Remark 4.4.8. Actually, condition (4.45) is obviously fulfilled for ν = 2 (compare Theorem 4.4.3). For the case ν = 3 the bound of l.h.s. of condition (4.45) is 1 and, hence always fulfilled.
to a model with intercept by removing the interaction term x1x2. As it was pointed out in Remark4.4.5the functionfβ(x) is invariant w.r.t. simultaneously scaling of x, i.e., fβ(λx) =fβ(x). Let λ= 1/(x1x2) then we obtain
fβ(x) =β1x1+β2x2+β3x1x2−1
x1 x2 x1x2
(4.48)
=β1t2+β2t1+β3−1
t2 t1 1
=f◦β(t) (4.49)
where t = t1, t2
>
, tj = 1/xj, j = 1,2. The range of t = t(x), as x ranges over X = [a, b]2 is a cube given by T = h(1/b), (1/a)i2 with f◦(t) = (t2, t1,1)>. One can rearrange the terms of (4.49) by making use of the 3×3 anti-diagonal transformation matrixQ. So we havef˜(t) = Qf◦(t) = (1, t1, t2)> and ˜β =Q>−1β =β3, β2, β1>. Hence, f˜β˜(t) = f˜>(t) ˜β−1f˜(t) and rewrites as
f˜β˜(t) = β3+β2t1+β1t2−1
1 t1
t2
, t ∈h(1/b), (1/a)i2. (4.50)
Since (4.50) coincides with that for a gamma model with intercept the D-criterion is equivariant (see Radloff and Schwabe (2016)) with respect to a one-to-one transforma-tion fromT =h(1/b), (1/a)i2 to Z = [0,1]2 where
tj →zj = 1
(1/a)−(1/b)tj− 1/b
(1/a)−(1/b), j = 1,2. (4.51) For a given transformation matrix
B=
1 0 0
−(1/b) (1/a)−(1/b)
1
(1/a)−(1/b) 0
−(1/b)
(1/a)−(1/b) 0 (1/a)−(1/b)1
with B−1 =
1 0 0
1 b
1
a− 1b 0
1
b 0 a1 −1b
we have f˜(z) =Bf˜(t) = (1, z1, z2)> withβ˜˜=B>−1β˜ = (β˜˜0, β˜˜1, β˜˜2)> where
˜˜
β0 = β3 + (1/b)(β1 +β2) , β˜˜1 = β2((1/a)−(1/b)) and β˜˜2 = β1((1/a)−(1/b)). It follows thatf˜β˜˜(z) =f˜>(z)β˜˜−1f˜(z) which rewrites as
f˜β˜˜(z) = β˜˜0+β˜˜1z1+ β˜˜2z2)−1
1 z1 z2
,z∈[0,1]2. (4.52)
Let M(x,β) = fβ(x)f>β(x), M˜(t,β) =˜ f˜β˜(t)f˜>β˜(t) and M˜˜(z,β) =˜˜ f˜β˜˜(z)f˜>β˜˜(z) be the information matrices for the models which corresponding to (4.48), (4.50) and (4.52), respectively. It is easy to observe that
M(x,β) = Q−1M˜(t,β)Q˜ −1 =B−1Q−1M˜˜(z,β)Q˜˜ −1B−1,
thus the derived D-optimal designs on X, T and Z, respectively are equivariant. Ac-cording to the mapping of x to t in the line following (4.49) and the mapping from t toz in (4.51) each component is mapped separately: xj →tj →zj without permuting them. Therefore, one modifies the direct one-to-one transformationg :X → Z where
xj →zj = 1/xj
(1/a)−(1/b) − 1/b
(1/a)−(1/b), j = 1,2. (4.53) Letξg∗ be a design defined onZ that assigns the weights ξ(x) to the mapped support points g(x), x ∈ supp(ξ∗). In fact, ξ∗ on X is locally D-optimal (at β) if and only if ξg∗ onZ is locally D-optimal (at β). Note that the function˜˜ f˜β˜˜(z) is injective and thus the optimal design on Z is only supported by the vertices (cp. Theorem 4.2.1). It is worth noting by transformation (4.53) we obtain
(b, b)>→(0,0)>, (b, a)> →(1,0)>, (a, b)>→(0,1)>, (a, a)>→(1,1)>.
