385
Views
0
CrossRef citations to date
0
Altmetric
Articles

On the construction of balanced repeated measurements designs with good circular properties

, &
Pages 121-129 | Received 26 Feb 2022, Accepted 20 Feb 2023, Published online: 13 Mar 2023

Abstract

Several fields, such as biological, medical, public health, agricultural sciences, etc., require circular balanced repeated measurement designs with fewer unequal number of repeated measurements than the number of treatments. Also, the availability and high cost of experimental subjects in these fields prefer the design in fewer experimental units. However, balancing the carryover effects of the treatments in minimal experimental subjects is one of the problems in this case. In this paper, several new series of minimal circular nearly strongly balanced RMDs in periods of two and three different sizes are constructed. The proposed construction of designs has high efficiency and, therefore, can save the cost of experimentations due to a fewer experimental subjects. Most of the designs are very useful because of the unavailability of strongly balanced RMDs for these combinations of parameters. A list of sets of shifts for the construction of minimal circular nearly SBRMDs has also been mentioned in the Appendix.

1. Introduction

Experimental design deals with the arrangement of experimental units and the assignment of treatments to them in such a way that the comparisons among the treatments are unbiased and as precise as possible. The precision of the experiment depends to a large extent on the size of the experiment and the variations present in the experimental units. In designing experiments, keeping in view the limitation on experimental resources, efforts are made to minimize standard errors of estimated contrasts of interest. Mostly error variation arises because of the variation in experimental units under the same treatments. Therefore, it is suggested to use the same experimental unit for different treatments in an experiment but in different periods. This results in a type of design known as repeated measurements design (RMD) which makes them different from other designs presented in the literature which make use of different experimental units for different treatments. Repeated measurements design (RMD) is the type of experimental design in which experimental subjects are repeatedly measured by giving a sequence of treatments. RMDs are a special kind of carryover design in which each subject is influenced by a treatment applied in a current period called treatment effects and treatment applied in a previous period called carryover effects. Cross-over designs are RMDs in which each unit receives the sequence of each treatment over different periods. RMDs in unequal period sizes are very useful if there is a restriction on the total number of treatments, and some experimental units can receive on the total length of time, while some experimental units can remain in the trial which results in RMDs of different or unequal period sizes. RMDs balanced in treatments and carryover effects are useful in several fields such as medicine, pharmacology, animal sciences and psychology where carryover effects are natural.

2. Literature review

Cheng and Wu (Citation1980) proposed the construction of BRMDs and CSBRMDs, especially for unequal block sizes. With the help of advanced technology in computer systems, the uses of block designs of unequal sizes k are beneficial, and hence it is very useful in industrial and agricultural experiments. The utilization of different block sizes in biological experiments was mentioned by Pearce (Citation1964). Afsarinejad (Citation1994) constructed minimal CSBRMDs for different period sizes. Iqbal and Jones (Citation1994) proposed the construction of efficient RMDs and CSBs for two different period sizes. Iqbal and Tahir (Citation2009), Iqbal et al. (Citation2010) and Bashir et al. (Citation2018) constructed CSBs for some classes. Rasheed et al. (Citation2018) recently developed some generators to obtain MCSBs in a period of three different sizes for some cases of t, where 2 ≤ k3 < k2 ≤ 10. However, there do not exist MCSBs for certain combinations of treatments and periods when several treatments are less than the number of periods. Here, minimal circular nearly CSBs are useful. A design is said to be CNBs (Circular nearly strongly balanced repeated measurement designs) if every treatment is immediately followed by every other treatment including itself except the treatment (labelled as t − 1) which is not preceded by itself. In this article, some generators are developed to construct CNBs in periods of (i) two different sizes and (ii) three different sizes. These designs have their own importance in extended scientific and biological experimentations with different block sizes which consist of clinical experimentations of human and also animal behaviour responses for the comparison of several non-curative treatments for their effectiveness. With the advancement of technological computer systems, block designs with unequal sizes p have become very useful in huge industrial and agricultural experiments. In biological experiments, the utilization of different block sizes has been studied significantly. RCTs become prone to the unravelling of sizes of the block when it remains the same throughout the trial, so investigators use different block sizes, i.e. they randomly vary the block size to lower the chances of unbiasedness (Schulz & Grimes, Citation2002; Schulz, Citation1995). There exist different scientific research areas such as animal husbandry and genetic experimentation. RCTs, where these designs are applied as within-subject treatment comparisons, are more efficient than between-subject treatment comparisons. These proposed designs are additions to the literature as they are incorporating different period sizes and also possess good efficiency in estimating direct and residual effects.

