第36卷第4期 2010年10月 曲阜师范大学学报 Journal of Qufu Normal University Vo1.36 No.4 Oct.2010 Minimum .aberrati0n Criteri0n for 2-1e e1 Fractional Factorial Designs .ZHAO Sheng,li①, LIU Xi0② (①School of Mathematical Sciences,Qufu Normal University,273165。Qufu; ②No.49 Middle School in Qingdao,266043,Qingdao,Shandong,PRC) Abstract:The lower order effects are more important than the higher order effects according to the effect hier- archy principle.Under such circumstances,minimum E.aberration(MEA)criterion based on the structures of the alias sets is suggested to rank 2…designs.Relations between this new criterion and other optimality criteriai.e., ,,the m ̄imum resolution,minimum aberration,clear effects and mDximum estimation capacity criteriaare studied. MEA designs with 16 and 32 runs are tabulated. Key words:Resolution;minimum aberration;clear;generalized alining pattern;estimation capacity CLC number:0212.6 Document code:A Article ID:1001-5337(2010)04-0001-08 1 Introduction Design of experiments has a long history of successful use in scientific investigations and industrial experi— ments.A full factorial design involves all possible treatment combinations.The number of such combinations grows rapidly as the number of factors increases.For obvious economic reasonsa full factorial experiment with large size ,may not be feasible.A practical solution is to choose a rfaction of the full factorial design for experimentationThen .how to choose the optimal fraction is what an experimenter concems.In order to draw valid statistical inference a. bout the factorial effects,the experimenter should select a“good”fractional factoria1 design.Many literature fo— cused on finding an optimality criterion for selecting optimal designs.The most populr craiteria include maximum resolution…,minimum aberration(MA) ,clear effects[ ]and maximum estimation capacitv[4・5]cfite1"‘1a.Muker. jee and Wu 。 gave a good summary and upgrade on these criteria. A regular 2 一 fractional factorial design d includes n factors,and it is determined by m independent defining words ̄The group formed by the m independent defining words is the defining contrast subgroup,denoted by G.Let A denote the number of -order effects in G.The vector W=(A,,A ,…,A ) (1) is called wordlength pattern(WIJP)of d.The resolution of d is defined to be the minimum i such that A ≠0.A design with the maximum resolution is said to be a maximum resolution designThe maximum resoluti0n criteri0n .selects designs with the maximum resolution as the optimal ones.Any design d with resolution II is not regarded as good because at least two of its main effects are fully confounded.Hence we only consider A for 3 as in(1)and the designs with resolution at least 111 in this paper. Suppose dl and d2 are two 2 designs and r is the smallest i such that Ai(d1)≠A (d2),then dl is said to have less aberration than dE ifA,(d1)<A,(d2).A design d is said to have MA if no other design has less aberra一 Received data:2010-02-09 Foundation item:This work was supposed by the NNSF of China(10826059,10901092),the NSF of Shandong Province of China(Q2007A05) and the China Postdoctoral Science Foundation(20090451292). Biography:ZHAO Sheng-li,male,1974一,PhD,Associate Professor;Major in:Mathematical Statistics.E-mail:zhaoshli758@126.c0m. 2 曲阜师范大学学报(自然科学版) 2010卑 tion than d.An MA design is regarded as the optimal one under the MA criterion. A main effect of a factor is said to be clear if it is not aliased with any main effects of the other factors or any two— factor interactions(2fi’s).A 2fi is said to be clear if it