Corollary 4.5.2. Consider f(x) = x1, x2, x1x2> on X = [a, b]2, 0 < a < b. De-note the vertices by v1 = (b, b)>, v2 = (b, a)>, v3 = (a, b)>, v4 = (a, a)>. Let β= (β1, β2, β3)> be a parameter point satisfying (4.4). Then the unique locally D-optimal design ξ∗ (at β) is as follows.
(i) If β32+b12(β12+β22) + (b12 −a12 +a b2 )β1β2+2bβ3(β1+β2)≤0 then ξ∗ assigns equal weights 1/3 to v1,v2,v3.
(ii) If β32+b12β12+a12β22+2bβ3β1+2aβ3β2+ (b12 +a12)β1β2 ≤0 then ξ∗ assigns equal weights 1/3 to v1,v2,v4.
(iii) If β32+b12β22+a12β12+2bβ3β2+2aβ3β1+ (b12 +a12)β1β2 ≤0 then ξ∗ assigns equal weights 1/3 to v1,v3,v4.
(iv) If β32+a12(β12+β22) + (a12−b12 +ab2)β1β2+a2β3(β1+β2)≤0 then ξ∗ assigns equal weights 1/3 to v2,v3,v4.
(v) If none of the cases (i) – (iv) applies then ξ∗ is supported by the four vertices
ξ∗ =
v1 v2 v3 v4 ω1∗ ω2∗ ω3∗ ω∗4
, where ω∗` >0 (1≤`≤4), P4`=1ω`∗ = 1.
Proof. The proof is obtained by verifying the D-optimality ofξg∗ on Z = [0,1]2 under transformationg given by (4.53). The regression vectorf˜β˜˜(z) given by (4.52) coincides with that for the two-factor gamma model with intercept onZ = [0,1]2 whose intensity function is defined asuβ˜˜(z) = (β˜˜0+β˜˜1z1+ β˜˜2z2)−2 for all z∈ Z. Denote
c1 =uβ˜˜((0,0)>) = β˜˜0−2 = (β3+ 1
b(β1+β2))−2, c2 =uβ˜˜((1,0)>) = (β˜˜0+β˜˜1)−2 = (β3 +β11
b +β21 a)−2, c3 =uβ˜˜((0,1)>) = (β˜˜0+β˜˜2)−2 = (β3 +β11
a +β21 b)−2, c4 =uβ˜˜((1,1)>) = (β˜˜0+β˜˜1+β˜˜2)−2 = (β3+ 1
a(β1+β2))−2.
Leth, i, j, k ∈ {1,2,3,4} are pairwise distinct such that ck = min{c1, c2, c3, c4} then it follows from Theorem 3.3.1 that if c−1k ≥ c−1h +c−1i +c−1j then ξ∗ is a three-point design supported by the three vertices vh, vi, vj, with equal weights 1/3. Hence, straightforward computations show that the condition in case (i) of the corollary is equivalent toc−14 ≥c−11 +c−12 +c−13 . Analogous verifying is obtained for the cases (ii), (iii), (iv). For case (v) the four-point design according to Theorem 3.3.1 is locally D-optimal if c−1k < c−1h +c−1i +c−1j which applies implicitly if non of the conditions (i) – (iv) of saturated designs is fulfilled by a given β (cp. Remark 2.2.4).
It is noted that the optimality conditions provided in parts (i)–(iv) of Corollary 4.5.2 depend on the values of a and b. Changing these values might affect the D-optimality of a design. To see that, more specifically, let a= 1 and b = 2, i.e., the experimental region is X = [1,2]2 and define γ1 = β1/β3 and γ2 =β2/β3, β3 6= 0.
