UNCLASSIFIED11 II RANDOMIZED FRACTIONAL REPLICATION DESIGNS Ii by S. Zacks TECHNICAL REPORT NO. 86...
Transcript of UNCLASSIFIED11 II RANDOMIZED FRACTIONAL REPLICATION DESIGNS Ii by S. Zacks TECHNICAL REPORT NO. 86...
UNCLASSIFIED
ARMED SERVICES TECICAL INFORMAMARLIIGEN HALL STATIWMUM 12, VIRGINIA
UNCLASSIFIED
3INOT When goveinunt or other drawings., sped-ficaticias or other data ane used for any purposeother than In connection with a dafinitely relatedgovernunt prcrntOperation, the U. 8.
Goverinmnt threby Incurs no responsibility, nor anyobligstion mbatsoever) and the fact that the Govern-ment my have formilasted, furnished, or in ay veysupplied the said drawings, specifications, or otherdata is not to be regarded by implication or other-vise as in any manner licensing the holder or anyother person or corporation., or conveying any rightsor permission to mmanfactzre,, use or sell anypatented Invention that my In any vey be relatedthereto.
APPLIED MATHEMATICS AND STATISTICS LADORATORIES
STANFORD UNIVERSITYCALIFORNIA
GENERALIZED LEAST SQUARES ESTIMATORS FOR
RANDOMIZED FRACTIONAL REPLICATION DESIGNS
BYS. ZACKS
TECHNICAL REPORT NO. 86March 15, 1963
PREPARED UNDER CONTRACT Norw-225(52)(NR-34 2-022)
FOROFFICE OF NAVAL RESEARCH
(
11
IIRANDOMIZED FRACTIONAL REPLICATION DESIGNS
Ii byS. Zacks
TECHNICAL REPORT NO. 86
March 15, 1963
i I PREPARED UNDER CONTRACT Nonr-225(52)(M-342-022)II: ASTIA
IFOR 1-7r7,'nm' 7I OFFICE OF NAVAL ]RESEARCH APR 10 1963
J ~J L LaTISIA
Reproduction in Whole or in Part is Permitted forany Purpose of the United States Government
APPLIED MATHIEATICS AND STATISTICS LABORATORIES
STANFORD UNIVERSITYI
STAI•iORD, CALIFORNIAII
I
1 Generalized Least Squares Estimators
ifor Randomized Fractional Replication Designs
by
I S. Zacks
1. Introduction
Fractional replication designs have become of great importance,
I especially for industrial experimentation. A missile whose operation
is affected simultaneously by dozens of interacting factors would pro-
duce a full factorial experiment of impractical size. Indeed, if there
are more than 20 factors which may affect the operation of a missile,
and if we like to attain complete information on all the main effects
and interactions of the controllable factors we would run the experi-
20ments over more than 2 = 1,048,576 treatment combinations. Fractional
replication designs are planned to attain information about some of
the main effects and interactions by a relatively small number of trials.
If the operation of a missle can be controlled with some information on
I the main effects and some low order linear interations, it might be
I sufficient to run only 32 or 64 trials at a time. These however, should
be chosen from those possible in some optimal manner.
f The problem of choosing a 1/ 2 m-s fractional replication and an
appropriate estimator of the parameters characterizing the factorial
model (main effects and interactions) has been studied by A. P. Dempster
(1960, 1961), K. Takeuchi (1961), S. Fhrenfeld and S. Zacks (1961,
1962) , S. Zacks (1962), B. V. Shah and 0. Kempthorne (1962a,b). In
all these studies the type of estimators considered is that which yields,
under a randomized procedure with equal probabilities of choice, unbiased
I
estimates of a specified linear functional of a subvector of parameters,
which lies in the range of the design matrix (the matrix of the corres-
ponding normal equations).
In the present study statistical properties of the generalized
least squares estimators, under randomized fractional replication de-
signs, are studied. The term generalized least-squares estimators (de-
noted henceforth by g.l.s.e.) is used since the matrices of the nor-
mal equations corresponding to these designs are singular. The fac-
torial models corresponding to the type of fractional replication de-
signs studied in the present paper is presented in section 2. For this
sake we start from the factorial model for a full factorial system.
Then we present the required algebra, and the method of constructing
the orthogonal fractional replications. The linear spaces of all g.l.s.e.
associated with the various orthogonal fractional replication designs
are characterized in terms of the linear coefficients of the corres-
ponding factorial models. Some statistical properties of the g.l.s.e.
under procedures of choosing a fractional replication at random, are
then studied. First we prove that there is no g.l.s.e. which yields
unbiased estimates of the entire vector of in parameters. However,
there are g.l.s.e.s which estimate unbiasedly subvectors of parameters.
The trace of the mean-square-error matrix corresponding to a g.l.s.e.
applied under certain randomization procedure is used as a loss
function for the decision problem of choosing a g.l.s.e. and a ran-
domization procedure. It is shown that when the parameters of the
factorial system may assume arbitrary values, the randomization pro-
cedure which assigns equal probabilities to various fractional
2
I
replications (denoted by R. P.*) is admissible. Bayes g.l.s.e., relative
to a-priori information available on the parameters, are then studied.
This leads to a minimax theorem, which specifies a minimax and admis-
sible g.l.s.e, under R.P.o*
The relationship between the generalized inverse of the matrix of
normal equations and g.l.s.e. as given by A. Ben-Israel and J. Wersen
(1962), and by C. R. Rao (1962) is studied. It is shown that these
are particular cases in the general class of g.l.s.e. studied presently.
Finally, it should be remarked that although the present paper
deals with factorial system of order 2m all the important results
hold in more general factorial systems of order pm (p > 2).