is not aliased itwh any main effects or any other 2fi’s.A clear main effect or clear 2fi can be estimated itwhout bias under the assumption of absence of three or higher order effects. Let E (d)be the number of models containing all the main effects and (1 sn n(n一1)/2)2fi’s which can be estimated by the design d.A design which maximized E (d)for all M is said to have maximum estimation capacity(MEC).The MEC criterion selects the designs with MEC as the optimal ones.Recent results on the MEC criteiron include[4,5,7]. Under the effect hierarchy principle,the lower order effects are more likely to be important than the higher or— der effects.A“good”optimality criterion should sequentially minimize the confounding between the lower order effects.In this paper,a new characterization,called generalized aliasing pattern(GAP),is proposed to rank 2 designs.Minimum E—aberration(MEA)criterion based on GAP is suggested to select‘‘good”2”一 designs.Sec・ tion 2 is devoted to establish the new criterion.Relations between this new criterion and those exsiting criteria are given in Section 3.MEA designs with 16 and 32 131ns are tabulated in Section 4. 2 Minimum E-aberration Criterion eLt d be a 2 ~design.We use 1,2,…,n to denote the factors in d.Let P n—m,then 2 一1 alias sets of d can be obtained from the defining contrast subgroup.Each of the 2 一1 alias sets contains 2 diferent effects which are confounded with each other and the defining contrast subgroup G contains 2 一1 effects(besides the grand mean,denoted by I)which are aliased with I. Zhang and Park proposed to use the following series C=(1C1,l C2,2C2,I C3,2C3,3C3,1 C4,2C4,3C4,4C4,…) (2) as a characterization to choose optimal designs,where Cl is the number of pairs of k・and/-order effects aliased with each other.Designs which sequentially minimize the components of C are more desirable.Note that Zhang and Parkt。]considered only the number of pairs of k-order and/-order effects aliased with each other but neglected the number of k-order effects aliased with Z—order effects and the number of Z.order effects aliased with k-order effects. Now let us also consider the aliasing of k—and/-order effects with k Z.Let A( .f) be the number of k-order effects which are confounded with/-order effects,A( 1),the number of/-order effects which ale confounded with k— .order effects,and A( f) the number of aliased pairs of k—and/-order effects.Note that the k-order effects are more ,important than the/-order effects under the hierarchical assumption.And the most important thing that the experi— menters concern is to estimate the effects,so the numbers of effects aliased with each other are more important than the number of aliased pairs.Hence a good design should minimize A( f)l,A( 1)2 and A( )3 sequentially. ,,, Example 1 To illustrate A( .,).(i=1,2,3)more clearly,let us consider the 2 一 design D determined by the defining contrast subgroup,=124=135=2345.Using 1,2,3,4,5,23 and 25 to multiply each elements of hte defining contrast subgroup,we get the following alias relations of D: 1 24 = 35 = 12345. 2 14 = 345 = 1235. 3 15 : 245 = 1234. 4 12 = 235 = 1345. 5 13 = 234 = 1245. 23 45 = 134 = 125. 25 34 = 145 = 123. 第4期 赵胜利,等:两水平部分因析设计的最小低阶效应混杂准则 3 Note that the 5 main effects,i.e.,1,2,3,4 and 5,are all confounded with 2fi’s,we get (12)l=5.Since there ,are 6 2fi’s,i.e.,24,35,14,15,12 and 13,which are all confounded with main effects,we have A(12)2=6. ,Furthermore。