Here, the parameter space which is depicted in Panel (a) of Figure4.7is characterized by γ2 +γ1 > −1, 2γ2 +γ1 > −2 and γ2 + 2γ1 > −2. It is observed from Panel (a) of Figure 4.7 that the design given by part (i) of Corollary 4.5.2 is not locally D-optimal because the corresponding optimality condition in part (i) of the corollary;
1
4(γ12+γ22) + 14γ1γ2+γ1+γ2 ≤ −1 can not be satisfied.
Let us consider another experimental region with a higher length by fixinga= 1 and takingb= 4, i.e., the experimental region isX = [1,4]2. The parameter space which is
depicted in Panel (b) of Figure4.7is characterized byγ2+γ1 >−1, 4γ2+γ1 >−4 and γ2+ 4γ1 >−4. In this case all designs given by Corollary4.5.2 are locally D-optimal at particular values of γ2 and γ1 as it is observed from the figure. It is obvious that along the diagonal dashed line, γ2 = γ1, there exist at most three different types of locally D-optimal designs.
(a) (b)
Figure 4.7: Dependence of locally D-optimal designs on γ1=β1/β3 and γ2=β2/β3 where for Panel (a) X = [1,2]2 and for Panel (b) X = [1,4]2. The diagonal dashed line represents the caseγ2=γ1. Note that supp(ξijk∗ ) ={vi,vj,vk} ⊂ {v1,v2,v3,v4}and
supp(ξ∗1234) ={v1,v2,v3,v4}.
.
For arbitrary values of a and b, 0 < a < b let us restrict to case γ2 =γ1 = γ, i.e., β1 =β2 =β, β3 6= 0 and the next corollary is immediate.
Corollary 4.5.3. Consider f(x) = (x1, x2, x1x2)> on an arbitrary square X = [a, b]2, 0< a < bin the positive quadrant. Letβ = (β1, β2, β3)>be a parameter point satisfying (4.4) with β1 =β2 =β and β3 6= 0. Define γ = ββ
3. Then the locally D-optimal design ξ∗ (at β) is as follows.
(i) If −a2 < γ ≤ −3b−aab , then ξ∗ assigns equal weights 1/3 to v2,v3, v4.
(ii) If b−3a >0 and γ ≥ b−3aab , then ξ∗ assigns equal weights1/3 to v1,v2,v3. (iii) Ifb−3a >0and−3b−aab < γ < b−3aab then the designξ∗ is supported byv1,v2,v3,v4
with optimal weights given by
ω1∗ = ab−(a−3b)γ
4b(a+ 2γ) , ω2∗ =ω∗3 =
ab+ (a+b)γ2
4ab(b+ 2γ)(a+ 2γ), ω4∗ = ab−(b−3a)γ 4a(b+ 2γ) .
Proof. Consider the experimental region X = [a, b]2, 0 < a < b. By assumption β1 =β2 =β, β3 6= 0 the range ofγ = ββ
3 is given by (−a/2,∞). Assumptionb−3a >0 implies that −a2 <−3b−aab < b−3aab . According to Corollary 4.5.2 we show the following under the assumptions of Corollary 4.5.3. Both conditions provided in parts (ii) and (iii) of Corollary 4.5.2are not fulfilled by any parameter point thus the corresponding designs are not D-optimal. In contrast, the designξ∗ given in (i) of Corollary 4.5.3 is locally D-optimal if the condition provided in part (iv) of Corollary 4.5.2 holds true.
That condition is equivalent to
(3b2+ 2ab−a2)γ2+ 4ab2γ+a2b2 ≤0.
The l.h.s. of above inequality is polynomial inγ of degree 2 and thus the inequality is fulfilled by −a2 < γ ≤ −3b−aab .
Similarly, the design ξ∗ in (ii) of Corollary 4.5.3 is locally D-optimal if the condition provided in part (i) of Corollary 4.5.2 holds true. That condition is equivalent to
(3a2+ 2ab−b2)γ2+ 4a2bγ+a2b2 ≤0.
The l.h.s. of above inequality is polynomial inγ of degree 2 and thus the inequality is fulfilled by γ ≥ b−3aab if b−3a >0.
The four-point design given in (iii) has positive weights on −3b−aab < γ < b−3aab if b−3a >0 and hence it is locally D-optimal in view of Remark 2.2.4.