Important definitions with abbreviations used that will be discussed throughout the article are discussed below briefly.

BRMDs=

Balanced repeated measurement designs

CSBs=

Circular strongly balanced repeated measurement designs

CNBs=

Circular nearly strongly balanced repeated measurement designs

2.1. BRMDs

In an RMD each experimental subject or unit receives a series of several treatments in successive periods.

2.2. CSBs

If every treatment is directly preceded by every other treatment (including itself) then it is called CSBs.

2.3. CNBs

In this type of design, every treatment is immediately followed by every other treatment exactly once as well as with itself except the treatment labelled as v−1, which is not preceded by itself.

In Section 3, the model and information matrix are described along with the formula for the computation of the efficiency for CNBs. In Section 4, the Method of Cyclic Shifts (MOCS) is described to generate CNBs in different period sizes. A series of generators is proposed in Section 5 to obtain CNBs in two different sizes of periods. Also, a series of generators for the construction of CNBs in three different sizes of periods are proposed in Section 6. Section 7 contains the concluding remarks about the proposed designs. A list of sets of shifts for the construction of CNBs in two and three different sizes of periods with many treatments smaller than thirty is provided in the Appendix.

3. Model and efficiency for CNBs

The model used for repeated measurements designs in the literature is the conventional model proposed by Magda (Citation1980) (1) yijk=μ+τd(k, j)+γd(k1, j)+πk+ξij+ϵijk,(1) where yijk is the ith observation from the subject jth with the sequence i in period k for which treatment d(k, j) is given.

μ = Overall mean, τd(k, j) =  treatment d(k, j) effect, γd(k1, j) =  treatment d(k − 1, j) effect in the period k which was applied in the period k−1 to the same experimental subject, πk = kth period effect, ξij  = jth subject effect of sequence i, and ϵijk is i.i.d normally distributed residual term having mean 0 and fixed variance σ2.

The joint information matrix of treatment and carryover effects for the CNB is expressed by A(τ,γ)=(A11A12A12A22),where A11, A12 and A22 in two different period sizes are expressed by A11=RM1k1N1N11k2N2N2+r_r_n1k1+n2k2J,M=[(s2(k21)+(s1)2n1+n2+(k1k2)n1)Jt1(t1)(s1)n1+n21_t1(t1)(s1)n1+n21_t1(t1)2n1+n2], A12=ZM¯1k1N1N¯11k2N2N¯2+r_r¯_n1k1+n2k2J, M¯=[(s2(k22)+2s(s1)n1+n2+(k1k2)n1)Jt1s(t1)n1+n21_t1s(t1)n1+n21_t10], A22=R¯M1k1N¯1N¯11k2N¯2N¯2+r¯_r¯_n1k1+n2k2J,

and A11, A12 and A22 in periods of three different sizes are expressed by A11=RM1k1N1N11k2N2N21k3N3N3+r_r_n1k1+n2k2+n3k3J, M=[(s2(k31)+(s1)2n1+n2+n3+(s1)2(k2k3)n1+n2+(s2)2(k1k2)n1)Jt1n3(s1)n1+n2+n31_t1n3(s1)n1+n2+n31_t1n32n1+n2+n3], A12=ZM¯1k1N1N¯11k2N2N¯21k3N3N¯3+r_r¯_n1k1+n2k2+n3k3J, M¯=[(s2(k32)+2s(s1)n1+n2+n3+(s1)2(k2k3)n1+n2+(s2)2(k1pk2)n1)Jt1sn3n1+n2+n31_t1sn3n1+n2+n31_t10], A22=R¯M1k1N¯1N¯11k2N¯2N¯21k3N¯3N¯3+r¯_r¯_n1k1+n2k2+n3k3J.Here,

