This article deals with estimation of finite population mean using the auxiliary proportion under simple and two phase sampling scheme utilizing two auxiliary variables. Mathematical expressions for the mean squared errors of the proposed estimators are derived under first order of approximation. We compare the proposed class of estimators "theoretically and numerically" with the usual mean estimator of Naik and Gupta [
Citation: Xuechen Liu, Muhammad Arslan. A general class of estimators on estimating population mean using the auxiliary proportions under simple and two phase sampling[J]. AIMS Mathematics, 2021, 6(12): 13592-13607. doi: 10.3934/math.2021790
[1] | A. El-Mesady, Y. S. Hamed, M. S. Mohamed, H. Shabana . Partially balanced network designs and graph codes generation. AIMS Mathematics, 2022, 7(2): 2393-2412. doi: 10.3934/math.2022135 |
[2] | Doris Dumičić Danilović, Andrea Švob . On Hadamard 2-(51,25,12) and 2-(59,29,14) designs. AIMS Mathematics, 2024, 9(8): 23047-23059. doi: 10.3934/math.20241120 |
[3] | Menderes Gashi . On the symmetric block design with parameters (280,63,14) admitting a Frobenius group of order 93. AIMS Mathematics, 2019, 4(4): 1258-1273. doi: 10.3934/math.2019.4.1258 |
[4] | Muhammad Sajjad, Tariq Shah, Huda Alsaud, Maha Alammari . Designing pair of nonlinear components of a block cipher over quaternion integers. AIMS Mathematics, 2023, 8(9): 21089-21105. doi: 10.3934/math.20231074 |
[5] | James Daniel, Kayode Ayinde, Adewale F. Lukman, Olayan Albalawi, Jeza Allohibi, Abdulmajeed Atiah Alharbi . Optimised block bootstrap: an efficient variant of circular block bootstrap method with application to South African economic time series data. AIMS Mathematics, 2024, 9(11): 30781-30815. doi: 10.3934/math.20241487 |
[6] | Hye Kyung Kim . Note on r-central Lah numbers and r-central Lah-Bell numbers. AIMS Mathematics, 2022, 7(2): 2929-2939. doi: 10.3934/math.2022161 |
[7] | Cui-Xia Li, Long-Quan Yong . Modified BAS iteration method for absolute value equation. AIMS Mathematics, 2022, 7(1): 606-616. doi: 10.3934/math.2022038 |
[8] | Ahmad Y. Al-Dweik, Ryad Ghanam, Gerard Thompson, M. T. Mustafa . Algorithms for simultaneous block triangularization and block diagonalization of sets of matrices. AIMS Mathematics, 2023, 8(8): 19757-19772. doi: 10.3934/math.20231007 |
[9] | Shakir Ali, Amal S. Alali, Atif Ahmad Khan, Indah Emilia Wijayanti, Kok Bin Wong . XOR count and block circulant MDS matrices over finite commutative rings. AIMS Mathematics, 2024, 9(11): 30529-30547. doi: 10.3934/math.20241474 |
[10] | Abd El-Raheem M. Abd El-Raheem, Mona Hosny . Saddlepoint p-values for a class of nonparametric tests for the current status and panel count data under generalized permuted block design. AIMS Mathematics, 2023, 8(8): 18866-18880. doi: 10.3934/math.2023960 |
This article deals with estimation of finite population mean using the auxiliary proportion under simple and two phase sampling scheme utilizing two auxiliary variables. Mathematical expressions for the mean squared errors of the proposed estimators are derived under first order of approximation. We compare the proposed class of estimators "theoretically and numerically" with the usual mean estimator of Naik and Gupta [
Regular two-level designs are widely used in factorial experiments. When the size of experimental units is large, inhomogeneity of experimental units has bad influences on estimating treatment effects (see, [1,2]). A useful way to reduce such bad influences is to block the experimental units into categories known as blocks. Thus, choosing optimal blocked regular two-level designs becomes an important issue.
As pointed out by [1], there are two kinds of blocking problems. One is called the single block variable problem which involves only one block variable and the other is called the multi block variables problem which considers two or more block variables. In the last decades, most of the literature were concerned with the single block variable problem. Sitter et al. [3], H. Chen and C. S. Cheng [4], R. C. Zhang and D. K. Park [5], and S. W. Cheng and C. F. J. Wu [6] respectively proposed different minimum aberration (MA) criteria which are suitable for selecting blocked designs with single block variable. Under these MA criteria, the construction methods of blocked designs with single block variable were discussed in [7,8,9].
Zhang et al. [10] proposed the general minimum lower-order confounding (GMC) criterion for choosing optimal regular two-level designs. The GMC criterion is preferable when there are some prior knowledge on the importance ordering of treatment effects. R. C. Zhang and R. Mukerjee [11] extended the GMC criterion to blocked designs with single block variable, referred as B-GMC criterion, and gave the construction methods of B-GMC blocked designs via complementary designs. From different considerations, [12] proposed another GMC criterion for blocked designs with single block variable, referred as B1-GMC. Zhao et al. [13] and Zhao et al. [14] studied the construction methods of B1-GMC designs. Zhang et al. [15] proposed multi-stage differential evolution algorithm for constrained D-optimal design. Gashi [16] considered symmetric block design.
Compared to the experiments involving a single block variable, the experiments involving multi block variables are often encountered in practice. As has been mentioned in [1], in the agricultural context, when designs are laid out in rectangular schemes, both row and column inhomogeneity effects probably exist in the soil. Another example of multi block variables problem is from [2], which considers comparing two gasoline additives by testing them on two cars with two drivers over two days. In this experiment, three variables, cars, drivers and days, have to be considered to partition the experimental units.