3
2. The statistical model for fractional replication designs.
2.a. The statistical model for a full factorial experiment
of order 2 m
A full factorial experiment of order 2m is a set of 2m treat-
ment combinations, consisting of m factors X0 ,..., X each at
two levels. Such a system can be characterized by 2m parameters
ao,..., , which are the coefficients of the multilinear re-2 m1l
gression function:
(2.1) E(Y(Xo,...,Xm.i ) X 0 u(, Xo Xl.. " m" 1
where X 0,1 (j=0,...,m-l); Uxo~,...,x 1 ) 2J;S 0 ----
and Y(Xo,...,X m_1) is a random variable representing the "yield" of
the experiment at treatment combination (X ,...,X M_). Denote by
X ,0 and X ,l (j=O,...,m-l), XJ, 0 < Xj,I, the two specified levels
of factor X . By changing variables according to the transformation
Xj~ i (Xj~ + XI(2.2) ZJ,k 1Jk I (x- 2 Jil (k=0,l;j--O,...,m-l)
f (xil XJ, 0 )
the regression function (2.1) is reduced to the form:
(2.3) E(Y(Zo,.. z )) = Pu Z 0 Z1 " m-Io _ 1m u--O u M-
where Z =1-, +1 (J=O,...,m-l); and u=U(o,...,X )
j 'm-4
I
Writing Zj = (-1) J with i j= 0,1 for all J = 0,...,m-1 , the
regression function (2.3) can be represented in the form:
M-12l M_ 1 E % (1-i )
(2.4) E(Y(i 0 ,..,i m-1 = ou(-l ) J=Ou=O
rm-1Furthermore, denote by xv (1.i M-ir ), v I . 2J , the
J=O2 m treatment combinations of the factorial system under considera-
tion; and let
m-1
(2.5) c( 2m) (-l)j=O , for all v,u = 0,..., 2 m 1vu/
then (2.4) is reduced to the form:
2m-1 (2m) m(2.5) E[Y(x)) = c ,u for all v = O,
Let Y' = (Y(X 0 ),...,Y(x .)) be the vector of observations at all
the 2m treatment combinations; and let 1' = ( PO 2m-l) be the
vector of parameters of (2.5). Thus, if
(C() = cm) , (vu = ,...,2 M),
denotes the matrix of the coefficients of the P's in (2,5), then
the statistical model for a full factorial system can be written as:
(2.6) Y = (C (2m)) P + C
where E is a random vector, with E c = 0 and E CE' = a2 1 ( 2m)
(In) denoting the identity matrix of order n).
5
!
2.b. The algebra of factorial experiments.
2.b.l. Properties of the matrices (C(2m))
The properties of the matrices (C (2m)) will be presented without
proofs. For details see S. Ehrenfeld and S. Zacks (1961).
The matrices (C (2m)) , m = 1,2,..., of (2.6) can be obtained
recursively by the following relationship:
(2.7) (C() = -1f (c(2) ,- m = 1,2,....
(where (C(0)) 1 (scalar), and where A®B is the Kronecker's
direct multiplication of the matrix A by the matrix B from the
left, defined as follows: If A is an n x m matrix ilaiji1, and
B is a k x 1 matrix lb rsi, then A(B is the nk x ml matrix:
• I o
anl B a 'a BI " I a*
I I
From the associative property of the Kronecker's direct multiplica-
tion it follows that every matrix (C(2m)) can be factorized into
2m'sx 2m-s (1 < s < m) submatrices of order 2s x 2s , according to
the relationship:
(2.8) (C(2m)) = (C(2m-s)) (C(2)) , (1 < s < M)
6
1From this relationship it follows that the elements of (C (2m)) are
related to those of (C ( 2 m-s) and (C(2s)) according to:
(2.9) C (2m)o =C (2 ) c(2m's)(.)i+J2It irt Jqt
for all i = 0,...,2 -1; j = 0,...,2 -1; and where t = rt + t2r
(r t = 0,...2S8-l ; qt = Op'"".2m-'sl) .
Another useful relationship among the elements of (C ( 2 m)) is
the following one: Let ulu 2 = 0,...,2 M- be given by
Ui (%oi, X i,..., I(ml)i) (i=1,2); 0j,1 = O,1 (J=O,...,m-1), and
define u u (o' Q' where Xm X + X (mod. 2);1 2 o M- 1 J 1 .Jl 2
then
(2.10) C(2m) = ( 2 m) ' (2m) for every v=O,...,2m-iVU1 U2 v1u1 v'u 2
The properties of (C(2)) are extended into (C(2m)) by the
recursion relationship (2.7), and are summarized as follows:
(i) C(2) = 1 for every v =v,0
(ii) C(2m) 1 for every u = 0,...,2m-12m-l,u
01 (2m)m(iii) I C = 0 for every v = 0,...,2 -2u=O v
u011• (2m)_m(iv) I C - 0 for every u = l,...,2m-i for every
v=O m
U = 1,...,2-1.
!
2(-) 7 (2m) c(2m) 2M)ifU(v) E~ fv if vv=O l u2 0 , if u 1 u 2
and
(2m) f2m if v1 'v2( 2T) 7 M Vlu vM -
(vi) u--O 2 0, if v 1 v2
(v) and (vi) can be expressed also in the form:
(C(2m)) (c(2)) - (c(2m))(c(2m)) = 2m(2)
2.b.2. The group of parameters.
Every parameter Pu (u = 0 ,.o., 2 m-1 ) of the statistical model
(2.6) can be represented by an m-tuple Pu = (ko' ,''X M-l) where
k = 0,U (J = O,.,m-1)
The set of all 2m parameters constitutes a group, B with
respect to the operator Q, defined as follows:
m-1 m-1Let u X k' 2j and u 2 = • X•21 then k
j=0 j=O 1 2
if, and only if k = 1 where V X + X' (mod. 2) forJ=O
all j = 0,...,m-l.
The unit element of the group B is P - (0,0,...,0) and the
1
inverse of M- (Xo. ., i.) is - (2-Xo,2.XlI 2-% l)(mod. 2)."" -l oU )•* , -, iu. ,
A set of n parameters uI )u 2'...'"u is called dependent if there2 n
I
exist n constants a k (k-l,...,n), not all of which are zero
(ak = 0,1), such that:
(2.11) [u11 a 2 2 [Un = poU1 a12 U 0
In
O , if a = 1where whr , if a = 0
If relationship (2.11) is valid only when all ak = 0 (k=l,...,n)
then Ul'''"Un are called independent. It is easy to check that
every set of n independent parameters (1 < n < m) generates a sub-
group of order 2 in B.
2.c. The construction of a 1 / 2 m-s (1 < s , m) fractional
replication.