n1=

subjects measured repeatedly up to k1 size

n3=

subjects measured repeatedly up to k3 size

s=

number of sets of shifts

N1=

t incidence matrix vs 1 to n1

N2=

t incidence matrix vs n1+1 to n1+n2

N3=

t incidence matrix vs n1+n2+1 to n1+n2+n3

N¯1=

residual incidence matrix vs 1 to n1

N¯2=

residual incidence matrix vs n1+1 to n1+n2

N¯3=

residual incidence matrix vs n1+n2+1 to n1+n2+n3

r_=

vector of replication of treatments

r¯_=

vector of replication of residual

R=

diagonal matrix of replication of t

R¯=

diagonal matrix of replication of residual

J=

matrix of 1s

Then the information matrix of the treatment and carryover effects are Aτ=A11A12A22A21 and Aγ=A22A21A11A12, respectively.

RMDs should be analysed for their capability of discriminating the effects of treatment from carryover effects. Hanford (Citation2005) has given the criteria to compare the RMDs based on this ability. Divecha and Gondaliya (Citation2014) gave a convenient method for calculating the efficiency of separability (ES) for the BRMDs. This formula considering the provided constraints of our proposed type of RMDs is given by (2) ES=[tt11tt1]×100%.(2)

4. Construction methodology

Construction of CNBs from all the series of generators requires an understanding of rule II of an MOCS which was proposed by Iqbal and Jones (Citation1994). MOCS (Rule II) is defined here briefly for constructing CNBs in periods of different sizes.

Let S1 = [q11, q12, … , qp11], S2 = [q21, q22, … , qp21] and S3 = [q31, q32, … , qp32]t be the sets of shifts, where 0 ≤qij ≤ v − 2. If every element 0, 1, … , v − 2 appears exactly once in S*, it is CNB in periods of sizes p1, p2 and p3, where S* = [q11, q12, … , qp11, q21, q22, … , qp21, q31, q32, … , qp31, v − 1 − (q11+ q12+ ··· + qp11) mod (v − 1), v − 1 − (q21+ q22+ ··· + qp21) mod (v − 1)]. In these sets of shifts, the sum of any 2, 3, … , or (p − 3) successive elements should not be 0 mod (v − 1). If so, rearrange them.

Consider here the Rule II of an MOCS briefly for the construction of CNBs with the help of an example as t = 13, S1 = [3, 4, 5, 9, 8] and S2 = [6]t.

Take t − 1 experimental subjects for one set of shifts [3,4,9,5,8]. Write 0, 1, … , t − 2 in the first period of t − 1 experimental subjects, respectively. For the second period of every subject, add 3 mod (t − 1) to every element of the first period of each subject. Then add 4 mod (t − 1) to every element of the second period which gives the treatment number of the third period in each subject. Similarly add 9, 5 and 8, respectively.

Take t − 1 more subjects for the second set of shifts [6]t. Allocate 0, 1, … , t − 2 to every subject in the first period, respectively. To obtain the elements of the second period for every subject, add 6 mod (t − 1) to every element of the first period for all subjects. Then insert t − 1 (i.e. 12) in each element of the third period. The above design is CNB in t = 13, k1 = 6, k2 = 3, n1 = 12, n2 = 12 and ES = 98%.

5. Generators to obtain CNBs in two different sizes of periods

Let S1 = [e11, e12, … , ep11] and S2 = [e21, e22, … , ep22]t be two sets of shifts, where 0 ≤eij ≤ t − 2. Define, S* = [e11, e12, … , ep11, e21, e22, … , ep22, t − 1 − (e11+ e12+ ⋅⋅⋅ + ep11) mod (t − 1)]. If every element 0, 1, … , t − 2 appears exactly once in the new set of shifts S*, then the design from the set of shifts will be CNB in periods of sizes k1 and k2. The sum of any 2, 3, … ,(k − 3) successive elements of any set of shifts should not be 0 mod (t − 1). If so, rearrange the elements.

Example:

Sets of shifts S1 = [1, 7, 4, 3, 9, 6] and S2 = [2, 5, 8]t give following CNB for t = 11 in k1 = 7 and k2 = 5 with ES = 97%.