Despite the wide application background, the multi block variables problem is less studied due to its complexity. In particular, constructing optimal designs with multi block variables is considerably challenging. Under the MA criterion, [17] developed some rules for constructing optimal regular two-level blocked designs with multi block variables. Zhang et al. [18] extended the idea of the GMC criterion to the case of multi block variables problem and developed the blocked GMC criterion, called B2-GMC criterion. Inheriting the advantage of the GMC criterion, a B2-GMC design is particularly preferable when some prior information on importance ordering of treatment effects is present. Zhang et al. [18] tabulated some B2-GMC designs with small numbers of treatment factors and small run sizes by computer search. When n or N is large, computer search becomes computationally challenging, where N=2n−m. Zhao et al. [19] studied the B2-GMC criterion and constructed a small number of B2-GMC designs. In this paper, the B2-GMC designs with the number of treatment factors n all over 5N/16+1≤n≤N−1 are constructed. The construction results cover all that of [19]. The structures of the constructed B2-GMC designs are concise and easy to implement.
The rest of the paper is organized as follows. Section 2 reviews the B2-GMC criterion and introduces some notation. The construction methods of B2-GMC designs are provided in Section 3. Section 4 gives concluding remarks. Some proofs are deferred to Appendix.
Let q=n−m and N=2q. Denote the regular two-level saturated design as
Hq={1,2,12,3,13,23,123,…,123⋯q} |
in Yates order, where the columns 1, 2,…, and q are q independent columns in the form of
1′=(1,−1,1,−1,…,1,−1)2q,2′=(1,1,−1,−1,…,1,1,−1,−1)2q,⋯q′=(1,…,1,−1,…,−1)2q, |
where the superscript of each column denotes transpose, in 1 the entry 1, followed by −1, is repeated 2q−1 times, and in 2 the two successive entries 1's, followed by two successive −1's, are repeated 2q−2 times, …, and in q the 2q−1 successive entries 1's are followed by 2q−1 successive −1's. The remaining columns in Hq are generated by taking the component-wise products of any k of the q independent columns, where k=2,3,…,q. For example, the column 12 is generated by taking component-wise products of the independent columns 1 and 2. Denote H1={1},Hr={Hr−1,r,rHr−1} for r=2,…,q, where rHr−1={rd:d∈Hr−1} and Hr consists of the first 2r−1 columns of Hq. Let Fqr={q,qHr−1} for r=2,…,q, then Fqr consists of the first 2r−1 columns of Fqq and Fqq consists of the last 2q−1 columns of Hq.
Suppose that the inhomogeneity of the units of an experiment comes from s different sources, i.e., s block variables, denoted as b1,b2,…,bs. Suppose the block variable bj has 2lj levels, i.e., the 2n−m units are grouped into 2lj blocks with respect to the block variable bj. Then there should be lj independent columns of Hq to implement such a blocking schedule. Such columns are called block columns in the following. Denote Fj as the collection of the lj independent block columns corresponding to the block variable bj. It is worth noting that the columns in Fj have the following two relations:
(i) the lj columns in Fj are independent of each other for j=1,2,…,s;
(ii) a column from Fj is not necessarily independent of the columns from Fi, i≠j.
Clearly, when s=1, the blocking problem with multi block variables reduces to that with a single block variable. For simplicity, we consider only the case of lj=1 for j=1,2,…,s. Then there are s block columns and each of them blocks the 2n−m experimental units into 2 groups. Here, we would like to emphasize that the s block columns are not necessarily independent of each other.
Throughout the paper, we use D=(Dt,Db) to denote a blocked regular 2n−m:2s design, where Dt is a regular 2n−m design, and Db consists of s block columns. We will not differentiate block variables, block columns, and block factors in the following. B. Tang and C. F. J. Wu [20] introduced the concept of isomorphism which helps to narrow down the search for optimal blocked designs in this paper. An isomorphism ϕ is a one-to-one mapping from Hq to Hq such that ϕ(xy)=ϕ(x)ϕ(y) for every x≠y∈Hq. Two 2n−m:2s designs D1=(D1t,D1b) and D2=(D2t,D2b) are isomorphic if there exists an isomorphism ϕ that maps D1t onto D2t, and D1b onto D2b.
Zhang et al. [18] put forward the effect hierarchy principle for blocked designs with multi block variables as follows:
(i) Lower order treatment factorial effects are more likely to be important than higher order ones, and treatment factorial effects of the same order are equally likely to be important.
(ii) Lower order block factorial effects are more likely to be important than higher order ones, and block factorial effects of the same order are equally likely to be important.
(iii) All the interactions between treatment and block factors are negligible.
With the effect hierarchy principle and weak assumption of effects involving three or more factors are usually not important and negligible, [18] proposed the B2-GMC criterion which considers only the confounding among the main effects and two-factor interactions. As a common assumption in the blocking issues, if a treatment effect is confounded with a potentially significant block effect, the treatment effect cannot be estimated. Thus, confounding of the main effects of treatment factors with any potentially significant block effect is not allowed. In the following, we always suppose the main effects and the two-factor interactions of the block factors are potentially significant.
Denote #1C(p)2(D) as the number of main treatment effects which are aliased with p two-treatment-factor interactions (2tfi's) but not with any potentially significant block effects, where p=0,1,2,…,P, P=n(n−1)/2. Similarly, #2C(p)2(D) denotes the number of 2tfi's which are aliased with the other p 2tfi's but not with any potentially significant block effects. Denote
#1C2(D)=(#1C(0)2(D),#1C(1)2(D),…,#1C(P)2(D)),#2C2(D)=(#2C(0)2(D),#2C(1)2(D),…,#2C(P)2(D)),#C(D)=(#1C2(D),#2C2(D)). | (2.1) |
A blocked design D=(Dt,Db) is called a B2-GMC design if D sequentially maximizes (2.1). Let #iC(p)j(Dt) be the number of i-th order effects which are aliased with p j-th order effects of Dt. Let
#1C2(Dt)=(#1C(0)2(Dt),#1C(1)2(Dt),…,#1C(P)2(Dt)),#2C2(Dt)=(#2C(0)2(Dt),#2C(1)2(Dt),…,#2C(P)2(Dt)),#C(Dt)=(#1C2(Dt),#2C2(Dt)). | (2.2) |
A 2n−m design Dt is called a GMC design if Dt sequentially maximizes (2.2).