Let ( '0dod 1 Pdm-s1 ) be any set of m-s independent
parameters in B. The 2 m treatment combinations can be classified
into 2m's disjoint subsets Xv (v=O,...,2m-s-l) of equal size, re-
lative to the specified m-s independent parameters, in the following
manner: Let Pdk E (X od kIXdk,..., XMlid k = 0,...,m-s-l be
a defining parameter, and let X a (io,...,i M.) be a treatment com-
bination, then x E X if, and only if,
m-1(2.12)O j 'j, ak (mod. 2), and where
-s-i ak 2
k=O
I
In order to perform the classification of treatment combinations into
the blocks X (v = 0 ,..., 2 ms- 1 ) we do not have to solve the linear
equations (2.12), but it suffices to compare the rows of the matrix
of coefficients (C( 2 m)) under the columns corresponding to the
special independent parameters (0d '0''"d ). These two procedureso mn-s-i
are equivalent (see S. Khrenfeld and S. Zacks (1961)). The m-s inde-
pendent parameters, relative to which the classification takes place,
are called defining parameters. The answer to the question, which of
the parameters should be specified for the role of defining ones de-
pends on the objectives of the experiment. The choice of a set of de-
fining parameters will generally effect the bias of estimators and
their variances, and might have other effects on the properties of
statistics and procedures (see 0. Kempthorne (1952); S. Ehrenfeld and
S. Zacks (1961).
The term fractional replication, in its broadest sense, relates
to any subset of treatment combination from a full factorial system.
K. Talnkuchi (1961) considers designs of randomly combined fractional
replications. We shall consider in the present paper only fractional
replications which consist of one block of treatment combinations, Xv,
chosen from the set of 2ms blocks constructed according to the pro-
cedure outlined above. These fractional replications are called
orthogonal. A randomized fractional replication procedure is one in
which a block Xv is chosen with a probability vector ý' = ( .toA 2m-s 1 )
2.d. The statistical iodol for a 1/ 2 m's fractional replication.
Let P 'd 0* d*']3d ) be a set of defining parameters;2m r-s I
(Xv ; v = 0,..., 2Ms- 1) the corresponding blocks of treatment combinations,
10
I
and Y(Xv) the vector of observations associated with the 2s treatment
combinations in Xv. The order of the components of Y(Xv) is deter-
mined by the standard order of the corresponding x's in Xv, e.g. if
Xv = (Xo, x5, xs, x6 ) then Y(Xv)' = (Y(xo), Y(x3), Y(x5 ), Y(x6 )).
Let P(0)= (Po' 1. t2sI) be any specified vector of 2s
parameters independent of the defining parameters (except for the "mean"
Po); with t < t for all k = l,...,2 S-1. Let (1* ; u=O,...,2m'-i)o k k+1 u
be the subgroup of 2m-s parameters, generated by the m-s defining
parameters. Define by P (u) (u = 1,...,2m's-l) the vector of 2s
parameters obtained by multiplying each of the components of 0(0) by
P, i.e., 0(u) = (Po(D , P t ® 3),... t2l@ ) . Then, the
statistical model for Y(Xv) can be written in the form:
2m-s.-
(2.13) Y(X) = v (p( 2S)) P + E = (p )P* + Cu=O
where; as proven by S. Ehrenfeld and S. Zacks (1961)
=( 2 -s " " b ( 2 M -s ) ) ( : O- 2 s )(2.lh) 1 (i, bvl ,.M_ v( 2m!l)
is a 2 sx( 2m- 2 s) matrix; (P(2)) is a 2s x 2s matrix obtainedvO
from (C(2m)), by picking the elements of (C(2m)) corresponding to
treatment combinations in X and the parameters in P(0) andV
arranging them in the standard order. The scalars b(2), by which-•(2s)p(2s)vu
we multiply (P 2)) to obtain (P (2) are given by the formula:vo vum-s-i
( 2 . 5 )( ) ( i 5 ( i j - L ( d j ) )
(2.15) b( u (-1)J:°
11
Irn-B-i mn-s-i m r-i j
where v = 1 ij2J;u= u iý 2j and L(dj) X ,d (mod. 2),
J=o Jýo k=O '
for every defining parameter and where.
and E is a random vector of order 2s, with Ec = 0 and EEE' a c2 1(2s).
(p_(2s))It can be readily proved that, the rows of ( vu ) (v'u--O'''"2')'
as well as its columns, are orthogonal, i.e.,
(2.16) (P(u) (P(2) = u(2s))((2s)) = 26 2
for all v,u=O,...,2m-l; and that a similar property holds for the
matrix (Bd "m))' whose elements are the coefficients b (2m-s) u,dS.. .,d S-)
defined by (2.15), i.e., j
(2.17) (B do,.? .,dm-s)l))t (B(do,...,dns~ i)) = 2m-s I( 2 mrs)
for every choice of m-s defining parameters. For the sake of simpli-
fying notation, let S - 2s and M = 2m-s, N = S-M = 2 m. Furthermore,
assume, without loss of generality, that the defining parameters are thek(o ) ' s l
"main effects" (PSP2S ,."'N/2) and that P = ()•, P,.
then, the blocks of treatment combinations are:
(2.18) (Xi+vS ; i = 0,...,S-1) for all v = 0,..., M-I
and the statistical model for Y(Xv) is given by:
(2.19) Y(Xv) = [(i, C(M) C(M) )D (C (S) ]Pvvi "'" v(M-1)(D((S)
(c(S))0(0) + M-1 0(M) P(u) + EU=l
where P (u) =(pus' Pl+us''" P(u+l)s-i)'
12
3. Generalized least squares estimators for fractional rplications.
t 3.a. The set of all least squares estimators.