Remark 4.5.1. One may note that from Corollary 4.5.3 when β = 0 the uniform design on the vertices v1,v2,v3,v4 is locally D-optimal.
4.5.3 Model of complete product-type interactions
Schwabe (1996b) developed an approach to construct optimal designs for linear models of complete product-type interactions by making use of optimal designs under marginal models. This approach is independent of the actual structure of the influence of the single factors and, hence, covers models with both qualitative and quantitative factors as well as purely qualitative or purely quantitative models (see Schwabe (1996b), p.35).
In this subsection we extend that approach for gamma models of complete product-type interactions. We will show that locally optimal designs for gamma models of complete product-type interactions can be obtained from locally optimal designs under the marginal counterparts.
We consider K marginal models each is containing νk factors, k = 1, . . . , K. The marginal νk-factor model is defined with a power link as in (4.2) where
µρk(xk) = f(k)>(xk)β(k), xk = (xk1, . . . , xkνk)> ∈ Xk ⊆Rνk (4.54) with exponentρ∈R, ρ >0. So all marginal gamma models are having equal exponent ρ, or, equivalently, all link functions of the marginal models are the same. The positivity assumption (4.4) of the expected meanµk is to be satisfied, i.e., f(k)>(xk)β(k)>0 for allxk= (xk1, . . . , xkνk)>∈ Xk where
f(k):Xk →Rpk, β(k)= (β1(k), . . . , βp(k)
k )> ∈Rpk (1≤k ≤K).
For each k, the marginal model (4.54) has intensity function
uk(xk,β(k)) =f(k)>(xk)β(k)−2, xk ∈ Xk (1≤k ≤K). (4.55) Letνdenotes the total number of factors in all marginal models, i.e., ν=PKk=1νk. The resulting ν-factor gamma model of complete product-type interactions with exponent ρis thus defined by
µρ(x) =f>(x)β
=
p1
X
i1=1
· · ·
pK
X
iK=1
fi(1)1 (x1)·. . .·fi(K)
K (xK)βi1,...,iK (4.56) where x = (x>1, . . . ,x>K)> ∈ X = X1× · · · × XK, in which f(x) collects all K−fold products of the componentsfi(k)(xk) which belong to the regression functionsf(k)(xk), k = 1, . . . , K and x = (x11, . . . , x1ν1, . . . , xK1, . . . , xKνK)> is a ν-tuple. The un-known parameter βi1,...,iK is equal to QKk=1βi(k)k and β is a p-dimensional parameter vector, i.e., β ∈ Rp where p = QKk=1pk. Note that β collects the parameters βi1,...,iK, ik= 1, . . . , pk, k = 1, . . . , K and in lexicographic order β rewrites as
β= (β1,...,1,1, β1,...,1,2, . . . , β1,...,1,pK, β1,...,2,1, . . . , βp1,...,pK−1,pK)>.
Note that f(x) : X →Rp and can be described by Kronecker products “⊗” as in the following;
f(x) = f(1)(x1)⊗ · · · ⊗f(K)(xK) =
K
O
k=1
f(k)(xk), (4.57) with β =β(1)⊗ · · · ⊗β(K)=
K
O
k=1
β(k). (4.58)
Therefore, model (4.56) rewrites as µρ(x) =
K
O
k=1
f(k)(xk)
> K
O
k=1
β(k)
(4.59) and of coursef>(x)β>0 for all x∈ X. The latter positivity assumption is obtained from that in the marginal models; that is because
f>(x)β =
K
O
k=1
f(k)(xk)
> K
O
k=1
β(k)
=
K
Y
k=1
f(k)>(xk)β(k) >0.
The intensity function u(x,β) for model (4.56) is determined by the product of the intensity functions (4.55) in the marginal νk-factor models (4.54).
Lemma 4.5.1. The intensity function for model (4.56) is given by u(x,β) =
K
Y
k=1
uk(xk,β(k)).