Let
U(Db)={γ∈Hq:γ∈Db or γ=ab with a,b∈Db}, |
i.e., U(Db) contains the potentially significant block effects. As aforementioned, confounding between main treatment effects and potentially significant block effects is not allowed which leads to Dt∩U(Db)=∅, the empty set, and consequently #1C(p)2(D)=#1C(p)2(Dt), p=0,1,2,…,P.
As a preparation of deriving B2-GMC designs, we introduce one more piece of notation. For Dt⊂Hq and γ∈Hq, define
B2(Dt,γ)=#{(d1,d2:d1,d2∈Dt,d1d2=γ}, |
where # denotes the cardinality of a set, and d1d2 stands for the two-factor interaction of d1 and d2. Thus, B2(Dt,γ) equals the number of 2tfi's of Dt appearing in the alias set that contains γ.
To construct B2-GMC designs, one should first consider the first part in (2.1), i.e., #1C2(D). Recall that #1C(p)2(D)=#1C(p)2(Dt) for p=0,1,…,P, then choosing D to maximize #1C2(D) reduces to choosing Dt to maximize #1C2(Dt).
A 2n−m design Dt is said to have resolution R if no c-factor interaction is confounded with any other interaction involving less than R−c factors (see, [21]). Note that a 2n−m design Dt with resolution at least IV has #1C(0)2(Dt)=n and #1C(p)2(Dt)=0 for p=1,…P. This implies that a 2n−m design Dt with resolution at least IV must maximize #1C2(Dt). When 5N16+1≤n≤N2, if Dt has resolution at least IV, then Dt⊂Fqq (see, [22]). In the remaining part of this section, we suppose Dt⊂Fqq. By Lemma 1 in [13], to choose Db from Hq, there are two possibilities: (i) Db∩Fqq=∅, and (ii) Db∩Fqq≠∅. As has been pointed out, when constructing B2-GMC designs, there should be Dt∩U(Db)=∅. This leads to the constraint
#{U(Db)∩Fqq}≤N2−n. | (3.1) |
Certainly, for Db in the case (i), U(Db)⊂Hq−1 and thus Db satisfies the constraint (3.1). For Db in the case (ii), there must be U(Db)∩Fqq≠∅ resulting in the necessity to investigate the number of columns in U(Db)∩Fqq. The following lemma addresses this question.
Lemma 1. Let Db be any s-projection of Hq with Db∩Fqq≠∅. If 2k≤s≤2k+1−1 for some k≤q−2, then #{U(Db)∩Fqq}≥2k.
The proof of Lemma 1 is lengthy and thus deferred to Appendix.
Lemmas 2 and 3 below are straightforward extensions of some results in [19] and [23], respectively. These two lemmas are helpful in deriving the construction methods of B2-GMC designs.
Lemma 2. Let Db be any s-projection of Hq with 2k≤s≤2k+1−1 for some k≤q−1.
(i) If Db∩Fqq=∅, then #{U(Db)∩Hq−1}≥2k+1−1 and the equality holds when Db has k+1 independent columns.
(ii) If Db∩Fqq≠∅, then #{U(Db)∩Hq−1}≥2k−1 and the equality holds when Db has k+1 independent columns.
(iii) If Db⊂Hk+1, then U(Db)=Hk+1.
(iv) If Db⊂Hk∪Fq(k+1), then U(Db)=Hk∪Fq(k+1).
Lemma 3. Suppose Dt consists of the last n columns of Hq. For any two columns γ1 and γ2 in Hq−1, if γ1 is ahead of γ2 in Yates order, then B2(Dt,γ1)≥B2(Dt,γ2).
Combining Lemmas 1, 2 and 3, the following theorem provides the constructions of B2-GMC designs with 5N16+1≤n≤N2, where N≥16.
Theorem 1. Suppose D=(Dt,Db) is a 2n−m:2s design with 2r≤N2−n≤2r+1−1 for some r≤q−3 and 2k≤s≤2k+1−1 for some k≤q−2. The design D=(Dt,Db) is a B2-GMC design if Dt consists of the last n columns of Fqq and
(i) Db is any s-projection of Hk∪Fq(k+1) when 1≤k≤r,
(ii) Db is any s-projection of Hk+1 when r+1≤k≤q−2.
Proof. Let D∗=(D∗t,D∗b) be a 2n−m:2s design with D∗t⊂Fqq and U(D∗b)⊂Hq∖D∗t. According to Lemma 1 of [22],
B2(D∗t,γ)={0, if γ∈Fqq,B2(ˉD∗t,γ)+n−N4, if γ∈Hq−1, |
where ˉD∗t=Fqq∖D∗t. Note that Dt has resolution at least IV, then #1C2 is maximized by D. Therefore, we consider only #2C2 in (2.1) in the following.
For (i). From Theorem 2 of [22], Dt is a GMC 2n−m design and thus Dt maximizes (2.2) among all D∗t. From Lemma 2 (iv), if Db is any s-projection of Hk∪Fq(k+1), then U(Db)=Hk∪Fq(k+1). Suppose γ0 is the last column of Hk in Yates order and B2(Dt,γ0)=p0, where p0≥n−N4≥N16+1≥2. From Lemma 3, for any γ∈Hk, we have B2(Dt,γ)≥p0. Since Dt⊂Fqq, we have B2(Dt,γ)=0 for any γ∈Fqq. By the definition of #2C(p)2(D), when p≤p0−2, we have 1≤p+1≤p0−1 and
#2C(p)2(D)=(p+1)#{γ∈Hq:γ∉U(Db),B2(Dt,γ)=p+1}=(p+1)#{γ∈Hq:B2(Dt,γ)=p+1}=#2C(p)2(Dt). | (3.2) |
From Lemma 3, for any γ∈Hq∖Hk, we have B2(Dt,γ)≤p0. Therefore, when p≥p0, we have
#2C(p)2(D)=(p+1)#{γ∈Hq:γ∉U(Db),B2(Dt,γ)=p+1}=(p+1)#{γ∈(Hq−1∖Hk)∪(Fqq∖Fq(k+1)):B2(Dt,γ)=p+1}=0. | (3.3) |
From (3.2), we obtain that D sequentially maximizes
(#2C(0)2(D),#2C(1)2(D),…,#2C(p0−2)2(D)) |
among all D∗ since Dt is a GMC design.