jGiven a block of treatment combinations X (v=O,...,M-1) and the
associated vector of observations Y(Xv), the "normal equations" cor-
responding to the linear model (2.19) are given by:
(3.1) (C_)', tCv) = (Cv)' Y(XV) I v=O,...,M-I
where (Cv) is the S x N matrix of the coefficients of (2.19), i.e.,
(5.2) (C ) = t (, ) c(M) ) ( s)v vk '"'" v(M-l)
A generalized least squares estimator (g.l.s.e.) of P is any
linear operator (Lv), on E(S) (Eucliden S-space), so that (Lv)
is an N x S matrix satisfying the equation:
(353) (C V)' (Cv)(Lv) = (Cv )' , v=O,...,M-l
Let (Lv)' - ((Lyo)' (Lv)' (L(Ml))') where (L)
(u=O,...,M-l) are square matrices of order S x S. Substituting
from (3.2) for (Cv) in 3.3 and decomposing (Lv) as indicated
we arrive at the matrix equation:
(Lvo) 1
(3.4) s[(Q(M) )®(D(IS) ".. ........ (S)),
(L ) C(M)Lv(M-1) Lv(m-1)
where (Q(M)) (l,C(M.) C(M) ), (lC(M) C(M) is aS 'vl'" v(M-1) vi "'' v(M-I)
13
square symmetric matrix of order M x M, whose (i,J)-th element is
qM(M) M) C(M) (iJ=O,...,M-l). Since C(M) = + 1 and (C(S)qij = vi vj vu --
is non-singular, the linear equations in the matrices (L vu) can be
expressed in a form equivalent to (5.4) as,
M-( (M) (c(S)) I(s)
u=O
Since the unique solution to the equation (H)(C(S)) I I(s) is
(H) =A (C (S) it follows that the M matrices ( v ), J=0,...,S
M-l, whose submatrices are given by:
1 c(M) (c(S)), if u = j-l(3.6) S vuIvu (0) , otherwise
constitute a basis of M independent solutions of (3.3). Thus,
every g.l.s.e. (L ) can be represented as a linear combination of
the M linearly independent operators (L(J) ) (J=O,...,M-l), so
that the coefficients of (LvJ)) add up to 1. Formally, the set of
all g.l.s.e., given X is:v
M-1 Lj -(.7) •(C) = (L") : (Lv) = Ml M-j ) =
1 =0 j =o
Every g.l.s.e. can thus be represented by M coordinatesM-1 A
(Xo, 1%,...,X . 1 ) such that 2 Xj = 1. Furthermore, if P v de-J=O
notes the vector of g.l.s.e. of P then we have, according
to (5.6) and (3.7)
14
!
X 0 C(M)
(3.8) v 1 c vu (Q (c (SY)) Y(Xv) for every v=O,...,M-l.C (M-1)SM-1 vu
S3.b. Som__.e statistic-al properties of glse
In the present section we prove that there are no unbiased g.l.s.e.
of P, and derive an expression for the trace of the mean-square-error
Amatrix of a g.l.s.e. P.
Consider a fractional replication design in which a block Xv
(v=O,...,M-l) is chosen with probability tv (tv > 0 for allM-1
v=O,...,M-l; 2 tv = 1). A randomized fractional replication pro-v=O
cedure is thus represented by an M-dimensional probability vector t.
This class of randomization procedures contains, in particular, the
fixed fractional replication design, in which one of the X blocksv
is chosen with probability one.
Theorem 1.
Let (u) = C(M) (C(S))' Y(X) then E (•(u)) P •(u) for
all u=O,...,M-l if, and only if, E* = , 1
Proof.
The expected value of ý(u) under randomization procedure t
is given by:
15
I
(3-9) E• (P (u)) 1 t •v C S Y(X)
' M- 1 (M) (S) Lw c(M) (c(S))P(w)I tv C'U (C) IL + (c)(
-1 M-1 cC (M) C(M) = •(u) + M1 M-1
w=O v=O vu w0 (w,)= O t v
• (M) C(M) (w)'vu! vV
Clearly, if tv = 1 for all v=O,...,M-l, thenMM-I
M-1 c(M) (M) i 1 c(M) c(M)
v=O v=O
for all u / w by the orthogonality of the column vectors of (C(M)).
Thus, E. ((u)) = (u) for all u=O,...,M-l. On the other hand,
if Et (•(u)) = 1(u) for all u--O,...,M-I then, in particular
(3.10) E() P() M-l M-1 '(M) P (w)(3.zo)E•{•(O I I t(o ÷ v c
w=l v=0O
Bu te onitonM-l •(M)B e1 C = 0 for all w=l,...,M-I is equivalent
v= VW
to the condition:
Multiplying both sides of (3.11) by (C(M)) we get:
16
I
(3.12) Mt = (c(M))(c(M))I t =(C(M))(:)= 1 (M)
where 1 (M) is an M-dimensional vector with unity in all its components.
It follows that a necessary condition for the unbiasedness of (u is
that (M) i.e., each block is chosen with the same probability.M
(Q.E.D.)
Returning to the g.l.s.e. we have:
(3.13) E {A = Et'(( 0 p (O)' X lP)))
1 X(M) X 1 ) '
where t* = 1 = ( ,M
M-1Since Z X = 1 we conclude that there is no unbiased g.l.s.e, of P.
u=O
The g.l.s.e. in which X 0 = 1 and X = 0 for all u > 0 yields0 u
unbiased estimates of the components of P(0) only. Similarly whenX= i (J:O,...,M-1) and Xj, = 0 for all j' i J, the corresponding
g.l.s.e. yields unbiased estimates of the components of ' (W only.
AThe mean-square-error dispersion matrix of a g.l.s.e. P , under
randomization procedure t, is defined by E . Let
M(f,X;P) denote the trace of the mean-square-error dispersion matrix
of a g.l.s.e. represented by a vector X, such that V'1 (M) = 1, under
randomization procedure • ; i.e., M(t,X;P) = EL((•-$)' (t- )P
17
Theorem 2. 1The trace of the mean-square-error dispersion matrix under ran-
domization procedure t is given by the expression:
2M112 X M-1 •(u) E2(3.14) M(t,X P) = ( + iZ) (2- u-I)l +
u=O u=OM-1 (-1 M-1 (M) (M) (u1)' (u2)
+uI =07 (2X u +1) (uuI__ L v-° CvuI C2 0
where i1i 2 = Pp, (u)12 = P(u),p(u) (u=O,...,M-1)
Proof. I
According to (3.8)
M-1 2(3.16) E• 03'P) 2 E _X u Y(Xv)
Substituting (2.19) for Y(Xv) in (3.16) we get:
1 1 M-1 (S) (ul)
(3.17) E (Y(X)' Y(x)) = is E[[ Z (C )PS t v u! =0 V
M-1 C ) (C(S)) P(u2)+ Eu2 =0 vu2
2 M-1 M-1 (u 1 (u2)2 + E( I C) c(M) (u()' ,C() ) (
ul=O u=0 vu 2
a 2 + M•1 p(u)12 + Mu1 t u (M) C(M) (u 1 (u2)
U-O u 1 /u2 v=O 1 2
18
!Furthermore, E = (X Et •(O)'P ()1 X E0M:)' )
Substituting (3.9) for Et ((u)), we arrive at
(1 M-12 M-1 M-(I) (
E3.18= P E iv C c()u=O U =10 u 2u 1
1 1 2
(Ul)' (u2)]
Thus, from (3.15)-(3.18) the result holds.
