2) Oe' = I. In this case, the experimenter would have assumed, after the first nine runs, that insignificant. and S2 3 were He would then do runs 10 to 12 and determine that 0 ^ 2 was significant. After these twelve runs, he would then complete the experiment by testing for three factor interactions'in exactly the same way as in the previous situation where 0 E was assumed to be .5 . That is, after completing runs I through 1 2 , he would immediately do only four additional runs, I9 through 22. 13 through'-18, would not be needed. The.other six runs, Conclusions regarding significance of order effects are identical to those obtained when a E was assumed equal to .5. 7P The problem of estimating a G is difficult, and the experimenter should be aware of the consequences of a wrong estimate. The more conservative approach is to under­ estimate and as a result more runs may be necessary to determine whether or not order effects are significant. CHAPTER V ONE AT A TIME PLANS FOR 2P PSD ■, . Introduction In the previous chapter, one at a time sequences for the 2 PSD.were examined. Using the systems for order effects to form equivalence classes, the number of runs necessary to estimate order parameters was reduced from twenty-four to as few as ten.or twelve depending on whether or not the.experimenter was using Case I or Case 2 of the system for order effects developed in Chapter I I . ■ When p > 3j> the number of units and number of tests required increase rapidly. For example, a single replication of a h full 2 PSD requires 24 units and 120:tests. The use of equivalence classes is very helpful in reducing the number of runs required for estimation of order effects. In the' previous chapter, all possible instances of a 2^ PSD were given to illustrate the procedure for deciding on additional sequences of' runs on the basis of the infor­ mation derived from the funs already completed. enumeration of distinct possibilities for the 2 not be given in this thesis. A complete 4 PSD will However, a,sample situation is presented to illustrate the procedure for deciding on the runs necessary to estimate the order effects of inter­ est. The. situation is presented without data and at suc­ cessive stages of the experiment, some effects are arbi­ trarily assumed to be significant in order to illustrate the use of the system in determining whether order effects are present or not. ■ h An Example of a One at a Time Plan for a 2 FSD V When planning one at a time plans for a 2 .FSD, the experimenter has considerable latitude in selecting the initial sequence of runs. There are ( ^ ) = 6 sets of two factor sequences from which the experimenter could form . • ■■■ * ■ preliminary .estimates of two factor order effects. Be­ cause there are only four factors, two must be replicated in order to get the first eighteen runs.. These replications will provide greater precision in estimating these two main effects as well as provide two degrees of freedom for an estimate of experimental error. Suppose'.A and B are the main effects which are most important, then the first eighteen"runs are: .. - 73 ROT UNIT TEST S E Q . NO. NO. 