Suppose D is not a B2-GMC design, then there exists a D∗ and some p1≥p0−1 such that
(#2C(0)2(D),#2C(1)2(D),…,#2C(p1−1)2(D))=(#2C(0)2(D∗),#2C(1)2(D∗),…,#2C(p1−1)2(D∗)), | (3.4) |
and
#2C(p1)2(D∗)>#2C(p1)2(D). | (3.5) |
Recall the definitions of #2C(p)2(D) and #2C(p)2(Dt), we have
1p+1#2C(p)2(D)=#{γ∈Hq:γ∉U(Db),B2(Dt,γ)=p+1}=#{γ∈Hq−1:B2(Dt,γ)=p+1}−#{γ∈Hk:B2(Dt,γ)=p+1}, | (3.6) |
where the second equality is due to B2(Dt,γ)=0 for any γ∈Fqq. From (3.6), it is obtained that
P∑p=01p+1#2C(p)2(D)=P∑p=0#{γ∈Hq−1:B2(Dt,γ)=p+1}−P∑p=0#{γ∈Hk:B2(Dt,γ)=p+1}=#{Hq−1∖Hk}=2q−1−2k. | (3.7) |
By (3.2), (3.3) and (3.7), it is obtained that
P∑p=01p+1#2C(p)2(D)=p0−2∑p=01p+1#2C(p)2(D)+1p0#2C(p0−1)2(D)=2q−1−2k. | (3.8) |
Similarly, for D∗ we have
P∑p=01p+1#2C(p)2(D∗)=P∑p=0#{γ∈Hq−1:B2(D∗t,γ)=p+1}−P∑p=0#{γ∈U(D∗b)∩Hq−1:B2(D∗t,γ)=p+1}=2q−1−1−#{U(D∗b)∩Hq−1}. | (3.9) |
From (3.2)–(3.5), we obtain
P∑p=01p+1#2C(p)2(D)=p1−1∑p=01p+1#2C(p)2(D)+1p1+1#2C(p1)2(D)<p1−1∑p=01p+1#2C(p)2(D∗)+1p1+1#2C(p1)2(D∗)≤p1−1∑p=01p+1#2C(p)2(D∗)+P∑p=p11p+1#2C(p)2(D∗)=P∑p=01p+1#2C(p)2(D∗). |
Then it leads to #{U(D∗b)∩Hq−1}<2k−1 from Eqs (3.8) and (3.9). This contradicts Lemma 2 (ii) and completes the proof of (i).
For (ii). From Lemma 1, if D∗b∩Fqq≠∅, then #{U(D∗b)∩Fqq}≥2k≥2r+1 which implies U(D∗b)∩D∗t≠∅. This is not allowed as has been pointed out. Therefore, if r+1≤k≤q−2, there should be D∗b⊂Hq−1. According to Lemma 2 (iii), if Db is an s-projection of Hk+1, then U(Db)=Hk+1. The remainder of the proof is similar to that of (i) and omitted. This completes the proof.
In the following, an example is provided to illustrate the constructions of B2-GMC 2n−m:2s designs with 5N16+1≤n≤N2.
Example 1. Consider constructing B2-GMC 212−7:22 and 212−7:29 designs. For both B2-GMC designs to be constructed, we have q=n−m=5, N=32 and r=2 as 22≤N2−n≤23−1. The values of the parameters N and n satisfy 5N16+1≤n≤N2. Therefore, to construct these two B2-GMC designs, Dt should be the last 12 columns of F55.
For the case 212−7:22, we have s=2 which gives k=1. Therefore, we should choose 2 block columns according to Theorem 1 (i) as k<r. Without loss of generality, let Db1={1,5} be the 2-projection of H1∪F52. Then D=(Dt,Db1) is a B2-GMC 212−7:22 design.
For the case 212−7:29, we have s=9 which gives k=3. Therefore, we should choose 9 block columns according to Theorem 1 (ii) as k>r. Without loss of generality, let Db2={1,2,12,3,13,23,123,4,14} be the 9-projection of H4. Then D=(Dt,Db2) is a B2-GMC 29−4:29 design.
Similar to the discussion in the first paragraph of Section 3.1, when constructing B2-GMC designs with n>N2, we should also first maximize #1C2(Dt). Suppose the number of columns in Hq∖Dt satisfies 2r≤N−1−n≤2r+1−1 for some r≤q−2. According to the first paragraph of Section 3.2 in [22], when n>N2, if Dt maximizes #1C2(Dt), then Hq∖Dt⊂Hr+1 up to isomorphism. This implies that Db⊂Hr+1.
Suppose 2k≤s≤2k+1−1 for some k≤q−2. According to Lemma 2 (i) and (ii) combined with Lemma 1, we have #U(Db)≥2k+1−1 no matter Db∩Fqq=∅ or Db∩Fqq≠∅. Therefore, there should be k=r with N−1−n=2k+1−1 or k<r, otherwise U(Db)∩Dt≠∅. For the case of k=r with N−1−n=2k+1−1, it is trivial since Dt=Hq∖Hk+1 and thus Db⊂Hk+1. This obtains that the design D=(Dt,Db) with Dt=Hq∖Hk+1 and Db being any s-projection of Hk+1 is a B2-GMC design. The following theorem considers the constructions of B2-GMC designs for the case of k<r.
Theorem 2. Suppose D=(Dt,Db) is a 2n−m:2s design with 2r≤N−1−n≤2r+1−1 for some r≤q−2 and 2k≤s≤2k+1−1 for some k<r. The design D=(Dt,Db) is a B2-GMC design if Dt consists of the last n columns of Hq and Db is any s-projection of Hk+1.