Q.E.D.
Corollary; When each block X (v--O,...,M-l) is chosen with equalv
probabilities (f=t*) we have
2M-1I M-1 i(u) 12(3.19) M(f*,'X;) = (2+ IP12) M-l - I (2Xu.1)
u=O u--O
3.c. Optimum strategies
A strategy of the Statistician is a pair of two M-dimensional
t vectors (f,X) such that t is a probability vector, and %' 1 (M) 1.
Every strategy (f,%) represents a randomization procedure and a
g.l.s.e. The decision problem is to choose (f,X) optimally, with
respect to the loss function M(f,X;P).
Comparing (3.14) to (3.19) it is easily verified that for every
(f,X) there exists fo in En su2h that M(t,X;P°) > M(t*,kX;°) .
Thus, whenever P is arbitrar, t* represents an admissible randomi-
zation procedure. For this reason we shall restrict the discussion
from now on to strategies with randomization procedure t*, and turn
now to the problem of deciding upon an optimum g.l.s.e. under •*.
19
We notice in (3.19) that M(t*,X;•) depends on P only through
the M values 1P(u) 12. An a-priori information concerning these Ivalues might thus be utilized for the choice of X. Thus, let 7(u)
be an a-priori distribution of 1P(u) 12 , defined over the half-line
[0,oa).
Theorem 5.
The Bayes g.l.s.e. of 0, with respect to the a-priori distribu-
tions [FT(O),...,TT(Ml)} , under randomization procedure V is
(M-1))determined by the vector X=, (X(O), ... I , , where
(320 (u) E 7(u)(5.20) iiu M-1 ( for all u=O,...,M-I j
u=0 (u)_I E MuOP 12
Proof:
The risk P'unction under (•*,X) and 7 is
(3.21) R(t*, X;TT) = ( 02 + M-1 E O P(u 12)) M _- 1 (2x2u -l)
I ý~u) Z :(2 1U=O E)Q1-u)2] u0 U=O
E (u)(0 •12)
It is easily verified that X(U) (u=O,...,M-l), given by (3.20),yrM-1
minimize (5.21) under the constraint I x = 1.u=O
Q.E.D.B•)(u)2 M-Il~ )te
Let R(u) = E (u)[C 12) (u=O,...,M-l) and R = thenyr u=O
the Bayes risk with respect to an a-priori distribution 7- is
20
2 M-1 (u2(3.22) R(t*,X (Mf) - (02 + R) it (x(u)) 2
-
U-0
- 1 (2x(U) R(u)- -( ' 2 M-. (U) 2
u=O ix it
In particular, when all 1P(u) 12 (u--O,...,M-l) have the same a-priori
distribution, with R(u) = R* for all u=O,...,M-l, then the Bayesi It
g.l.s.e. is represented by X* = ( 1 ., with a Bayes risk
2(3.23) R(-*,x*,T) =.-+ (M-1)R*
M i
Theorem 4.
* = 1 l represents the minimax and admissible g.l.s.e.M
under randomization procedure k* relative to the class of allM-I
a-priori distributions T , such that R = R (u)
2u=O(C < R i< ). The minimax risk is given by (5.23).
Proof:
The minimax risk is the maximal Bayes risk, with respect to all
the a-priori distributions 7 in the class considered. The BayesM-1 (u)
risk for any of these v's is given by (3.22) where R = uu=O
is a given constant. Set the Lagrangian
(3.24) L(R(o),...,R (M1) ;P) = H' L 2.. M-( (u) )2+
+ -M-1
R(u)uu=O
21
By differentiating partially with respect to R(uj (u=O,...,M-l) and
p and equating the derivatives to zero we arrive at the system of
linear equations:
2 i-2 R(U)(- 1 2 2 • ) = R for all u--O,...,M-l
(3.25) (--
-I R~u Ru=O
The solution of this system of linear equations is given by R(u) R
2 M
for every u=O,...,M-l. Furthermore, since R > a , all the second
order partial derivatives with respect to R.u) are negative. Thus all ia-priori distributions n such that R(u) =
v W - for every
u=O,...,M-I are minimax strategies for Nature. As mentioned before, I1 l1 (M)
* = lM is then the unique minimax strategy for the Statistician.Mi
The Bayes risk corresponding to X* is given by (3.23). The admissi-
bility of X*, relative to the class of a-priori distributions con-
sidered, follows from the fact that it is the unique minimax.
Q.E.D.
22
4. The genraized inverse and the g.l.s.e.
4.a. The g.l.s.e. of minimum norm
A. Ben-Israel and S. J. Wersan (1962) proved that the g.l.s.e.
(L v) with minimum norm, i.e., min tr (L v),(~ ,v is a particular
v ~(L) v
generalized inverse (C v)I of the matrix of coefficients in (2.19),
namely (C) (1,1 C M ...1 (M~) ) X (C (S) ). The generalized in-
verse (C v)t always exists, it is unique, and given in general by
the formula:
(4.1) (C ) = [I (N- (D )(D' D f(D )'( vvC
for all v=O,...,M-l; where (E v) is a product of elementary trans-
formations, which transforms (C )'(C)v into:
(4.2) (E v)(X )'(X) ..[ . ..... (v=0 . . . M-l)
L(0) (0)-and where
(4.3) (D v) v0..M
The generalized inverse matrix, (C V) has the properties:
(4.4) (C v)(C v)t(C v) (C V) for all v=0,.,.,M-l
and
(C V) (C )(C )' P (C v)
A straightforward computation of (C v)t according to formula (4.1)
yields the result
23
"[(o) 1(4.5) (C, M" I
nM-d)
That is, (Cv)t is a g.l.s.e. represented by X* = (M) and hasM
the optimal properties mentioned in the previous section. This re-
sult can be obtained in the present framework more easily. According
to (3.8) a g.l.s.e. is given by
(M)
[M-i C(v(Ml1
Accordingly, the norm of (Lv) is
(4.7) tr. (Lv)'(L): tr. (L M-l x2 (C(S))(C(S))P)
S u--O
M-I M-I=Etr. I xU I I "
u=O u=O
Since the vector X* = l mi M-1 %2 under the constraintSine heveto k = minimizes I
M-1MZO x = 1 , it follows that the g.l.s.e. represented by X* minimizes
= U
the norm of (Lv), (v=O,...,M-l).