'NO. SPEC. ESTIMABLE FUNCTIONS (at run Indicated) ' ESTIMATORS e 12+2(AB-ABC-ABU+ABCO). ab-a-b+lg C 6.23+2 (BC-ABC-BCD+ABCD) bc-b-c+lg -6 13+2 (AC-ABC-ACD+ABOD) c a— c—a+1^ -9g^+2(CD-ACD-BCD+ABCD) dc-d-c+l^ I I ■ I' 2 I '2 a 3 I 3 aU 4 2 I 5 2 2 6 2 3 7 3 .1 8 3 '■ 2 9 3 3 ca IO 4 ■T 1|| ll 4 . 2 d. 12 4 ■ 3 dc 13 5 l4 5 ,2 15 5 '3 16 6 I- 17 . 6 2 b 18 6. 3 bd I'. 1I R "be 1B 1S ? . _ . ' a ad. 9 ^ + 2 (AD-ABD-ACD+ABCD) ad— a—d+1^ 8 2 h+2 (BD-ABD-BCDtABCD)' bd-b-d+1^ 1G 74 After these l8 runs, the experimenter has available estimates of all two factor order effects. As in the 2^ F S D , the estimates of two factor order effects are biased, but if there is some prior knowledge of the random error, some of the order effects and the corresponding interactions can be judged insignificant. Suppose that and are possibly significant on .the basis of the first eighteen runs . Then the following nine runs should be completed. to find unbiased estimates of e 1 2 5 ©2 4 and ©3 ^ ® RUN UNIT TEST S E Q . . NO. SPEC. NO. NO. 19 7 .I 20 7 . 2- 21 7 ■3 ba 22 8 I 1S 23 8 2 d 24 8 3 Sb 25 9 'I 26 9 2 . C 27 9 3 cd ESTIMABLE FUNCTIONS (at run indicated) .' ESTIMATORS 4 b ^12 . ab-l-^-ba+l^ 2e24 bd-lg-db+lg 34 cd-l^-dc+1 ^ 1S 20 After these runs, assume the estimates of 0 2 4 are significant, but that S ^2 is insignificant. and . 75 Up until this point in the example, the 2 7 runs completed are suitable for either Case I or Case 2 assumptions. The remainder of the discussion will assume Case 2, where non­ significant order effects may commute regardless of where they appear.in the application sequence. Thus, using Lemma 2.1,- and Definition 2.2, the following equivalence classes of three factor order effects can be derived assum­ ing @24 ^ £ 4 3 and 8 2 4 7^ djjg, ^ut that all other effects are equal. I9-1239 6 1325 6 3125 0 321> e2 3 1 5 9 213^ • {0 ^ 4 , 8 3 1 4 , 0 3 4 %}, t0 i4 3 5 9 4135 9 431 ^ {0124i 62l45 024l^ 9 (9l42* 94125 9421-^ ^ 92 3 4 9324^9 ^ 9432ji 9423^9 ^9342^ 5 ^ 9243^ Similarly, the following equivalence classes of four factor order effects can now be derived, using only information about two factor order effects. {9 12345 9 1324^ 9 312459 32l45 9 3241' 9 2134' 92314) {9 1423' 9 4l23' 9 4l32' 9 4312' 9 4213' 9 4231' 9 l432) ^91243' 9 2l43' 924l3' 9243l) {0 34l2' 8 3 4 2 1 ' 93i42' 9 1 3 4 2 ) 76 These equivalence classed may be further condensed using information concerning significant three factor order ef­ fects. All. six three factor order effects with subscripts which are permutations of the symbols I 5 2 5 and 3 appear in the same equivalence class. The fact that all of these effects can be considered equal5 in conjunction with the model condition (3 .2 ) that all of these