Proof. In general, the intensity function of gamma models with linear predictorf>(x)β is defined byu(x,β) = (f>(x)β)−2. In view of (4.57) and (4.58) it follows that
u(x,β) =
K
O
k=1
f(k)(xk)
> K
O
k=1
β(k)
−2
=
K
O
k=1
f(k)>(xk)β(k)
−2
=
K
O
k=1
f(k)>(xk)β(k)
−2
=
K
Y
k=1
uk(xk,β(k)).
Our aim is deriving an optimal design for model (4.56) as a product type design which is supported by the cross-product of the finite sets of design points of the designs under marginalνk-factor models and the weights are given by the product of the weights of those designs. To be more specific, denote byξka design defined onXkfor a marginal νk-factor model (4.54) (1≤k ≤K). We introduce ξk as in (2.9);
ξk =
xk1 xk2 . . . xkrk ωk1 ωk2 . . . ωkrk
, (4.60)
where ξk has rk design points xkj and corresponding weights ωkj,j = 1, . . . , rk. Then the product type design ξ = NKk=1ξk is defined on X = X1 × · · · × XK and has r = QKk=1rk design points xi1,...,iK = (x1i1, . . . , xKiK)> with corresponding weights ωi1,...,iK =QKk=1ωkik,ik= 1, . . . , rk, k= 1, . . . , K.
In order to reduce the problem of locally optimal designs at a given β for model (4.56) to the locally optimal designs for the marginalνk-factor models at a givenβk it is required to use a factorized information matrix. The information matrix and, hence, the variance-covariance matrix of a product type design ξ factorizes into its marginal counterparts as it is given by the next lemma.
Lemma 4.5.2. The information matrix of a product type design ξ = NK
k=1
ξk for a gamma model of complete product-type interactions (4.56) is
Mξ,β=
K
O
k=1
Mkξk,β(k)
Proof. In general, the information matrix ofξ for the model (4.56) is given by
M(ξ,β) =RX u(x,β)f(x)f>(x)ξ(dx). In view of Lemma4.5.1with (4.57) and (4.58) it follows that
Mξ,β=
Z
Xu(x,β)
K
O
k=1
f(k)(xk)
K
O
k=1
f(k)(xk)
> K
O
k=1
ξk(dxk)
=
Z
X1· · ·
Z
XK
K
Y
k=1
uk(xk,β(k))
K
O
k=1
f(k)(xk)f(k)>(xk)
K
Y
k=1
ξk(dxk)
=
K
O
k=1
Z
Xkuk(xk,β(k))f(k)(xk)f(k)>(xk) ξk(dxk)
=
K
O
k=1
Mk
ξk,β(k).
As a consequence of Lemma 4.5.2 we get
detM(ξ,β)= det
K
O
k=1
Mkξk,β(k)
=
K
Y
k=1
det
Mkξk,β(k)
K
Q
j=1 j6=k
pj
. For example; letξ =ξ1⊗ξ2⊗ξ3 then
det(M(ξ,β)) = det(M1(ξ1,β(1)))p2p3det(M2(ξ2,β(2)))p1p3det(M3(ξ3,β(3)))p1p2. Therefore, for a given parameter pointβ that is evaluated from the parameter points of the marginal models, i.e.,β = NK
k=1
β(k) the best product type designξ∗ = NK
k=1
ξk∗ with respect to the D-criterion is generated from the locally D-optimal designsξk∗ atβ(k) for the marginalνk-factor models (1≤k ≤K).
Theorem 4.5.1. Let ξk∗ be a locally D-optimal design (atβ(k))for a marginalνk-factor model (4.54) on the experimental region Xk (1 ≤ k ≤ K). Then ξ∗ = NK
k=1
ξk∗ is a
locally D-optimal design (at β = NK
k=1
β(k)) for model (4.56) on the experimental region X =×Kk=1Xk.