Proof. Let D∗=(D∗t,D∗b) be a 2n−m:2s design with Hq∖D∗t⊂Hr+1. Then we have U(D∗b)⊂Hq∖D∗t. From Lemma 3 of [22],
B2(D∗t,γ)={n−N/2, if γ∈Hq∖Hr+1,B2(ˉD∗t,γ)+N2−2r, if γ∈Hr+1, |
where ˉD∗t=Hq∖D∗t.
By Lemma 2 (iv), if Db is any s-projection of Hk+1, then U(Db)=Hk+1. Suppose γ0 is the last column of Hk+1 in Yates order and B2(Dt,γ0)=p0, where p0≥N2−2r>n−N2≥1. From Lemma 3, if γ∈Hk+1, then B2(Dt,γ)≥p0. By the definition of #2C(p)2(D), when p≤p0−2, we have 1≤p+1≤p0−1 and
#2C(p)2(D)=(p+1)#{γ∈Hq:γ∉U(Db),B2(Dt,γ)=p+1}=(p+1)#{γ∈Hq:B2(Dt,γ)=p+1}=#2C(p)2(Dt). | (3.10) |
From Lemma 3, for any γ∈Hq∖Hk+1, we have B2(Dt,γ)≤p0. Therefore, when p≥p0, we have
#2C(p)2(D)=(p+1)#{γ∈Hq:γ∉U(Db),B2(Dt,γ)=p+1}=0. | (3.11) |
Since Dt consists of the last n columns of Hq, from Theorem 3 of [22], Dt is a GMC 2n−m design. Then Dt sequentially maximizes (2.2). Recall that #1C(p)2(D)=#1C(p)2(Dt) for p=0,1,…,P. Thus, (3.10) implies that D maximizes
(#1C(0)2(D),#1C(1)2(D)…,#1C(P)2(D),#2C(0)2(D),#2C(1)2(D)…,#2C(p0−2)2(D)) |
among all D∗.
Suppose D is not a B2-GMC design, then there exists a D∗ and some p1≥p0−1 such that
(#1C(0)2(D),#1C(1)2(D),…,#1C(P)2(D),#2C(0)2(D),#2C(1)2(D)…,#2C(p1−1)2(D))=(#1C(0)2(D∗),#1C(1)2(D∗),…,#1C(P)2(D∗),#2C(0)2(D∗),#2C(1)2(D∗),…,#2C(p1−1)2(D∗)) |
and #2C(p1)2(D∗)>#2C(p1)2(D). With a similar argument to the proof of Theorem 1 (i), such a D∗ results in #{U(D∗b)∩Hq−1}<2k+1−1 which contradicts Lemma 2 (i). This completes the proof.
In the following, an example is provided to illustrate the constructions of B2-GMC 2n−m:2s designs with n>N2.
Example 2. Consider constructing B2-GMC 29−5:22 and 212−8:23 designs.
For the case 29−5:22, we have q=n−m=4, N=16 and r=2 as 22≤N−n−1≤23−1. Since s=2, we obtain k<r as k=1. According to Theorem 3.2, let Dt1 be the last 9 columns of H4, and Db1={1,2} be a 2-projection of H2. Then D=(Dt1,Db1) is a B2-GMC 29−5:22 design.
For the case 212−8:23, we have q=n−m=4, N=16 and r=1 as 21≤N−n−1≤22−1. Since s=3, we obtain k=r as k=1. As discussed in the second paragraph in Section 3.2, let Dt2 be the last 12 columns of H4, and Db2={1,2,12}=H2. Then D=(Dt2,Db2) is a B2-GMC 212−8:23 design.
Regular two-level factorial designs are widely used in factorial experiments. Inhomogeneity of the units has bad influences on estimating the treatment effects when size of experimental units is large. To reduce such bad influences, a useful way is to block the experimental units into categories. As has been pointed out in [1], there are two types of blocking problems. One is the single block variable problem and the other is the multi block variables problem. In the last decades, the single block variable problem was maturely investigated in the literature.
As has been exemplified in Section 1, multi block variables problem is more widely encountered in practice compared to the blocking problem with a single block variable. However, the studies on multi block variables problem are relatively primitive. The GMC criterion is welcome in the situations where the importance ordering of treatment effects is present. Zhang et al. [18] proposed the B2-GMC criterion for choosing optimal blocked regular two-level designs. Construction methods on B2-GMC designs can only be found in [19] in which the B2-GMC designs of some n from 5N16+1≤n≤N2 are constructed. In this paper, the B2-GMC designs of n all over 5N16+1≤n≤N−1 are systemically constructed. The structures of the constructed B2-GMC designs are concise and easy to implement.
This work was supported by the National Natural Science Foundation of China (Grant No. 11801331).
The authors declare that there is no conflict of interest.
We only need to prove #{U(Db)∩Fqq}≥2k for the case of s=2k. Recall that
U(Db)=Db∪{γ∈Hq:γ=ab,a,b∈Db}. |
We have
U(Db)∩Fqq=(Db∩Fqq)∪({γ∈Hq:γ=ab,a,b∈Db}∩Fqq). |
Denote A=Db∩Fqq and E=Db∩Hq−1. Then Db=A∪E, A∩E=∅,
{γ∈Hq:γ=ab,a,b∈Db}∩Fqq=A⊗E, |
and
U(Db)∩Fqq=A∪A⊗E, |
where A⊗E={γ∈Hq:γ=ab,a∈A,b∈E}. Thus, it suffices to prove
#{A∪A⊗E}≥2k | (A.1) |
for #A+#E=2k.
Regarding to the columns in A and E, there are two cases:
(B1) #A>#E, or
(B2) #A≤#E.
The global line of the remaining proof for Lemma 1 is as follows. In Lemma A.1, we first show that if (A.1) holds for the case (B1), then (A.1) holds for the case (B2). Afterwards, with Lemma A.2–A.5, we prove that (A.1) indeed holds for the case (B1).