4.b. The g.l.s.e. suggested by C. R. Rao.
C. R. Rao (1962) defines the g.l.s.e. of P by the operator
(Lv) = [(Cv)'(C)v" (Cv)' where [(Cv)'(C )v] is a generalized in-
verse of (Cv)'(Cv). In case (C)' (Cv) is invertible, (Lv)" is the
24
!
unique g.l.s.e. of A. We shall prove now that under the present model
of fractional replication designs, Rao's g.l.s.e., (Lv)- , is represented
by the vector X" = (i,0,0,...,0). For this purpose, define the matrix
of elementary transformations
l 11c(M) o1 Vl . (s)(4.8) (E ) o ". (E S
_C(M) 1v(M-1)
then we have, for every v=O,...,M-l,
(4.9) (Ev)[(C)'(C)](E )' = (o) +1(0 (0)]
Hence,
(4.10) [Cv)'(Cv)] : (Ev)'(E)
To show this, consider the relationship
(4.11) [(Cv),(Cv)][(Cv),(c)]Cv) ,(cv)] = [(Cv)'(Cv)]
Multiply both sid"s of (4.11) from the left by (E) and from the
right by (Ev)'. Then,
(4.12) (Ev)(C)'(Cv)I(Ev)'(Ev )[(Cv)'(Cv)](E) =v
Or according to (4.9)
25
'I
I r rS0'(l (0)o
. [(o) (0o)] 1o0) (0 1 (0) (0) 1
Accordingly,
(4.14) (L) (Ev)'CEv)(Cv
M -CvM ... Cv(M.±) 1
_(M)1 "vl 0 '@ (s)S(M) . c(M)
•(M-) .0 • v(M-1) j"Ucv(M-l)
1 0 (C(S))' [~0]
That is, Pao's g.l.s.e. (L v is represented by %_ = (1,,...,0)'.
From theorem 1 it follows that (L )_ is an unbiased estimator
of (P(0),O). The trace of the dispersion matrix of (Lv)_ Y(XV)
under randomization procedure t* is given according to (3.19) by
(4.15) M(t*X;) = 02 + 2 Ila2 - 21P(O)12
2 M-1 P(u)12
u=l
Thus, in case all 1P(u) 12 have the same a-priori distribution, the
risk under strategies (l*,X) and n will be
26
1. (4.16) 2 0 + 2(M-1) R*
IComparing (4.16) to (3.23) we conclude that Rao's g.l.s.e. (Lv
1 might be very far from the optimum g.l.s.e. in case all the subvectors
of P have approximately the same average effect. On the other hand,
in case the effects of P(1) to 0(M-1) are negligible relative to
the effect of P(0) the Bayes g.l.s.e. will be very close to Rao's
g.l.s.e.
Acknowledgment
The author gratefully acknowledges Professors H. Solomon and
H. Chernoff for their helpful comments and encouragement.
27
References
1. A. P. Dempster (1960), "Random allocation designs I: On general tclasses of estimation methods", Ann. Math. Stat. Vol. 31,pp. 885-905.
2. A. P. Dempster (1961), "Random allocation designs II: Approximativetheory for simple random allocations", Ann. Math. Stat. Vol. 32,pp. 387-405.
3. S. Ehrenfeld and S. Zacks (1961), "Randomization and factorial ex-periments", Ann. Math. Stat. Vol. 32, pp. 270-297.
4. S. Ehrenfeld and S. Zacks (1962), "Optimum strategies in factorialexperiments", submitted for publication in the Ann. Math. Stat.
5. 0. Kempthorne, The Design and Analysis of Experiments, John Wiley,New York (1952).
6. C.R. Rao (1962), "A note on a generalized inverse", Jour. Roy.Stat. Soc. (B) Vol. 24, p. 152.
7. B.V. Shah and 0. Kempthoren (1962a), "Some properties of randomallocation designs", Technical Report, Stat. Lab. Iowa Stat. Univ.
8. B.V. Shah and 0. Kempthorne (1962b), "Randomization in fractionalfactorial", Technical Report, Stat. Lab. Iowa State Univ.
9. K. Takeuchi (1961), "On a special class of regression problems andits applications: Random Combined Fractional Factorial Designs",Rep. Stat. Appl. Res. JUSE, Vol. 7, No. 4, pp. 1-33.
10. K. Takeuchi (1961), "On a special class of regression problems andits application: some remarks about general models", Rep. Stat.Appl. Res. JUSE. Vol. 8, No. 1, pp. 7-17.
11. S. Zacks (1962), "On a complete class of linear unbiased estimatorsfor randomized factorial experiments", submitted for publication inthe Ann. Math. Stat.
28
STANFORD UNIVERSITY
TECHNICAL REPORTS DISTRIBUTION LIST
Are tvcsTcnclComreandigfficer Document LlearwyArmd oresl Aec Flanklerd Arsenal U.S. Atomic Envroy Coneetslosnt
Arington alls Sageonc Lihrary bench, 0270, Bldg. 40 J.9th. and Constitution Ayes. N.W.
Aurlngton 12, Virginia 10 Bridge and Tacon-1 Streets Washington 25,0D. C.