Here, 106 means that treatment 10 was applied in the current period, while the subscript ‘6’ is the treatment applied in the previous period.

Generator 5.1:

If t = 2zi + 2a − 1, i is an integer, k1 = 2z, k2 = 2a, a ≠ z and a, z > 1, then CNBs can be obtained from the sets given below where j = 0, 1, … , i − 1. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a1,t1(zi+1),t1(zi+2),,t1(zi+a2), 0]t.

Example 5.1:

CNBs for t = 13, k1 = 8 and k2 = 6 are constructed from the sets of shifts given below with 98% efficiency. S1=[1, 2, 3, 4, 11, 10, 9],S2=[5, 6, 7, 0]t.

Generator 5.2:

If t = 2zi + 2a, i is an integer, k1 = 2z, k2 = 2a + 1 and a, z > 1, then CNBs can be obtained from the sets given below where j = 0, 1, … , i − 1. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a1,t1(zi+1),t1(zi+2),,t1(zi+a1), 0]t.

Example 5.2:

CNBs for t = 16, k1 = 10 and k2 = 7 are constructed from the sets given below with 98% efficiency. S1=[1,2,3,4,14,5,13,11,12],S2=[6,7,9,8,0]t.

Generator 5.3:

If t = 2zi + 4a − 1, i is an integer, k1 = 2z, k2 = 2a and az, a, z > 1, then CNBs can be obtained from the sets given below. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a1,t1(zi+a1),t1(zi+a2),,t1(zi+1)];Si+2=[zi+a+1,zi+a+2,,zi+2a1,t1(zi+a+1),t1(zi+a+2),,t1(zi+2a2), 0]t.

Example 5.3:

CNBs for t = 25, k1 = 10 and k2 = 8 are constructed from the sets given below with 99% efficiency. S1=[1,2,3,4,5,23,22,21,20],S2=[6,7,8,17,9,18,16],S3=[10,11,12,14,13,0]t.

Generator 5.4:

If t = 2zi + 4a + 1, i is an integer, k1 = 2z, k2 = 2a + 1 and a, z > 1, then CNBs can be obtained from the sets given below. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a,t1(zi+1),t1(zi+2),,t1(zi+a)];Si+2=[zi+a+1,zi+a+2,,zi+2a,t1(zi+a+1),t1(zi+a+2),,t1(zi+2a1)]t.

Example 5.4:

CNBs for t = 21, k1 = 8 and k2 = 7 are constructed from the sets given below with 99% efficiency. S1=[1,2,3,4,19,18,17],S2=[5,6,7,15,14,13],S3=[8,9,10,12,11]t.

6. Generators to obtain CNBs in three different sizes of periods

Let S1 = [e11, e12, … , ep11], S2 = [e21, e, … , ep21] and S3 = [e31, e32, … , ep32]t be the sets of shifts, where 0 ≤eij ≤ t − 2. Define, S* = [e11, e12, … , ep11, q21, e22, … , ep21, e31, e32, … , ep31, t − 1 − (e11+ e12+ ··· + ep11) mod (t − 1), t − 1 − (e21+ e22+ ··· + ep21) mod (t − 1)]. If every element 0, 1, … , t − 2 appears exactly once in the new set of shifts S*, then the design from the set of shifts will be CNB in periods of sizes k1, k2 and k3. The sum of any 2, 3, … , (k − 3) successive elements of the set of shifts Si + 1 should not be 0 mod (t − 1). If so rearrange the elements.

Example

Sets of shifts S1 = [3, 4, 9, 5, 8], S2 = [1, 2, 11, 10] and S3 = [6]t give the following CNBs for t = 13 in k1 = 6, k2 = 5 and k3 = 3 with ES = 98%.

Here, 116 is the treatment 11 applied in the present period, while the subscript ‘6’ is the treatment applied in the previous period.

Generator 6.1:

If t = 2zi + 2a + 2b, i is an integer, k1 = 2a, z ≠ a, z, a, b > 1, k2 = 2a + 1 and k3 = 2b, then CNBs can be obtained from the sets given below where j = 0, 1, … , i − 1. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a,t1(zi+1),t1(zi+2),,t1(zi+a)];Si+2=[zi+a+1,zi+a+2,,zi+a+b1,t1(zi+a+1),t1(zi+a+2),,t1(zi+a+b1)]t.