Lemma A.1. Suppose that (A.1) holds for the case (B1), then (A.1) holds for the case (B2).
Proof. For the A and E in the case (B2), without loss of generality, we suppose q∈A. Then,
#{A∪A⊗E}=#{A⊗({IN}∪E)}=#{(qA)⊗({q}∪qE)}=#{({IN}∪˜A)⊗˜E}=#{˜E∪˜E⊗˜A}, |
where ˜A=q(A∖q)⊂Hq−1 and ˜E={q}∪qE⊂Fqq.
Note that #˜A=#A−1≤2k−1−1, #˜E=#E+1≥2k−1+1 and #˜A+#˜E=2k, then #˜E>#˜A and the case (B2) converts into the case (B1). Therefore, if (A.1) holds for the case (B1), then (A.1) holds for the case (B2). This completes the proof.
In the following, we only need to prove that (A.1) holds for the case (B1).
For A in the case (B1), we have #A≥2k−1+1. Then, A has at least k+1 independent columns. We suppose A has h+1(≥k+1) independent columns. Let e denote the hth independent column in Hq in Yates order. Up to isomorphism, A can be expressed as
A=A1∪{ea1,ea2,…,eav}, | (A.2) |
where A1 has h independent columns with A1⊂Fqh and {a1,a2,…,av}⊂Fqh.
For E in the case (B1), if #E=0, then #A=2k and (A.1) holds. In the following, we consider only 1≤#E≤2k−1−1. Up to isomorphism, there are three cases for the columns in E:
(C1) all the columns are from Hh−1;
(C2) some columns are from Hh−1 and the others are from Hh∖Hh−1;
(C3) some columns are from Hh−1, some are from Hh∖Hh−1 and the others are from Hq∖Hh.
We first consider the case (C2) with h>k. Note that A∪A⊗E=A⊗({IN}∪E). Denote B={IN}∪E. Then B can be represented as
B=B1∪{eb1,eb2,…,ebw}, | (A.3) |
where IN∈B1, B1∖{IN}⊂Hh−1 and {b1,…,bw}⊂Hh−1.
Recall that 2k−1<#A≤2k−1<2h−1, we can always find a column r∈Hh−1 or r=IN such that at least one column in {ea1,ea2,…,eav}, say ea1, satisfies (re)(ea1)=ra1∈Fqh∖A1. Without loss of generality, we assume that there is some t1(1≤t1≤v) such that re{ea1,…,eat1}⊂Fqh∖A1 and re{eat1+1,…,eav}⊂A1. Meanwhile, there is some t2(0≤t2≤w) such that re{eb1,…,ebt2}⊂Hh−1∖B1 and re{ebt2+1,…,ebw}⊂B1.
Let A2={eat1+1,…,eav} and A3={ea1,…,eat1}. Then,
A=A1∪A2∪A3. | (A.4) |
Let B2={ebt2+1,…,ebw} and B3={eb1,…,ebt2}. Then,
B=B1∪B2∪B3. | (A.5) |
Let
A∗=A1∪A2∪A∗3 | (A.6) |
and
B∗=B1∪B2∪B∗3, | (A.7) |
where A∗3=reA3 and B∗3=reB3.
Lemma A.2. Suppose A,B,A∗ and B∗ are defined as in (A.4)–(A.7), respectively, then #{A∗⊗B∗}≤#{A⊗B}.
Proof. Let
Q1=(A1⊗B2)∪(A2⊗B1),Q2=(A1⊗B1)∪(A3⊗B3),Q3=(A1⊗B3)∪(A3⊗B1),Q4=(A2⊗B3)∪(A3⊗B2),Q∗3=(A∗3⊗B2)∪(A2⊗B∗3),Q∗4=(A1⊗B∗3)∪(A∗3⊗B1). |
Since reA2⊂A1 and reB2⊂B1, we have A2⊗B2⊂Q2. Thus
A⊗B=Q1∪Q2∪(A2⊗B2)∪Q3∪Q4=Q1∪Q2∪Q3∪Q4. |
Since Q1∪Q3⊂Fq(h+1)∖Fqh and Q2∪Q4⊂Fqh are mutually exclusive, thus
#{A⊗B}=#{Q1∪Q3}+#{Q2∪Q4}=#Q1+#Q2+#Q3+#Q4−#{Q1∩Q3}−#{Q2∩Q4}. |
Note that A∗3⊗B∗3=A3⊗B3, then
A∗⊗B∗=Q1∪Q2∪(A2⊗B2)∪Q∗3∪Q∗4=Q1∪Q2∪Q∗3∪Q∗4. |
Since Q1∪Q∗3⊂Fq(h+1)∖Fqh and Q2∪Q∗4⊂Fqh are mutually exclusive, thus
#{A∗⊗B∗}=#{Q1∪Q∗3}+#{Q2∪Q∗4}=#Q1+#Q2+#Q∗3+#Q∗4−#{Q1∩Q∗3}−#{Q2∩Q∗4}=#Q1+#Q2+#Q3+#Q4−#{Q1∩Q∗3}−#{Q2∩Q∗4}, |
where the third equality is due to reQ∗3=Q4 and reQ∗4=Q3. Therefore, to prove Lemma A.2, it suffices to prove
#{Q1∩Q∗3}+#{Q2∩Q∗4}≥#{Q1∩Q3}+#{Q2∩Q4} |
or equivalently
#{Q2∩Q∗4}−#{Q2∩Q4}≥#{Q1∩Q3}−#{Q1∩Q∗3}. | (A.8) |
Note that A2⊂reA1 and B2⊂reB1, then
A2⊗B3⊂(reA1)⊗B3=A1⊗B∗3 |
and
A3⊗B2⊂A3⊗(reB1)=(reA3)⊗B1=A∗3⊗B1 |
which implies that Q4⊂Q∗4. Similarly, we can obtain Q∗3⊂Q3. Therefore, (A.8) is equivalent to
#{(Q2∩Q∗4)∖(Q2∩Q4)}≥#{(Q1∩Q3)∖(Q1∩Q∗3)}. | (A.9) |
Thus, we only need to prove (A.9).