BIuo upisadAcut Philadelphia 37, Pennsylvania HeduatrCnde OW Conseenading Officer Oklahoma City Air Materiel AreaDepartment oE the Niavy Reck Island Arsenal United Stales Air ForceWashington 25, 0 C. 1 Reck Island, Illinois 1 Tinker Alr Farce Base,
OklahomaHead Logistics AM Mathemaetical - Coemmanding General
itstigllcs Braech Redatene Arsenal (ORODW-OCt Insltitte ye StatisticsOffice of Naval Reosoerch Huntsville, Alabama 1 North. Carolina Stats Callege of A & E
Cads 436 .3 Comnig eea Raleigh, North. Carolina
Washington 25, D.C.3 ilme Sndsn Genein real iiPv sinLbrtr
Commandieg Officer (ORDBS-TS-TiIS California Inst tate of Technoiogy
Office of lNeval Research Las Cruces, iNew UMics 1 Alte! A.J. SteutchBranchOffice4800 Oak Groeta Drive
Braney Nfie.10FleP.. Canmmending Generai Pasadena 3, CallfornisNavy Yor. 100 y.otP.0 2 Aite, Peaul C. Coo, Ord. Mission
ho ok t .White Sends Preraleg Groued Librarian
Commanding Officer Las Cruces, Now Mexico 1 The RAND CorporatienOffice of Naval Research 1700 Male SaWeBranch Office Coomendl?, General Santa Monica, Cailifornia
1000 Francsc 9, 1aiori Signal Corns Engineering Laboratory Library DivisoneSan rancsco9, CilfoiaFarn Maneenowts Now Jersey 1 Nasal Missile Cene Command
CCommening Officer U.S. Navel Missile Center
Office of Naval Researh eueinGnral tnJ .NceBranch Office Ordnance Wepos ommand Psint Mugu, California
10th. Fleor, The John Cgravr Ally: Research BranchLibrary Bldg Rock Island, Illinois 1 Mathematics Division
BA Eni Rndolh SteetCede 5077CEagst1 Randolph 1tee Commeanding General U.S. Nasal Ordnance Test Siatine
USAryElectronic Proving GrosundChnLaeCsifrlCommanding Officer F~Han wc ArizonaOff ice of fNaval Reseach Attn; Tec;e;Ical Llhtary 1 NASA
Braich Office Ail.: Mr. E.B. .Iacksen. Office
346, Breadwy Commander of Aaer IntelligenceNew York 13, N. Y. 1 ftlght Air Oeesmn etr17 24 F Sireet N. W.
Ally, ARL TTch. Li~brary, WCRR Washingon~ 25, D. C.
Commending Officer Wright-Psilarson Air Force Base. Ohio 1Diammond Ordence Fase Labs, Nationsal Applied Matheematics Labs.Washington 25, 0. C. 1 Connentrdee National Buresau of Standards
Western Development Division. WOSIT Washington 25. 0. C.Commanding Officer P.O. Box 26 2
ictlre rsenal IORDBB-THSil Inglewood, Califorsia 1 Naval Inspector of OrdnanceDowr,lj~~e 1s" U.S. Naval Gon Factory
Chief, Research Division Washington 25. 0. C.
Commradirtg Officer Office of Research & Develepment Attint Mrs. C. D. HachWtatertwn Arsenal IOMRO) oltic of Chief of StallWaterfman 72, Massachusetts 1 U S. Afrny Office, Asst. Chief ol Stalf. G-4
Washington -S, D. C. 1 Retsearch Branch, Rt & 0 Division
Commending Officor Depaertment of the Ant"AiM, W. A. Lahs Chief. Conrytlnri Laboratory Washington 25, 0. C.1
Wateltown Arsenal Ballistic Research LaboratoryWatertown 7?. Massachusetts 1 Akardeen Proving Ground, Maryland 1 S.peh~tedori
U.S. Navy Pastor datle School
Caynnanding Olilcor tilecttor Monterey, Calif=or[&WaeriltArenlNailor~i Securnt Age-.y Alty: LibraryI
Watervliet. Nrewa YIh Attn: REM4-1Waevit o okFort George C. Mend., Maryland 2 Technilcal information Officer
Commending Officer lNeval Research Laboratory
Alttn. Inspection Division Director ye Operations Washington 25, D. C. 6Springfield Armory Operattttio Analysis Div.. AFOOP
Springfield, Massachusetts 1 Hil., U.S. Air Force Techetcai Information ServiceWashington 25. 0. C. 1 Ally: Reference Branch
Conmndin 011cm.P.O. Box h2
Signai Corps Electronic Reseach Director OhRde enseUnit, EDL Seas, Ice & Permafrost Research
9560 Technical service unit Establishment Technical Library Branch
P.O. Box 205 Corps of Engineers Cad. 234Mountlain View, California 1 1215 Washington Avenue U.S. Naval Ordnance Lahoratory
Wilnefto, Illineis 1 Atln: Clayhorn Granes
Commanding Officer Coronea, CaillIorsiae9550 Technical Service Unit DirectorArm Liaison Grosp Project Michigan Lincoln Laboratory Institute for Defense AnalysesWillow Ron Research Car.r I. oxi,.gton, Massachusetts 1 Ceonmunications Research DivisionYpsilanti, Micbl'vba- 1 non Neumnann Haill
Departmeori of Mathematics Princeton, New Jersey
Commanding Officer Michigan State UniversityEngineering Research & Dnevelpment Lahs, East Lanning, Michigan1Fort Beinsir, Virginia
August, 196,2
Mr. Irmine B. Altman Professor Sol omon Kialkback Professor L. J. SeempInspection 11 QC Division Department of statistics uaimuDogeatmordOffice Ass'. Secretary i Defense GereWashington University Uiagt fUuapRoom iB87O, The PeteaWs= gton 7, D. C. 1ýh~ AI ~ Ie icuWashington 25, D. C. I1 rfso .H rsa opesrW .SolProfessor T. W. A'idarvan D.Paritmett of Statistics Statistics DopsormeadOeparmmnlof Statietlet The University of Chicago UniversityroNami CarollieColombia Uijoversity Chicago, hIinis@l 1 Chaof Hill, N"s CeolirNow York 27,New Yorh 1
Professor Eugen Lulkacs Dr. Milton SadProfessor Rabert Bachiehfer Department of Mothoolatics Shaalstlaa Dqor5magDeat. of Industiual and Engineering Catholic University University of Mineagata