effects must sum to zero5 implies that these six order effects are insignifi cant. Therefore the following runs sould be performed in order to estimate the remaining order effects of three and four factors. RUN UNIT TEST SEQ. NO. NO. .NO. SPEC. cad 5 4 ' adc 30 I 4 abd '8 " -4 dba 2 • 4 bed 33 4 4 deb 34 9 4 cdb 35 6 4 dbc 9 314 cad-I0-adc+Ic 3 9 cad-lg-abd+1^ 6124 " 9421 abd-1-^-d.ba+lg cad-lg-bcd+lg - 0432 bed—Ig-dcb+l^i bcd-lg-cdb+lg CO -dCM CD I -dCO CM CD 32 6 314 “ 9143 -3- 29 CD I 3 -=jCO CM CD I -3i —i OO CO CM CD CD 28 31 ESTIMATORS . .i CD UO -Pr PO ’■ 4 ESTIMABLE FUNCTIONS (at run indicated) . bcd-lg-bdc+lg 9234 77 RUN UNIT TEST SEQ „ NO. NO. NO. SPEC. 36 3 5 cadb 37 5 5 adcb 38 I 5 abdc 39 2 5 bcda ESTIMABLE.FUNCTIONS (at run indicated) ESTIMATORS c adb-1 0 - adcb+lr3 5 cadb-lg-abdc+1 ^ 6 3 142 " 9 1432 63142 “ 6 1243 93142 ™ 9 234l ■cadb-lg-bcda+lg T h e ■contrasts formed from runs 28 to 39 9 the equiva­ lence classes,and the assumptions of the model can then he used to form estimates of the order parameters. If the less restrictive assumptions for the system of order effects for Case I (symbols commute only on left) .had been used, more equivalence classes of order effects would have been formed and therefore, more runs would have been required to estimate the order effects. But the tech­ nique of examining two factor order effects for significance and then forming equivalence classes of higher level order effects is the same for both cases. illustration of this technique. This example is an One at a time plans for a 2^ PSD can be developed by using the system to form equiva­ lence classes of effects, and then estimating one effect from each class. 78 Example Example 5 . 1 : If the data from the example of a 1/2 x 2^ FSD appearing in [10], is subjected to a one at a time analysis assuming a £ = I 3 the first 1 8 runs are as follows: RUN I ' S E Q . SPEC. ' 1I OBSERVATION 8.51 . 2 a 15.67 3 ab 13.68 4 1S 10.77 5 b 10.08 6 be 7 8.77 9.86 8 h C 11.04 9 ca 15.39 - 10 14 11 d 12 dc ' 9 .6 6 1 0 .1 5 8.97 10.13. 13 14 a 15.53 15 ad 13.66 79 RUN S E Q . SPEC. 16 16 OBSERVATION IO.9 6 17 b 1 1 .0 6 18 bd 10.15 The following estimates are formed. After run 5: S12 .+ bias = 13.68 - 1 5 . 6 7 - 1 0 . 0 8 + 10.77 = -1.30 After run 8: . + b i a s = 8.77 - 1 0 . 0 8 - 11.04 + 9.86 = -2.49 After run 9: ■ - 0 1 3 +'bias 15.39 - 11..04 - 15.67 + 8.51 -2.8l ■ After run'.-12: -0 ^ + bias = 8 . 9 7 - 10.15 - 11.04 + 9.86 = -2. 3 6 ' After run 15: ®l4 + ' = 13.66 - 15.53 - 10.15 ■ 9.66 = -2.36 8o After run 18: + . M a s = 10.15 - 11.06 - 10.15 + 9.66 = -1.40 If the one at a time experimenter Is looking for ef­ fects three or four times the magnitude of the experimental error, none would appear to be significant ' 0 ^2 * To be safe. do three runs to estimate RUN S E Q . SPEC. OBSERVATION 19 I7 7.97 20 a 16.32 21 ac 16.63 The following unbiased estimate of 9 ^ 13' found after run 21, 8^2 = 1/2(16.63 - 7.97 - 15.39 + 9.86) = 1.565 The conclusion based on the assumption a£ = 1 insignificant. is the 8^^ is If Case 2 of the system for order effects is assumed, no two factor order effects would imply no ■ ■ 4 order effects for the 2 FSD. For any other two factor order effect which might be significant, an additional three runs will be sufficient to obtain an unbiased estimate. 