For the left hand side of (A.9), we have
#{(Q2∩Q∗4)∖(Q2∩Q4)}=#{Q2∩(Q∗4∖Q4)}. |
For the right hand side of (A.9), we have
#{(Q1∩Q3)∖(Q1∩Q∗3)}=#{Q1∩(Q3∖Q∗3)}=#{(reQ1)∩(reQ3∖reQ∗3)}=#{(reQ1)∩(Q∗4∖Q4)}. |
Since reQ1=(A1⊗(reB2))∪((reA2)⊗B1)⊂A1⊗B1, we have reQ1⊂Q2. Then (A.9) holds. This completes the proof of Lemma A.2.
Remark 1. Lemma A.2 indicates that for any A defined in (A.2) and B defined in (A.3), we can always find A∗, which has less columns out of Fqh than A, and B∗, which has no more columns out of Hh−1 than B, such that #{A∗⊗B∗}≤#{A⊗B}. Repeatedly applying Lemma A.2, we can finally find A∗∗⊂Fqh and B∗∗, which has no more columns out of Hh−1 than B, such that #{A∗∗⊗B∗∗}≤#{A⊗B}.
For simplicity of notation, we still denote A∗∗ as A and B∗∗ as B. Then we can assume that A⊂Fqh. Note that there might be t2=0 in the procedure above. Then, following Remark 1, B has the following cases:
(D1) B∖{IN}⊂Hh−1, or
(D2) (B∖{IN})∩(Hh∖Hh−1)≠∅,
For the case (D2), we write B as B=B1∪{eb1,…,ebt3}, where IN∈B1, B1∖{IN}⊂Hh−1 and {b1,…,bt3}⊂Hh−1. Note that 1≤#(B∖{IN})≤2k−1−1<2h−1−1. We can always find a column r1∈Hh−1 or r1=IN such that at least one column in {eb1,…,ebt3}, say eb1, satisfies (r1e)(eb1)=r1b1∈Hh−1∖B1. Without loss of generality, suppose there is some t4 with 1≤t4≤t3 such that r1e{eb1,…,ebt4}⊂Hh−1∖B1 and r1e{ebt4+1,…,ebt3}⊂B1. Denote B2={ebt4+1,…,ebt3} and B3={eb1,…,ebt4}, then
B=B1∪B2∪B3. | (A.10) |
Denote
B∗=B1∪B2∪B∗3, | (A.11) |
where B∗3=r1eB3.
Lemma A.3. Suppose A⊂Fqh, B and B∗ are defined as in (A.10) and (A.11), respectively, then #{A⊗B∗}≤#{A⊗B}.
Proof. Note that A⊗B1⊂Fqh and (A⊗B2)∪(A⊗B3)⊂Fq(h+1)∖Fqh. Therefore,
#{A⊗B}=#{A⊗B1}+#{(A⊗B2)∪(A⊗B3)}=#{A⊗B1}+#{A⊗B2}+#{A⊗B3}−#{(A⊗B2)∩(A⊗B3)}. |
Since (A⊗B1)∪(A⊗B∗3)⊂Fqh and A⊗B2⊂Fq(h+1)∖Fqh, we have
#{A⊗B∗}=#{(A⊗B1)∪(A⊗B∗3)}+#{A⊗B2}=#{A⊗B1}+#{A⊗B2}+#{A⊗B∗3}−#{(A⊗B1)∩(A⊗B∗3)}=#{A⊗B1}+#{A⊗B2}+#{A⊗B3}−#{(A⊗B1)∩(A⊗B∗3)}, |
where the third equality is due to r1e(A⊗B∗3)=A⊗B3. On one hand,
#{(A⊗B1)∩(A⊗B∗3)}=#{r1e((A⊗B1)∩(A⊗B∗3))}=#{(A⊗(r1eB1))∩(A⊗B3)}. | (A.12) |
On the other hand, B2⊂r1eB1, which leads to
(A⊗B2)∩(A⊗B3)⊂(A⊗(r1eB1))∩(A⊗B3). | (A.13) |
From (A.12) and (A.13), we obtain that
#{(A⊗B1)∩(A⊗B∗3)}≥#{(A⊗B2)∩(A⊗B3)}. |
This implies #{A⊗B∗}≤#{A⊗B} and completes the proof.
Remark 2. By repeatedly applying Lemma A.3, we can finally find B∗⊂Hh−1 such that #{A⊗B∗}≤#{A⊗B}. This result is also true for the cases (C1) and (C3) with h>k due to the following reasons. When {ea1,ea2,…,eav}=∅, the case (C2) reduce to the case (C1). For the case (C3), with a similar argument to Lemma A.3, we can find a T with IN⊂T and (T∖IN)⊂Hh such that #{A⊗T}≤#{A⊗B}. Then the case (C3) reduces to case (C2).
The following remark considers the case of h=k.
Remark 3. For h=k, up to isomorphism, A⊂Fq(k+1). In this situation, h should equal to k in the cases (C1), (C2) and (C3). Especially, in the cases (C1) and (C2), E is already a subset of Hk. By repeatedly applying Lemma A.3 to E in the case (C3), we can find a set, say P, such that P⊂Hk and #{A⊗({IN}∪P)}≤#{A⊗({IN}∪E)}.
In summary, for any A and B defined in (A.2) and (A.3), by repeatedly applying Lemma A.2 and A.3, we can always find A∗⊂Fq(k+1) and B∗∖{IN}⊂Hk with IN∈B∗, such that #{A∗⊗B∗}≤#{A⊗B}. Next, we denote A∗ as A and B∗ as B and prove that #{A⊗B}=2k for any A⊂Fq(k+1), B∖{IN}⊂Hk with IN∈B and #A+#B=2k. We first introduce a useful lemma from [24].
Denote IS as the set consisting of the distinct columns generated by taking component-wise products of any two columns of S.