Proof. The proof is obtained by making use of condition (2.11) of The Equivalence Theorem (Theorem2.2.2). To this end, denotef(k)β(k)(xk) =
q
uk(xk,β(k))f(k)(xk) and fβ(x) = qu(x,β)f(x). Since ξk∗ is locally D-optimal (at β(k)) we guarantee that f(k)>
β(k)(xk)M−1k (ξk∗,β(k))f(k)
β(k)(xk) ≤ pk for all xk ∈ Xk. Thus in view of Lemma 4.5.2 with (4.57) and (4.58) we obtain
f>β(x)M−1(ξ∗,β)fβ(x) =u(x,β)
K
O
k=1
f(k)(xk)
> K
O
k=1
Mkξk,β(k)
−1 K
O
k=1
f(k)(xk)
=
K
O
k=1
uk(xk,β(k))f(k)>(xk)M−1k ξk∗,β(k)f(k)(xk)
=
K
Y
k=1
f(k)>
β(k)(xk)M−1k (ξk∗,β(k))f(k)
β(k)(xk)≤
K
Y
k=1
pk =p
for all x ∈ X. The Equivalence Theorem, thus, proves the local D-optimality of the product designξ∗ = NK
k=1
ξk∗ at a parameter point β= NK
k=1
β(k).
In the following we focus on the local A-optimality. From Lemma 4.5.2 we obtain the next straightforward factorization for every product type designξ = NK
k=1
ξk.
trM−1(ξ,β)= tr
K
O
k=1
Mkξk,β(k)
−1
=
K
Y
k=1
tr
M−1k (ξk,β(k))
.
Hence, in analogy to the previous case of local D-optimality for a given parameter point β that is evaluated from the parameter points of the marginal models, i.e., β = NK
k=1
β(k) the best product type design ξ∗ = NK
k=1
ξk∗ with respect to the A-criterion is generated from the locally A-optimal designs ξk∗ at β(k) for the marginal νk-factor models (1≤k ≤K).
Theorem 4.5.2. Let ξk∗ be a locally A-optimal design (atβ(k))for a marginalνk-factor model (4.54) on the experimental region Xk (1 ≤ k ≤ K). Then ξ∗ = NK
k=1
ξk∗ is a locally A-optimal design (at β = NK
k=1
β(k)) for model (4.56) on the experimental region X =×Kk=1Xk.
Proof. The proof is obtained by making use of condition (2.12) of The Equivalence Theorem (Theorem 2.2.2 ). Sinceξk∗ is locally A-optimal (at β(k)) we guarantee that f(k)>
β(k)(xk)M−2k (ξk∗,β(k))f(k)
β(k)(xk)≤tr(M−1k (ξk∗,β(k))) for all xk ∈ Xk. Thus in view of
Lemma4.5.2 with (4.57) and (4.58) we obtain f>β(x)M−2(ξ∗,β)fβ(x) =u(x,β)
K
O
k=1
f(k)(xk)
> K
O
k=1
Mkξk,β(k)
−2 K
O
k=1
f(k)(xk)
=
K
O
k=1
uk(xk,β(k))f(k)>(xk)M−2k ξk∗,β(k)f(k)(xk)
=
K
Y
k=1
f(k)>
β(k)(xk)M−2k (ξk∗,β(k))f(k)
β(k)(xk)
≤
K
Y
k=1
tr(M−1k (ξk∗,β(k))) = tr
K O
k=1
Mk
ξk∗,β(k)
−1
= trM−1(ξ∗,β)
for all x ∈ X. The Equivalence Theorem, thus, proves the local A-optimality of the product designξ∗ = NK
k=1
ξk∗ at a parameter point β= NK
k=1
β(k). Example 4.5.1. Marginal single-factor models.
Here we consider model (4.54) with one factor, νk = 1, and two parameters, pk = 2,.