8l Prairie and. Zimmer in [10] ran a 1/2 x 2 Zi fraction of 6 0 runs to arrive at the same conclusions reached after only 21 runs using the one at a time plan under Case 2 assumptions. The assumption of Oe = 1 could have been re­ placed b y an estimated value of 1 . 6 obtained from runs I, 2 5 4 5 5.5 13 5 l 4 5 1 6 , and I/. If this estimated value had been used, the experimenter would have had no reason to conduct more than the first 1 8 runs. If this example is considered under assumptions of Case I then 52 runs are required, which is a saving of only eight runs (or tests). Using Case I, the following equivalence classes for three factor order effects are ob­ tained. {01235 6213^ 5f0i3256312^ 5 t63215 0231^ {0124.5 e2 l 4 ^ ^0 l425 04 l 2 ^ {0 1 3 4 , ^ 64215 9 24l^ 5 {0 ^ 3 5 8 4 1 3 ] 5 ^ e34lj’ 6 4 3 1 ^ C 8 2 3 4 ^. 0 3 2 4 ^ ^ ^■0243-’0 4 2 3 ^ 503 4 2 04 3 2 ^ Because there are twelve classes of three factor order effects, ■twelve units will need, to be tested. Only seven units were used, initially3 therefore, five additional units must be tested which incidently, give unbiased estimates of 82 order effects and a total of error. RUN SEQ . SPEC . OBSERVATION 22 abc ■ 14.25 23 bca 16.13 24 cab' 16.27 25 deb 11.03 . 26 adb 14.79 . 27 bdc 9-03 28 acd 14.18 29 1S ■ 30 b 10.37 . Si' ba 15.74 ■ 32 ■ bad 14.99 33 4 34 C 35 ’ cb 36 cbd 37 1IO 38 C 39 cd 4o c da 8.01 10.06 10.98 8.18 .1 0 .9 4 9.78 10.21 9.92 15.28 ' 83 rim- 41 S E Q . SPEC. OBSERVATION 8 .6 0 1Il 9.43 - 42 d 43 da 14.86 44 dac 13.87 45 R 2 9.32 46 d. 10.99 47 dh 11.86 . 48 dha 16.08 The following contrasts are formed. After run '24: (8123 e2 3 l) ^ 14.25 8.51 - 16.13 + 10.77 8.51 1 6 .2 7 > 9.86 8.01 14.79 + 10.13 8 .0 1 1 6 .0 8 + 9 .6 6 = .38 After run 25: ^123 ~ 8 312) 14.25 -.6 7 After run.32: (e2l4 “ ®l42) == 15.74 = 2.97 After run 48: (e2l4 ■“ 0421^ = 15.74 = 1.31 84 After run 44: (9134 " e4l3^ 14.18 7.97 - 1 3 .8 7 + 8.60 .94 After run 40: (0134 “ e34i) ~ l4.l8 7.97 - 15.28 + 9.78 = .71 After run 3 6 : (9324 " 8 2 4 3 ) 10.94 10.06 - 9 . 0 3 + 10.96 1.91 After run 3 6 : ( 8 3 2 4 "'8 4 3 2 ) ^ 10-94 10.06 - 1 1 .0 3 + 9 .6 6 = - .4 9 The contrasts5 the equivalence classes, and the con­ ditions of the model imply there are no three factor order effects.• Four equivalence classes of four factor order effects result by appling Lemma 2.1 and the fact that three factor order effects are insignificant. To estimate the possible significance of order effects from these four equivalence classes, an additional four runs are necessary. 