Lemma A.4. Let S be an s-subset of Fqq, s=2k−1+δ≥3 and 0<δ≤2k−1, then #IS≥2k−1, where the equality holds when the number of independent columns of S is k+1.
Lemma A.5. Suppose A⊂Fq(k+1), IN∈B, B∖IN⊂Hk with #A+#B=2k+1 and k≤q−2, then #{A⊗B}=2k.
Proof. Without loss of generality, suppose e is the (k+1)th independent column in Hq. Then #{A⊗B}=#{A⊗(e\bmqB)} and e\bmqB⊂Fq(k+2)∖Fq(k+1). Next, we show #{A⊗(e\bmqB)}=2k. Since A∩e\bmqB=∅, we have #{A∪e\bmqB}=#A+#{e\bmqB}=2k+1 and A∪e\bmqB has k+2 independent columns. By Lemma A.4, we have
#IA∪e\bmqB=#{IA∪I\bmeqB∪(A⊗(eqB))}=2k+1−1. | (A.14) |
Note that IA∪IeqB⊂Hk then #{IA∪IeqB}≤2k−1, and A⊗(eqB)⊂Hk+1∖Hk then #{A⊗(eqB)}≤2k. From (A.14), there should be #{IA∪IeqB}=2k−1 and #{A⊗(eqB)}=2k. This completes the proof.
Proof of Lemma 1. According to the proofs of Lemma A.2, A.3 and A.5, we can immediately obtain that #{U(Db)∩Fqq}≥2k. This completes the proof.
[1] | V. D. Naik, P. C. Gupta, A note on estimation of mean with known population proportion of an auxiliary character, J. Indain. Soc. Agric. Stat., 48 (1996), 151–158. |
[2] | H. S. Jhajj, M. K. Sharma, L. K. Grover, A family of estimators of population mean using information on auxiliary attribute, Pak. J. Stat., 48 (2006), 43. |
[3] | A. M. Abd-Elfattah, E. A El-Sherpieny, S. M Mohamed, O. F Abdou, Improvement in estimating the population mean in simple random sampling using information on auxiliary attribute, Appl. Math. Comput., 215 (2010), 4198–4202. |
[4] | N. Koyuncu, Efficient estimators of population mean using auxiliary attributes, Appl. Math. Comput., 218 (2012), 10900–10905. |
[5] | R. S. Solanki, H. P. Singh, Improved estimation of population mean using population proportion of an auxiliary character. Chil. J. Stat., 4 (2013), 3–17. |
[6] | P. Sharma, H. K. Verma, A. Sanaullah, R. Singh, Some exponential ratio-product type estimators using information on auxiliary attributes under second order approximation, Int. J. Stat. Econ., 12 (2013), 58–66. |
[7] | S. Malik, R. Singh, An improved estimator using two auxiliary attributes, Appl. Math. Comput., 219 (2013), 10983–10986. |
[8] | H. Verma, R. Singh, F. Smarandache, Some improved estimators of population mean using information on two auxiliary attributes, In: On improvement in estimating population parameter(s) using auxiliary information, Columbus: Educational Publishing, Beijing: Journal of Matter Regularity, 2013, 17–24. |
[9] | R. S. Solanki, H. P. Singh, S. K. Pal, Improved estimation of finite population mean in sample surveys, Columbia Int. Publ. J. Adv. Comput., 1 (2013), 70–78. |
[10] |
P. Sharma, R. Singh, Improved ratio type estimator using two auxiliary variables under second order approximation, Math. J. Interdiscip. Sci., 2 (2014), 179–190. doi: 10.15415/mjis.2014.22014
![]() |
[11] | M. Mahdizadeh, E. Zamanzade, Kernel-based estimation of p (x > y) in ranked set sampling, SORT-Stat. Oper. Res. T., 40 (2016), 243–266. |
[12] |
H. P Singh, S. K. Pal, R. S. Solanki, A new class of estimators of finite population mean in sample surveys, Commun. Stat. Theor. Methods, 46 (2017), 2630–2637. doi: 10.1080/03610926.2015.1030429
![]() |
[13] |
M. Mahdizadeh, E. Zamanzade, Smooth estimation of a reliability function in ranked set sampling, Statistics, 52 (2018), 750–768. doi: 10.1080/02331888.2018.1477157
![]() |
[14] |
S. Hussain, S. Ahmad, S. Akhtar, A. Javed, U. Yasmeen, Estimation of finite population distribution function with dual use of auxiliary information under non-response, PloS One, 15 (2020), e0243584. doi: 10.1371/journal.pone.0243584
![]() |
[15] |
S. Al-Marzouki, C. Chesneau, S. Akhtar, J. A. Nasir, S. Ahmad, S. Hussain, et al., Estimation of finite population mean under pps in presence of maximum and minimum values, AIMS Mathematics, 6 (2021), 5397–5409. doi: 10.3934/math.2021318
![]() |
[16] |
B. Kiregyera, A chain ratio-type estimator in finite population double sampling using two auxiliary variables, Metrika, 27 (1980), 217–223. doi: 10.1007/BF01893599
![]() |
[17] | S. Mohanty, J. Sahoo, A note on improving the ratio method of estimation through linear transformation using certain known population parameters, Sankhyā: Indian J. Stat. Ser. B, 1995, 93–102. |
[18] | A. Haq, J. Shabbir. An improved estimator of finite population mean when using two auxiliary attributes, Appl. Math. Comput., 241 (2014), 14–24. |
[19] | M. N. Murthy, Sampling theory and methods, Florida: CRC Press LLC, 1967. |
[20] | S. Singh, Advanced sampling theory with applications, Springer Science and Business Media, 2003. |
[21] | A. Sharmin, J. R. Sarker, K. R. Das, Growth and trend in area, production and yield of major crops of Bangladesh. Int. J. Econ. Financ. Manage. Sci., 4 (2016), 20–25. |
1. | Yuna Zhao, Gengxin Sun, General Minimum Lower-Order Confounding Designs with Multi-Block Variables, 2021, 2021, 1563-5147, 1, 10.1155/2021/5548102 |