We restrict ourselves to the case of two marginal modelsk = 2 where
µρ1(x1) = f(1)>(x1)β(1) =β0(1)+β1(1)x1, x1 ∈ X1 = [0,1], (4.61) µρ2(x2) = f(2)>(x2)β(2) =β0(2)+β1(2)x2, x2 ∈ X2 = [0,1] (4.62) such that β0(k) > 0, β0(k)+β1(k) > 0, k = 1,2. The resulting 2-factor gamma model of complete product-type interactions is thus written as
µρ(x) = f>(x)β=β0+β1x1+β2x2 +β3x1x2 (4.63) where x= (x1, x2)> ∈ X =X1× X2 = [0,1]2,β ∈R4,β=β(1)⊗β(2), hence
β=
β0 β1 β2 β3
=
β0(1)β0(2) β0(2)β1(1) β0(1)β1(2) β1(1)β1(2)
, (4.64)
where β satisfies f>(x)β>0 for all x ∈ X. Note that the 2-factor model with in-teraction (4.63) was considered in Subsection 4.5.1 where locally D- and A-optimal designs were derived. However, from Corollary 4.3.1 the marginal design ξk∗ on Xk = [0,1] with support {0,1} is equally weighted D-optimal (at β(k)). Thus the design ξ∗ = ξ1∗ ⊗ ξ2∗ is locally D-optimal (at β = β(1) ⊗ β(2)) on the experimental region X = [0,1]2 for model (4.63). The design ξ∗ assigns uniform weights 1/4 to the support (0,0)>,(1,0)>,(0,1)>,(1,1)>.
Moreover, from Corollary 4.3.1 the marginal design ξk∗ onXk = [0,1] with support {0,1} is A-optimal (at β(k)) with weightsωk0∗ = √
2β0(k)/((√
2 + 1)β0(k) + β1(k)) and ωk1∗ = (β0(k)+β1(k))/((√
2 + 1)β0(k)+β1(k)). For a given β =β(1)⊗β(2) from (4.64) the product type designξ∗ =ξ1∗⊗ξ2∗on the experimental regionX = [0,1]2 for model (4.63) is locally A-optimal and assigns weights ω∗10ω20∗ to (0,0)>, ω∗11ω20∗ to (1,0)>, ω∗10ω21∗ to (0,1)> and ω11∗ ω21∗ to (1,1)>. Clearly, the product type design w.r.t. A-criterion coincides with that from Corollary4.5.1 where for instance; the optimal weightω∗10ω20∗ of the point (0,0)> leads to the identity
2β0(1)β0(2) (3 + 2√
2)β0(1)β0(2)+ (1 +√
2)(β0(1)β1(2)+β0(2)β1(1)) +β1(1)β1(2)
= 2β0 c
wherec= (3 + 2√
2)β0+ (1 +√
2)(β1+β2) +β3.
Example 4.5.2. Marginal two-factor models.
Here we consider model (4.54) with two factors,νk = 2, and three parameters,pk= 3,.
The case of two marginal modelsk = 2 is also adopted where
µρ1(x1) = f(1)>(x1)β(1) =β0(1)+β1(1)x11+β2(1)x12, x1 = (x11, x12)> (4.65) µρ2(x2) = f(2)>(x2)β(2) =β0(2)+β1(2)x21+β2(2)x22, x2 = (x21, x22)> (4.66) where x1 ∈ Xk = [0,1]2, k = 1,2 such that β0(k) > 0, β0(k)+β1(k) > 0, β0(k)+β2(k) > 0, β0(k) +β1(k) +β2(k) > 0, k = 1,2. The resulting 2-factor gamma model of complete product-type interactions is thus written as
µρ(x) =β0+β1x11+β2x12+β3x21+β4x22+β5x11x21+β6x11x22
+β7x12x21+β8x12x22 (4.67)
where x = (x11, x12, x21, x22)> ∈ X = [0,1]4,β ∈ R9,β = β(1) ⊗ β(2). Note that β0 =β0(1)β0(2). From part (i) of Corollary 4.4.1 the design ξ∗k on Xk = [0,1]2 with support {(0,0)>,(1,0)>,(0,1)>} for the kth marginal model is equally weighted D-optimal (at β(k)) if and only if (β0(k))2 ≤ β1(k)β2(k), k = 1,2. The latter optimality conditions transfer to the product type design ξ∗ = ξ1∗ ⊗ξ∗2 which is thus locally D-optimal on the experimental region X = [0,1]4 for model (4.67) at β=β(1)⊗β(2) if (β0(1)β0(2))2 ≤(β1(1)β1(2))(β2(1)β2(2)), i.e., β02 ≤β5β8 and ξ∗ assigns weights 1/9 to
{(0,0)>,(1,0)>,(0,1)>} × {(0,0)>,(1,0)>,(0,1)>}=
{(0,0,0,0)>,(0,0,1,0)>, (0,0,0,1)>, (1,0,0,0)>, (1,0,1,0)>, (1,0,0,1)> ,(0,1,0,0)>, (0,1,1,0)>, (0,1,0,1)>}.