85 RUN S E Q . SPEC. OBSERVATION 49 ABCD 15.02 50 BADC 13.23 51 ACDB 14.20 52 CBDA 1 5 .5 6 After run 50: (61234 S2l43^ = 15.02 8.51 13.23 + 8 . 0 1 8.51 14.20' + 9 . 9 7 8.51 1 5 .5 6 + 1 0 .0 6 = 1 .2 9 After run 51: Ce 12.3k ~ 61 3 4 2 ) = 1 5 *0 2 = 2.28 After run 5.2.: (01234 “ ^324%) ^ 15.02 = 1.01 These contrasts together with the conditions of the model and: the equivalence classes imply no order effect. Also5 there are eight degrees of freedom for error which give an estimate based on the replications of each of the four factors for the first treatment on each unit.. The estimate of = . 6 5 or a = . 8 0 is close to, the initial A estimate5 cr = 1 . G These are the same conclusions con- cerning order effects that Prairie and Zimmer found in [10]. CHAPTER VI SUMMARY AND EXTENSIONS In experiments where the order of the application of ■ factors may he important, the 2p Factor Sequencing Design proposed in [9] is of use in determining order effects. The design considered for 2p factorial experiments where the factors are applied sequentially, the effects of the factors are permanent, each unit may be tested p+1 times without the test itself affecting the unit, and the low level of a factor is the absence of the factor. In this thesis, a model with explicit representation of the order parameters is formulated. In Chapter II, the underlying assumptions regarding order effects are examined and a system' with algebraic properties is developed to assist the experimenter in the estimation and interpretation of order effects. The model used is a mixed model with random unit effects and fixed treatment effects. The design and analysis of this model is presented in Chapter III along with examples to supplement the text. 87 Because the number of runs required, is (p+l)p'3 the size of the experiment becomes prohibative for p > 3. Methods for reducing the number of runs without losing the ability to estimate order effects are discussed in Chapters IV and. V. ' The one at a time plans proposed in those sec­ tions use the systems developed in Chapter.II to form, sequences of runs which allow for the estimation of members of equivalence classes of effects. These plans are very useful in situations where the experimenter can achieve quick results with small error. The possibility of extending these designs to factors with more than two levels is open to further study. 88 BIBLIOGRAPHY 1» Anderson, T. W. An Introduction to Multivariate Statistical Analysis. John Wiley and Sons, New York. 1958. 2. Cochran, W. C. and Cox, C. M. Experimental Designs. Wiley, New York. 1957. 3. Daniel, C. "One - at- a- Time Pla n s. 11 Journal of the American Statistical Association, ' 6 8 (1 9 7 3 ) : 353-360. 4. Fisher, R. A. The Design of Experiments. Boyd, Edinburgh. 1947• 5. Grayhill, F. A. Introduction- to Matrices with Appli­ cations in Statistics. Wadsworth Publishing Company Inc.-Belmont, California. '1 9 6 9 . 