note that there ale 6 pairs,i.e.,(1,24),(1,35),(2,14),(3,15),(4,12)and(5,13),each pair representing an alias relation of 1一and 2-order effects,we can obtain A(12)3=6.Similarly,we can get A(2,,2)l =A(22)2=6 and A(22)3=3,…. ,,Define A( f)=(A(k,1)l,A(k,t)2,A( ,z)3)for f.Note that a(i,f)2=A(1, )3 f0 resolution at least 111 design and ,A( f)】=A( z)2 when k=1.To keep the symmetry and consistency of the defining 0fA( ,,,z)for different pairs of(k, Z),we still reserve three components for all k Z.Then the vector A( 。f)reflects the confounded degree of k-and f— order effects.The three components of A( f)contain not only the number of pairs of effects aliased with each other .but also the number of such effects. Since main effects are the most important effects under the hierarchical assumption and 2fi’s are the most im- portant interactions,the aliasing between main effects and 2fi’s is the most severe one,and then the aliasing be— ,tween 2fi’s follows.Hence a“good”design should minimize A(1ifrst.2) and then A(2.2).Following similar argu— ments,we define We:(A(12),A(22),A(13),A(23),A(33),A(14),A(24),A(34),A(4,.,,,,,,,4),…) (3) as GAP,where A( is placed ahead ofA( , )ifj<t,or i<s andj=t.A“good”design should minimize the en- tries of sequentilaly.Based on We,we define MEA designs as follows.The MEA criterion selects designs having MEA as the optimal ones. Definition 1 Let be the ith component of We,and let (d1)and We(d2)be the GAPs of two designs dl and d2 respectively.Suppose is the first component such that (d1)and (d2)are different from each other.If (dI)is smaller than (d2),then d1 is said to have less E-aberration than d2.A design d is said to have MEA if no other design has less E-aberration than d. Remark 1 Note that the series C defined in equation(2)depends only on and = ( 3.Hence C is just a part of defined in equation(3).On this account,We can be more discriminatory than the series C. 3 Relations with other Optimality Criteria This section is devoted to study relations of the MEA criterion with other optimality criteria. 3.1 Connections with Maximum Resolution and MA Criteria Resolution is one of the most important indexes used to describe the optimality of a design.A good design should have the maximum resolution under the maximum resolution criterion.The following theorem shows that the MEA designs have the maximum resolution. Theorem 1 An MEA design must be a maximum resolution design. Proof For given n and m.suppose the maximum resolution of the 2 一 designs is R.Then there exists at least a design,say dl,which has resolution R,and A( ,f)(d1)=(0,0,0)for k+Z<R by the defining ofA( .j). Thus dI has minimized A( f)for k+l<R.Therefore,the MEA design d must have A( ,.f)(d)=(0,0,0)for后+l< R.Thus the resolution of d is no less than R.By the assumption that the maximum resolution of the 2 一 designs is R,we know that the resolution of d equals R.The proof of Theorem 1 is completed. Isomorphism is an important relation between 2“一 designs.Two 2 一 designs are said to be isomorphic if one can be obtained from the other through row permutations,column permutations,or relabeling of levels.There may exist many nonisomorphic 2“一 designs with the same resolution for given n and m and the maximum resolution cri— teflon can not distinguish them.This is the limitation of the maximum resolution criterion.As the core of MEA cri— teflon,GAP can distinguish most of the nonisomorphic 2 ~designs according to the following discussion. 4 曲阜师范大学学报(自然科学版) 2010年 For a 2一 design d,let W(d)and We(d)be the WLP and GAP of d respectively.The relations between W(d)and (d)can be obtained from their definitions.For convenience,we define Al:A2=An+l=0. Theorem 2 The WLP W(d)and the GAP (d)have the following relation: A(1 =,(n—i+1)A。