For A-optimality, from part (iv) of Corollary 4.4.2 the design ξk∗ on Xk = [0,1]2 with support {(0,0)>,(1,0)>,(0,1)>} and the following weights
ξk∗(0,0)>=√
3β0(k)/c, ξk∗(1,0)>= (β0(k)+β1(k))/c,
ξk∗(0,1)>= (β0(k)+β2(k))/c wherec= (√
3 + 2)β0(k)+β1(k)+β2(k), k = 1,2, is locally A-optimal (at β(k)) if and only if
(1 + 2/√
3)(β0(k))2+ (1/√
3)β0(k)(β1(k)+β2(k))−β1(k)β2(k) ≤0, k= 1,2. Hence, the product type design ξ∗ = ξ1∗⊗ξ2∗ which is thus locally A-optimal (at β = β(1) ⊗β(2)) on the experimental region X = [0,1]4 for model (4.67) assigns weights ξ1∗(x1)ξ2∗(x2) to the point (x>1,x>2)>∈ {(0,0)>,(1,0)>,(0,1)>} × {(0,0)>,(1,0)>,(0,1)>}.
Another result of this type treats the Kiefer Φs-optimality (for notational con-venience we use index s instead of index k) which covers D-optimality at s = 0, A-optimality at s = 1 and E-optimality at s → ∞. Therefore, we obtain the next lemma.
Lemma 4.5.3. For every product type design ξ= NK
k=1
ξk we have the next factorization
Φs(ξ,β) =
K
Y
k=1
Φs(ξk,β(k)), (0≤s <∞) Proof. Employing the definition of Kiefer Φs-criterion yields
Φs(ξ,β) =
1
ptrM−s(ξ,β)
1s
=
1 ptr
K
O
k=1
Mk(ξk,β(k))
−s
1 s
=
K
Y
k=1
1 pktr
M−sk (ξk,β(k))
1 s
=
K
Y
k=1
Φs(ξk,β(k)).
Theorem 4.5.3. Letξk∗ be a locallyΦs-optimal design (atβ(k))for a marginalνk-factor model (4.54) on the experimental region Xk (1 ≤ k ≤ K). Then ξ∗ = NK
k=1
ξk∗ is a locallyΦs-optimal design (at β = NK
k=1
β(k)) for model (4.56) on the experimental region X =×Kk=1Xk.
Proof. The proof is obtained by making use of condition (2.13) of The Equivalence Theorem (Theorem 2.2.2 ). Sinceξk∗ is locally Φs-optimal (at β(k)) we guarantee that
f(k)>
β(k)(xk)M−s−1k (ξk∗,β(k))f(k)
β(k)(xk)≤tr(M−sk (ξk∗,β(k))) for all xk ∈ Xk. Thus in view of Lemma4.5.2 with (4.57) and (4.58) we obtain
f>β(x)M−s−1(ξ∗,β)fβ(x) =u(x,β)
K
O
k=1
f(k)(xk)
> K
O
k=1
Mkξk,β(k)
−s−1 K
O
k=1
f(k)(xk)
=
K
O
k=1
uk(xk,β(k))f(k)>(xk)M−s−1k ξk∗,β(k)f(k)(xk)
=
K
Y
k=1
f(k)>
β(k)(xk)M−s−1k (ξk∗,β(k))f(k)
β(k)(xk)
≤
K
Y
k=1
tr(M−sk (ξ∗k,β(k))) = tr
K O
k=1
Mkξk∗,β(k)
−s
= trM−s(ξ∗,β)
for all x ∈ X. The Equivalence Theorem, thus, proves the local Φs-optimality of the product designξ∗ = NK
k=1
ξk∗ at a parameter point β= NK
k=1
β(k).