6. Graybill, F. A. An Introduction to Linear Statistical Models, Volume I. McGraw-Hill Book' Company, In c ., New- YorkV 1961. 7. Hunter, J . S. "A Sort of Least Squares Estimation After Each Run." Technometrics, 6(1964): 41-58. 8. Kempthorne, 0. The Design and Analysis of Experiments. John Wiley and Sons, New York. I9 6 7. 9. Prairie, R. R. and Zimmer, W. J . "2^ Factorial Experi­ ments with the Factors Applied Sequentially." Journal of the American Statistical Association, 59(1964): 1205-1216. Oliver and 10. Prairie, R. R . .and Zimmer, W. J . "Fractional Repli­ cations of 2'P Factorial Experiments with the Factors Applied Sequentially." J ournal of the American Statistical Association, 63(1968X7 644-652. 11. Searle, S . R. Linear Models. New York. 1971. John Wiley, and Sons, APPENDIX 90 The matrices which appear on the following pages are Q for one replication of a 2 ^ Factor Sequencing Design. The matrix formulation of the design and analysis is discussed in Chapter III. ' Equation (3.3) which appears on page I9 is: Y = X* Zm X + Wu + e. Equation (3.6) which appears on page 27 is: Equation (3.7) which appears on page 30 is: T = S ® I]. Equation (3.8) which appears on page 32 is: Z = Rg + e. Equation (3.9) which appears on page 33 is: I 0 91 Matrix X* I 0 0 0 0 0 0 0 0 0 I I 0 I 0 I 0 I 0 0 0 0 '0 I I 0 0 0 0 0 0 I 0 0 0 0 0 .0 0' 0 0 0 0 0 I. 0 0 0 0 ■0 ■ I 0 0 0 I I I .0 I I I 0 I 0 I 0 I I I I 0 0 0 0 0 0 I 0 0 0 0 O- 0 0 I 0 0 0 0 ■0 I I I 0 I 0 0 0 0 I I I 0 0 I 0 0 0 I 0 0 0 .0 0 0 .0 .0 0 I .0 0 0 0 0 0- '0 0 0 0 0 'I 0 I 0 0 0 0 0 I 0 0 Q '0 0 0 0 0 0 0 0 0 0 0 I •0 0 0 0 0 0 I 0 0 0 0 I 0 0 0 '0 0 0 0■0 0 0 0 I 0 0 0 0 0 0 0 I 0 0 0 I 0 0 0 0 0 0 I 0 0 0 0 ■1 0 0 0 0 0 0 0 I 0 0 0 0 0 I 0 0 0 0 I 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 I 0 0 0 ■0 0 0 0 0 0 0 0 '0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0' 0 0 0- 0 0 0 0 0 0 0 0 0 0 I 0 0 0 0 0 0 0 0 0 '0 0 0 0 0 0 0 :o 0 0 0 0 0 0 0 0 I 0 0 0 0 0 0 0 0 0 0 I 0 0 0 0 0 0 0 0 0 0 0 0 0 O O O 0 0 0. 0 O O O O O 0 0 O O O O O 0 0 O O O O 0' 0 .I O O O O O 0 0 O O O O O o. 0 O O O O O 0 0 O O O O O 0 ■0 I O O O O 0 0 ■ o. O O O O 0 0 O O O O O 0 0 O O O O O 0 0 O I O O O 0 ;o O O O O O 0 ' O'- O O O ■ O O 0 ■0. O . O . O O O 0 O O O I O O 0 '0. O O O O O 0 O .O O O O O 0. O O O O O O 0 '0 'O O O I O 0 O .0 O O O , O 0 O O O O •O O I -O O O O O O 0 O O O O O I 92 Matrix X I -.5 -.5- ;-.5 .5 .5 .5 -.5 O .0 0 0 0 O O O I .5 -.5 '-.'5 -.5 -.5 .5 .'5 O 0 0 0 0 O O O I »'5 .5- .5 -.5 - .5 -.5 I 0 0 0 0 O O O I .5 .5' .5 .5 .5 .5 O 0 0 I 0 O O O I -.5 .5 •5 .5 -.5 O 0 0 0 0 O O O I .5 -.5 -.5 -.5 -.5 .5 o5 O 0 0 '0 0 O O O I .5 -.5 -.5 •5 - .5 -.5 O I 0 0 0 O O O I .5 . *5 •5.- .5 .5 .5 .5 0 0 .0 I O O O I -.5 .5 •5 .5 0 0 0 0 O O O I -.5 -.5 ..5 ■ O -.5 O -.5 .5 -.5 o5 O 0 0 0 0 O O O I .5 .5 •5 -.5 -.5 -.5 -I 0 '0 0 0 O o- O I .5 •5 .5 .5 .5 .5 O 0 0 0 0 I O O I -.5 -.5. -.5 .5 .5 .5 o5 -.5 O 0 ■0. 