一I+(i+1)A +l (4) f0r 2≤i n. Proof For 2 i n,any/-order effect which is aliased with a main effect must correspond to an(i一1)一or (i+1).order effect which belongs to G.Note that each/-order effect in G contains i factors and each of the main effects of these factors is confounded with an( 一1)-order effect,and each of the main effects ofthe remaining n— i factors which are not contained in this/-order effect is confounded with an(i+1)一order effect,respectively.Thus (4)follows. Remark 2 From(4),one can find that there is a one—to-one correspondence between re(d)and(A(1.2),, A(1.3)2,…, (1㈤2).Thus (d1)= (d2)means W(d1)=W(d2)for any two designs d1 and d2.Ther efore designs GAP is a more detailed characterization of fractional factorial design than WLP. If two 2 designs are isomorphic,then they must have the same WLP and GAP.However,two 2 with the same WLP or GAP may be nonisomorphic.Two 2一 designs with the same WLP may have diferent GAPs. he fTollowing example gives two pairs of 2一 designs.The ifstr pair has the same WLP but diferent GAPs,and the second pair gives two nonisomorphic 2“一 designs with the same GAP. Example 2 The first pair includes the following two 2 designs: ,dll:,=126=137=238=12349:1235t0=45tl=12345t2; d12:,=126=137=248=349=125t0=135t1=145t2, where t0,tl,t2 represent the factors 10,11,12.The WLPs ofdl1 and dl2 are both W=(8,15,24,32,24,15,8,0, 0,1).The GAPs ofdIl and dl2 are (d。1)=((12,24,24),(66,66,45),(12,60,60),(66,212,456),…) and (dl2)=((12,24,24),(66,66,45),(12,60,6O),(66,204,456),…) respectively.Clearly,We(dl1)is different from (dI2).The second pair includes the following two 21 矗de— signs: d2l:,:127=138=149=25to=236tl=346t2:56t3=2456t4; d22:,=127=138=149=25to=236t1=346t2=56t3=2345t4, where to,…,t4 represent the factors 10,…,14.The designs d2l and d22 aye nonisomorphic .One can easily check up that d2l has the same GAP as d22.We omit the GAP ofd2I and d22 here for it is so lengthy. Since nonisomorphic 2 ~designs with the same resolution may not be equally good,the MA criterion based on WLP is often used to discriminate them.Most of the nonisomorphic 2 desinsg can be discriminated by WLP es— .pecially for designs with small runs.Although the MEA criterion still can not discriminate all of the nonisomorphic designs,it can further discriminate some nonisomorphic designs which the MA criterion can not discriminate,since GAP is more detailed than WLP. Remark 3 Zhang and Park obtained a generl faormula for calculating C with i as: ‘cJ= ( 一 二 +2 ’)( 一 z+2 At— + ,i,j=l,2,…,n, (5) where( )=1,( )=0 for <y or x<0,and A =0 ofr ≤2 or i>几・Since W(dl1)=W(dl2),0necan get that C(dl1)=C(dl2)by the formula(5),where dIl and dl2 are the first pair of desinsg in Example 2.By the results of Example 2,d11 and dl2 have different GAPs.Hence the GAP series C. contains more ifornmation of a desin tghan the 第4期 赵胜利,等:两水平部分因析设计的最小低阶效应混杂准则 5 3.2 Connection、 Ul Clear EtTects Criterion By the definition ofA( ,z)=(A( , )l,A( ,z)2,A( 。 )3),the smallest A( ,f)means the most slight confounding of k— and Z—order effects.Note that for k=1 and l=2,A(1.2)l just equals the number of non—clear main effects.There ex- ist resolutionⅣdesigns when n 2 。.Then the MEA design d must satisfyAf1.2)(d)=(0,0,0)ifn 2 ,and the main effects of d are a11 c1ear.H0wever.any 2 一 design with resolutionⅣhas no clear 2fi if rt>2p一 -4-1[ . Note that A(212) is just the number of non—clear 2fi’s for resolutionⅣdesigns,thus A(212) (d)= if a resolutionⅣ2 design and n>2 +1. When n 2 ~4-1.we can show that the MEA design must be ODe of the designs with the maximum number of clear 2fi’s. Lemma 1 Suppose dl and d2 are two 2一 designs with resolutionⅣ,then dl has less E-aberration than d2 ifd1 has more clear 2fi’s than d2. Proof For a resolutionⅣdesign d,A(1_2)(d)=(0,0,0),and the number of clear 2fi’s of d is 芝_= ・If dt has m。