Admmivnistration Washington 15, 0. C.1 MleeuhMmgaSibley School of Mechanical EnginreeriInO.CaiMairM.6.pSa&
tesNwNak1 2954 Winchester Way DIVslaw 5511Ithac. les~orhRancho Cordova, Cal ifornia 1 SandIa Corp., Sandia bam
Pynfassor Fred. C. Andrews Aluur ,Hv oiiiDeormn of Mtalls*ties Professor G. W. McElrath AbamraNwMaa
nieSity of Oregon Department of Mechanical Englunemne Picofess Donakll TreatEugene. Orngon 1 Unviersity Of Minneseta Dfni~tional a' Maalsmtlas
Minneapolis 14, Minnesoia 1 thrivemtaOgeProfstaor Z. W. BirbsuanDoparter of mathemsatics Dr. Knox T. MilloalimUniversity of Vhshington Enecutive Director Pi Ofeads Johen W. TeySeattle 5, Washington 1 Air Force lifilc* of Scientific Resgeach Dogealumaea of Mgladsaics
Df. iaod BlckinliWashington 2S,0D. C. 1 prisicat"s tkieasltyDoesartrent of Mathematical sciences 0. E. Newnham leJuaLhsiner,fty of Cailifornia Chief, Ind. Enter. Div. Comptroller Pirofegar C. S. 1114sbe
Barbivy , CaltorlaU1., San Dernardina Air Material Area Dam4.net a Mes'emotaAF. Noiton Air Force Base, California 1 Univenrsit of Teprohee
Professor Ralph A. Bradley PoesrEwnGOlsTeml 5,Outario, &aeadaFlorida State iUniversity Departmrent af Mathematic or. lHarry wamillsueeTallahassee, Florida 1 Colnegie iof Elecisn andSence Special Prjcs Offl, SP201b
O JohnW Cii o llei Pittsbunrghr 13, Pennsylvanila 1 Wanito 5,D. C.11Depaetment oil MatheatricsCoils Carolina S~tat college Dr. Williae Rt. Pabst Dr. F. J1. laryl, Dirgctor
Nalegir, North Caralina I Bureauo f Weapons MatleAmeical Scieme- DeimiioRooer 0306,, Main Navy Office of Ildval Regseach
Professor Willi~amG. Cecivan Day ienvtof te avay Weoimtee* 25, 0. C.Department or statistics Washingtonl 25., 0. C.1Harvard Unriversity Dr. John WIlle2 Divinity Avenue, Rtoom 311 Mr. Edward Paulson * Office of Nanai Reseerch, Cede 200Camdbridge 38, Mdassachusetts1 72-10 41 Aoe. Wabna.25,0D. C.
Woodside 77MiasasBst&f8 Day No.sYork, PNe'wYork 1 Professor S. S. WhiitBureau of Ships, Code 3420 Deartment of Mafh~itsRo-w3210 Main Nav H. Walter Price. Chief Pyltuate UnieersityDeparlorent of the Navy Reliaihlity Branch, 750 P,,cgtet, Now JeeseWashington 25, D C.1 Diamond tknnasce Fuze Laboratory
Rorinr 1015, Building 83 Mr. Silas MIWKllmOr. Waiter L. Daemean. Jr Washingtion 25, 0. C 1 Standards Braoch, Pride. Div.Operations Analysis Div., DCE/D Office, DC)S for Legistico;Hq., U.S. Air Force Proites,o Ronald Pyke Departmnt of thse ArmyWashington 25, 0. C. 1 Manthasics Departmentn Washington 25, 0. C.I
UnInensily v0 WashingtonProfessor ',yrns Deonrn Seattle 5, Washington 1 PoasracbW-flfomitDeer 01 Industrial Einritneelog Depgauss of MathematicsCoI..rrila University Or iaoi Rider COsrnefl Universityless York 27. New York 1 Wrighi Alt Denvelopmrent Center, WCRRM IOWAc, New York
M ight-Pattnrnscw, A. F.B. , Ohio 1Or. Donald P. Goner Mr. William W. WeimeaWesllvghouse Research Labs. Promiseor Herbert Riobbins CedeME= Blg. oT-2 ReamwC301
Bouiah Rd. - Churchill Bore. Dept. of Mathemnatical Statistic, 700 econPlac , . W.
Pittsbu~rgh 35, Pa. 1 Coilombia University Washinsgton 25. D. C.I
Mr. Harold GSwybarinrileHevad, Operations Research Group Professor fMlonay Rosenblatt Mothvin . RaeserhCneCode D1-2 Ovpartrve,, of Mathematics U. S. ArmyPacific Missile Range Brown Un~ivnrsity University of Wiscensie
CaBoonx I Providence 12, Rhode island 1 Madisont 6, WiscoensinPoin Mug, ClifoniaProfrssor Herman Robin
Or. Ivan Hersinser Oepartmernt of Statistics Additional cspies for meaiactOffice, Cisiaf of Research & Dow. Michigan Stair University leader aied assistants andU S. Army, Research Division 3E382 East Lanmina, Michigan 1 for future requilgments 50Washinggon 25, 0. C. 1.SProfessor W Hirsch Cvi leer of MediciveInstitute of Mothemaicai Sciencess University of CincinnsatiNiow York University Cincinnati, Ohio1
Now Yrk 3 NowYorkProfessor 1. R. SavageMr. Eugn HInoan School of Business AdministretionCoda 600.1 University of MivnesotaGSFC, NASA Minneapls, Mirnvesota1
Grenbet, aryandMiss Marion M. SaridonirProfessor Hlarold Hoteiling 2281 Cedar StreetDepartment of Statistics Berkeley 9, California1Uviversity of North CarolinaChapel Hiii, North Carolina1
Co Inoeatlee225152)Aagestl, 19b2
I
JOINT SERVICES ADVISORY GROUP
Mr. Fred Frishman Lt. Col. John W. Querry, ChiefArmy Research Office Applied Mathematics DivisionArlington Hall Station Air Force Office of Scientific
Arlington, Virginia 1 ResearchWashington 25, D. C.
Mrs. Dorothy M. GilfordMathematical Sciences Major Oliver A. Shaw, Jr.
Division Mathematics DivisionOffice of Naval Research Air Force Office of Scientific
Washington 25, D. C. ResearchWashington 25, D. C. 2
Dr. Robert LundegardLogistics and Mathematical Mr. Carl L. Schaniel
Statistics Branch Code 122
Office of Naval Research U.S. Naval Ordnance Test
Washington 25, D. C. 1 StationChina Lake, California
Mr. R. H. NoyesInst. for Exploratory Research Mr. J. WeinsteinUSASRDL Institute for Exploratory Research
Fort Monmouth, New Jersey 1 USASRDLFort Monmouth, New Jersey 1
iii