0 ■0 O O O I -.5 .5 -.5 -.5 •5 -.5 •5 O 0 0 0 0 O O O I -.5 .5 .5 -.5 -.5 .5 -.5 O 0 I 0 o' O O O I '.5 .5 ■■ .5' .5 ,5 .5 .5 O 0 0 0 O O I O I -.5 -.5 ’- .5 •5 •5 .5 - .5 O 0 0 0 O O O O I -.5 -.5 •5 c5 -.5' -.5 <>5 O 0■ 0 0 O O O O I .5 -.5 .5 -.5 .5 - .5 -.5 O -I 0 '0 O O O O I ..5 .5 .5 .5 .5 .5 .5 O 0 0 0 O- O O I I -.5 -.5 .5 .5' .5 -.5 O 0 0 0 O O O O l' -.5 -.5 -o5 . .5 •5 -.5 -.5 •5 O 0 0 0 O O O O I -.5 .5 .5 -.5 -.5 •5 -.5 O 0 -I' .0 O O O O I .5 •5 ' .5 ' .5 •5 .5 »5 O 0 -.5 .-.5 . -.5 •5' -.5 ■-.5" ' -.5 0 -I -I -I -I -I Matrix T 0 .7071 -.7 0 7 1 .4082 .4082 -.8 1 6 5 .2887 .2887 .2887 0 0 o 0 0 - .8 6 6 0 0 0 0 0 0 .7071 -.7 0 7 1 .4082 .4082 - .8 1 6 5 .2687 .2887 .2887 0 0 0 0 0 O 0 O O O O O O 0 O O O O O O O O O 0 0 O O O O O O O O 0 O O O O O O O O O 0 O O O O O O O O O O O O O O 0 0 0 O O O 0 , O P O O O O O P O O 0 0 O O O O O O O O O O O 0 O O O O O O O P O O 0 0 0 O O O O O 0 P O O O O 0 0 0 0 Q 0 0 0 0 0 0 0 0 O .7071 -.7 0 7 1 O O O O O O O 0 0 0 0 0 0 0 0 O .4082 .4082 -.8 1 6 5 O O O O O P O O O 0 O O 0 0 0 0 Q 0 0 O .2887 .2887 .2887 - .8 6 6 0 O O O 0 O O O P O O O O 0 0 0 0 0 0 0 0 O O O O .7071 -.7 0 7 1 O O O O P O O O O O 0 0 0 0 0 0 P 0 O O O O .4082 .4082 -.8 1 6 5 O 0 0 O O 0 O O O 0 0 0 0 0 0 0 O O O O O .2887 .2887 .2887 - .8 6 6 0 d O O O O O O O 0 0 0 0 0 0 P 0 O O O O 0 O O P .7071 -.7 0 7 1 O O O O O O 0 0 0 0 0 0 0 O O P O O O O .4 0 8 2 .4082 -.8 1 6 5 O 0 O O O 0 0 0 0 0 0 0 O O O O 0 0 O O 0 O O .2887 .2887 .2887 - .8 6 6 0 P O O O 0 0 0 0 0 0 0 O O O O O O 0 0 O O 0 0 0 P P 0 O': O O O O O O O O O 0 0 0 0 0 O O O P P O O O O O O 0 - .8 6 6 0 0 O O 0 .7071 -.7 0 7 1 O O O P O .4082 .4082 - .8 1 6 5 O O O O P .2887 .2887 .2 8 8 7 - .8 6 6 0 Matrix R .7 0 7 1 0 .4082 - .8 1 6 5 .2 8 8 7 - .5 7 7 4 .7071 0 .4082 0 .2 8 8 7 0 - .8 6 6 0 0 .7 0 7 1 0 - .4 0 8 2 .4 0 8 2 - .2 8 8 7 - .5 7 7 4 .7 0 7 1 .7 0 7 1 - .8 1 6 5 .4 0 8 2 - .4 0 8 2 - .5 7 7 4 - .5 7 7 4 - .2 8 8 7 - .8 6 6 0 0 0 0 .8 1 6 5 .4 0 8 2 - .2 8 8 7 - .5 7 7 4 0 0 0 8165 0 0 0 2887 0 0 0 0 - .7 0 7 1 0 0 0 0 0 0 0 0 8165 0 0 0 2887 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - .8 1 6 5 0 0 0 0 0 .2 8 8 7 - .8 1 6 5 .4 0 8 2 .8 1 6 5 - .4 0 8 2 .4 0 8 2 - .2 8 8 7 - .5 7 7 4 - .5 7 7 4 - .2 8 8 7 - .2 8 8 7 - .5 7 7 4 0 0 0 .8 6 6 0 0 0 - .4 0 8 2 0 0 - .8 6 6 0 0 0 - .8 6 6 0 0 0 0 0 0 0 0 0 .8 1 6 5 0 0 0 0 0 - .2 8 8 7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - .7 0 7 1 .7 0 7 1 .7 0 7 1 - .7 0 7 1 0 0 - .4 0 8 2 .8 1 6 5 - .4 0 8 2 .4 0 8 2 .4 0 8 2 0 - .2 8 8 7 - .2 8 8 7 - .2 8 8 7 - .5 7 7 4 - .5 7 7 4 0 .7 0 7 1 .7 0 7 1 - .7 0 7 1 0 0 0 0 0 .7 0 7 1 - .7 0 7 1 0 .4 0 8 2 - .7 0 7 1 0 0 - .7 0 7 1 0 0 0 .7 0 7 1 - .5 7 7 4 - .8 6 6 0 0 0 - .5 7 7 4 .5 7 7 4 0 0 .8 1 6 5 .8 1 6 5 0 .2 8 8 7 - .2 8 8 7 0 0 0 0 - .2 8 8 7 0 0 0 - .5 7 7 4 - .2 8 8 7 .8 6 6 0 0 0 - .2 8 8 7 .5 7 7 4 .7 0 7 1 0 0 0 .8 1 6 5 0 0 0 .4082 0 0 - .8 1 6 5 .7 0 7 1 0 0 0 -.4 0 8 2 0 0 .4 0 8 2 0 0 0 - .8 6 