re clear 2fi’s than d2,then (2,2) (d1)<A(2,2)1(d2)・The result foll。w s. Theorem 3 When n 2 +1.the MEA 2 design must be one of the resolution at least 1V designs with the maximum number of c1ear 2fi’s. ’ Proof For given n and m,ifthere exists a resolution at least V 2 ~design d,then A(1,2)(d)=A(2-2)(d)= (0,0,0)and the 2fi’s of d are all clear.The MEA 2一 designs must have the most clear 2fi’s.If there does not exist a resolution at least V 2…design,the result follows from Lemma 1. - Example 3 Let us consider two 2 designs.Let d1 be the design determined by I=1236=1247=1258= 13459 and d2 be the design determined by,=1236=1247=1348=23459.dl has MA and d2 has the maximum number of clear 2fi’s in resolutionⅣ2 -4 designs .The number of clear 2fi’s in d1 and d,are 8 and 15,re- spectively.And we can easily get that A(1,2)(d1)=A(1,2)(d2)=(0,0,0),A(2,2)l(d1)=28 and A(2,2) (d2):21. Thus d2 has less E—aberration than d1.So d2 is better than d1 under the MEA criterion. As an optimal criterion,the clear effects criterion cannot be used for many parameters.For example,there ex- ist resolutionⅢdesigns when >2 ~.but these designs do not contain any clear main effect or clear 2fi.Hence the clear effects criterion cannot select the“best”one from them.Thus the clear effects criterion cannot be used when n>2 _。.However,the MEA criterion can be used then.When 2 一 +1<l凡≤2 -。,resolutionⅣdesigns ex- ist but they make no difference under the clear effects criterion.However,the MEA criterion can discriminate them further. 3.3 Connection、】lrith MEC Criterion Now let us consider the relation between the MEC criterion and the MEA criterion. Suppose il…i andJl…Jz are two diferent effects,i1…i is said to be“smaller”thanJl…Jz ifk<l or ifk= Z and il…i should be listed ahead ofj1…Jz lexicographically.If i1…i is“smaller”than j1…J ,it will be written as i1…i Jl…J1.For a given alias set,the“smallest”effect in the alias set is referred to as the alias set leader. Applying to the alias set leaders.the alias sets can be ranked from the“smallest”to the“largest”with the “smallest”alias set G receiving rank 0 and the“ largest ”receiving rank 2 一1. For a 2 ~design of resolution llI or higher,let m denote the number of 2fi’s in the ith alias sets and m= (m +1,…,m +_,),wheref=2 一1一 .A vector M=(Ml,…,M )is said to be upper weakly majorized by = t t ( 1,…,Vs)if∑uⅢ i=l ∑ Ⅲfi=1 or 1 £ss,where u…sM …sn and Y Ell … are the or. dered components of u and respectively.A design dl is said to dominate d2 if E (d1) E (d2)for all and with strict inequality for at least one u.A sufifcient condition for d1 dominating d2 is given in the following lemma by T 6 曲阜师范大学学报(自然科学版)S U 2010年 Cheng,Steinberg and Sun Lemma 2 If m(d1)is upper weakly majorized by m(d2)and m(dj)cannot be obtained from m(d2)by permuting its components,then di dominates d2 with respect to the criterion of estimation capacity. n+, Lemma 2 shows that a design d will behave well under the MEC cirterion if∑ i=n+1 (d)is large and mn+l(d),…,m +r(d)are close to one another. Now let us consider design d with resol ution 1V.Since A(1_2)(d) =(0,0,0),we have n+f ∑一n+| n | , 一翌(翌= ! —。 m 一一 2 。 + 2A(2':) (n(n一1) ∑一d)=∑m m 一1)=∑肌 一 i=n+1 i=n+1 2 Henee for resoluti。