6 0 .8 1 6 5 0 0 0 0 - .4 0 8 2 0 0 - .7 0 7 1 - .7 0 7 1 - .8 6 6 0 0 - .7 0 7 1 - .8 1 6 5 - .4 0 8 2 .8 1 6 5 .4 0 8 2 - .4 0 8 2 .4 0 8 2 0 0 .8 1 6 5 - .5 7 7 4 - .2 8 8 7 - .2 8 8 7 - .5 7 7 4 - .2 8 8 7 - .5 7 7 4 0 0 - .2 8 8 7 .8 6 6 0 .8 6 6 0 .8 6 6 0 - .8 6 6 0 .8 6 6 0 0 0 0 - .8 6 6 0 .8 6 6 0 Matrix (R 1R )"1 .4271 -.1510 - .1 5 1 0 -.0 1 5 6 2 -.0 1 5 6 2 .03124 -.04168 -.1510 .4271 - .1 5 1 0 -.0 1 5 6 2 .03124 -.0 1 5 6 2 -.04168 -.1 5 1 0 -.1 5 1 0 .4271 .03124 -.01562 -.0 1 5 6 2 -.04168 -.01562 -.01562 .03124 . 26 o4 -.06772 -.0 6 7 7 2 - .0 1 5 6 2 .03124 -.0 1 5 6 2 -.0 6 7 7 2 .2604 .03124 -.01562 -.0 1 5 6 2 -.06772 -.06772 - . 04 168 -.04168 -.04168 .1563 .1563 0 -.1 5 6 3 0 0 0 -.1 5 6 3 -.0 9 3 7 7 .1563 - .1 5 6 3 -.0 9 3 7 7 .2500 - .0 6 2 5 3 -.1 8 7 5 .02085 .2500 - .1 8 7 5 -.06253 0 -.06253 .2500 - .1 8 7 5 -.1 0 4 2 .02085 0 -.09377 0 .09377 -.1042 .02085 -.0 8 3 3 4 0 .1563 .1563 .1563 0 0 .2500 .2500 .1563 - .6 2 5 3 -.1 8 7 5 -.1 8 7 5 -.0 6 2 5 3 0 - .1 5 6 3 - .1 5 6 3 0 -.0 9 3 7 7 -.0 9 3 7 7 - .6 7 7 2 0 -09 3 7 7 0 .09377 .2604 0 .09377 .09377 0 0 .2084 0 0 .09377 0 .8750 .06252 .09377 0 .06252 0 0 .02085 -.1 0 4 2 .08334 0 .8750 .06252 .1250 • 3751 .8750 .06251 .06251 .3751 .1250 0 1.417 .08334 0 .1250 .3751 0 -.0 8 3 3 2 -.1042 -.1 0 4 2 .08334 .02085 -.06253 -.1 8 7 5 .2500 -.8 3 3 4 .02085 -.1042 0 0 0 .375 0 .1250 0 0 0 .1250 0 -.06253 .08334 .3751 0 .2500 .02085 -.0 6 2 5 2 - .0 8 3 3 4 -.1 8 7 5 -.1042 - .3 7 5 1 0 0 -.06253 .2500 -.1875 .02085 .08334 -.1 0 4 2 0 -.3 7 5 1 0 - .1 8 7 5 .2500 -.0 6 2 5 3 - .1 8 7 5 - .0 6 2 5 3 .2500 -.1 0 4 2 .08334 .08334 .02085 6 - .1 2 5 0 0 .02085 -.1042 0 0 -.3751 .1 2 5 0 .3751 -.0 8 3 3 2 -.3 3 3 4 - .3 3 3 4 -.3334 1.417 - .3 3 3 4 - .3 3 3 3 -.3334 -.1 2 5 0 .1250 - .3 3 3 4 -.3334 1.417 -.0 8 3 3 2 -.3333 .3751 - .3 3 3 4 -.3333 -.0 8 3 3 2 1.417 -.3 3 3 4 -.1 2 5 0 - .3 3 3 4 -.3334 -.3333 -.3 3 3 4 1.417 '96 Matrix R n ~ J- 0 ■.7071" .4082 - - . 8 1 6 5 .2887 -.5 7 7 4 .7071 .4 0 8 2 . .2887 ' 0 0 0 -.8 6 6 0 0 0 -.8 6 6 0 0 • -.8 1 6 5 ' -.5 7 7 4 .7071 4.4082 ■-.2887 .7 0 7 1 . .4082 -.5 7 7 4 .8165 .■ - . 7 0 7 1 T.4082 .5 7 7 4 -.2887 ■0 0 -.07071 .-.4082 -.8 1 6 5 .4082 .8660 -.2887 -.5 7 7 4 -.5 7 7 4 0 0 -.8 6 6 0 0 0 0 -.7071 .8165 0 -.4 0 8 2 -.8 6 6 0 .5774 0 0 -.2887 -.8165 -.7 0 7 1 -.4 0 8 2 . 8 6 6 0 .- . 5 7 7 4 -.2887 0 ... .7071 -.4 0 8 2 -.2887 .7071 0 .8165 -.2887 0 .8165 -.2887 .7071 .4082 -.5 7 7 4 .7071 -.4082 -;2887 0 .8165 -.2887 0 .8165 -.2887 .7071 -O4082 -.2887 .7071 ■ .4 0 8 2 -.5774 0 -.7 0 7 1 .8165 -.2887 -.5774 0 -.7 0 7 1 .8165 -.2887 .4082 .4082 -.5 7 7 4 .7071 .4082 -.7 0 7 1 -.5 7 7 4 -O5774 . .7071 -.4 0 8 2 -.7 0 7 1 -.2887 -.5 7 7 4 .4082 .4082 .7 0 7 1 ’ - . 7 0 7 1 .4082 .4082 -.5774 -.5 7 7 4 .7071 -.4082 -.2887 -.7 0 7 1 .4082 -.5 7 7 4 4 D378 E198 Hansen, James Leonard cop. 2 One at a time plans for 2^ factor sequencing designs ^AMt ANP AbowEBa »Lwz