n IV designs with gim ,ven ( 22)l,the smallest (22)3 means the closest of mn+i,m,,n+2,…,m +, t。。He another・Then d1 tends to behave welm > l than d2 with respeet to MEC criterion ifA(2,2)1(d1)=A(2,2) (d2)and A(2,2)3(d1)<A(2,2)3(d2).Note that A(= 2,2)1(d)is the number of non—clear 2fi’s of resolution IV design d.The MEA criterion tends to select the designs with the maximum estimation capacity among those with the maximum number of clear 2fi’s as the optimal ones when n<2 ~∑一 m +1. 4 16-and 32-run MEA Designs Let a1,u2,03,a4 and a5 denote the five independent columns(10000) ,(01000) ,(OOLOO) ,(o0o10) and(00001) ,respectively.Then any sum mod 2 of a1,u2,Ⅱ3,a4 and a5 also co ̄esponds to a binary sequence, for example a1+u3+u5(mod 2)corresponds to(10101) .Table 1 converts these binary sequences into base..ten system.A 2 ~design can be obtained by selecting a subset of凡columns of S.For 16.run designs.S consists of the first 4 rows and 15 columns.For 32一run designs,S is the whole matrix.Independent columns are numbered I, 2,4,8 and 16. For brevity,we use n—m to denote the MEA 2 一design in Tables 2 and 3.To save space。the independent columns l,2,4 and 8 in Table 2 and I,2,4,8 and 16 in Table 3 are omitted.Designs for n=2 一i,i:I,2, 3 are unique up to isomorphism and hence omitted in both Table 2 and Table 3.Note that A(1, )2=AiI, ),for resolu— tion at least III design and A(22),=A(2,,2),.Only four entries of GAPs are given in Tables 2 and 3,i.e.,GAPs= (4(12)I,A(12)2,4(22)l,A(2,,,,2)3). Table 1 Design matrices for 16-and 32-run designs 第4期 赵胜利,等:两水平部分因析设计的最小低阶效应混杂准则 Table 2 16-run MEA designs withn=6,…,12 columns Table 3 32・run MEA designs tll n=7,…,28 colmuns 7 r}rL 1 2 3 1{4 1J 曲阜师范大学学报(自然科学版)5 1J rL 2010卑 References: Box G E P,Hunter J S.The 2 一 fractional factorila designs[J].Technometrics,1961,3:311.351.449-458. Fries A,Hunter W G.Minimum aberration 2 designs[J].Technometrics,1980,22:601-608. Wu C F J,Chen Y.A graph—aided method for planning two—level experiments when certain interactions are important[J].Techn ometrics,1992,34:162—175. Cheng C S,Mukerjee R.Regular fractional factorial designs with minimum aberation and maximum estimation capacity[J].Ann Statist,1998,26:2289—2300. Cheng C S,Steinberg D M,Sun,D X.Minimum aberration and model robustness for two.1evel factorial designs『J].J Roy Stat. ist Soc Ser B,1999,61:85—93. [6] Mukerjee R,Wu C F J. A modem theory of factoiral designs[M]. Springer Series in Statistics.Springer Science+Business Media Inc,New York,2006. [7] Ai M Y,Zhang R C. Muhistratum fractional factorial split—plot designs with minimum aberration aT1d maximum estimation capaeitv [J].Statist Probab Lett,2004,69:161 170. [8]Zhang R C,Park D K.Optimal blocking of two—level fractional factorial designs[J].J Statist Plann Inference,2000,91:107 121. [9]Zhu Y,Zeng P.On the coset pattern matrices and minimum M—aberration of2“一 designs[J].Statist Sinica,2005,15:717.730. [10]Chen H,Hedayat A S.2…designs with resolutionⅢand 1V containing clear two—factor interactions[J].J Statist Pl舢I r- ence,1998,75:147—158. [11]Wu H Q,Wu C F J.Clear two—factor interaction and minimum aberration[J].Ann Statist,2002,30:1496.1511. [12]Zhao S L,Liu Y L.Some results on 3-level factorial designs[J].Journal of Qufu Normal University,2008,34(1):1-5. 两水平部分因析设计的最dq¥阶效应混杂准则 赵胜利①, 刘 霞② (①曲阜师范大学数学科学学院,273165,曲阜市;②青岛市第49中学,266043,山东省青岛市) 摘要:根据效应分层原理,低阶效应比高阶效应更重要.在这种情况下,以别名集的结构为基础,提出了最小低阶效应混 杂准则对2.-m设计进行排序.研究了新准则与其它最优性准则,即最大分辨度、最小低阶混杂、纯净效应和最大估计容量准则 的联系.给出了具有16和32个水平组合的最小低阶效应混杂设计. 关键词:分辨度;最小低阶混杂;纯净;广义别名型;估计容量 中图分类号:O212.6 文献标识码:A 文章编号:1001-5337(2010)04-0001-08