Example 6.1:

CNBs for t = 13, k1 =  8, k2 = 7 and k3 = 4 are constructed from the sets given below with 98% efficiency. S1=[1,2,3,4,16,15,14],S2=[5,6,7,12,11,10]andS3=[8,9]t.

Generator 6.2:

If t = 2zi + 2a + 2b − 1, i is an integer, k1 = 2z, z ≠ a ≠ b > 1, k2 = 2a and k3 = 2b, then CNBs can be constructed from the sets given below where j = 0, 1, … , i − 2. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a,t1(zi+1),t1(zi+2),,t1(zi+a1)];Si+2=[zi+a+1,zi+a+2,,zi+a+b1,t1(zi+a+1),t1(zi+a+2),,t1(zi+a+b2),0]t,forb>2;Si+2=[(t1)/2, 0]t,forb=2.

Example 6.2:

CNBs for t = 23, k1 =  10, k2 = 8 and k3 = 6 are constructed from the sets given below with 99% efficiency. S1=[1,2,3,4,5,21,20,19,18],S2=[6,7,8,9,16,15,14] andS3=[10,11,12,0]t.

Generator 6.3:

If t = 2zi + 2a + 2b + 1, k1 = 2z, i is an integer, k2 = 2a + 1,k3 = 2b + 1, a ≠ b > 1 and z > 1, then CNBs can be obtained from the sets given below where j = 0, 1, … , i − 1. Sj+1=[zj+1,zj+2,,zj+z,t1(zj+1),t1(zj+2),,t1(zj+z1)];Si+1=[zi+1,zi+2,,zi+a,t1(zi+1),t1(zi+2),,t1(zi+a)];Si+2=[zi+a+1,zi+a+2,,zi+a+b,t1(zi+a+1),t1(zi+a+2),,t1(zi+a+b1)]t.

Example 6.3:

CNBs for t = 19, k1 =  8, k2 = 7 and k3 = 5 are constructed from the sets given below with 99% efficiency. S1=[1,2,3,4,17,16,15] ,S2=[5,6,13,7,12,11],S3=[8,9,10]t.

7. Concluding remarks

CSBs help in estimating direct and carryover effects independently and providing high efficiency of separability. Despite the importance of RMDs, there is no sufficient literature available on the construction of the CSBs and CNBs with high efficiency of separability for unequal period sizes. In situations where CSBs cannot be constructed, they are preferable as they give efficiencies close to CSBs. Through well-known MOCS Rule II, these designs have been constructed for unequal period sizes with high efficiency of separability which covers the lack of designs for unequal period sizes. The experimenters now have more choices when dealing with different period sizes. All the proposed designs possess an efficiency close to 100% and, therefore, they should be a better alternative to CSBs.

Acknowledgements

The authors are thankful to the reviewers for the valuable suggestions and corrections.

References

  • Afsarinejad, K. (1994). Repeated measurements designs with unequal periods sizes. Journal of the Italian Statistical Society, 3(2), 161–168. https://doi.org/10.1007/BF02589224
  • Bashir, Z., Ahmed, R., Tahir, M. H., Ghazali, S. S. A., & Shehzad, F. (2018). Some extensions of circular balanced and circular strongly balanced repeated measurements designs. Cozzunications in Statistics – Theory and Zethods, 47(9), 2183–2194. https://doi.org/10.1080/03610926.2017.1337140
  • Cheng, C. S., & Wu, C. F. (1980). Balanced repeated measurements designs. Annals of Statistics, 8(8), 1272–1283. https://doi.org/10.1214/aos/1176345200
  • Divecha, J., & Gondaliya, J. (2014). Construction of minimal balanced cross over designs having good efficiency of separability. Electronic Journal of Statistics, 8(2), 2923–2936. https://doi.org/10.1214/14-EJS979
  • Hanford, K. (2005). Lecture notes on mixed models. Spring.
  • Iqbal, I., & Jones, B. (1994). Efficient repeated measurements designs with equal and unequal period sizes. Journal of Statistical Planning and Inference, 42(1-2), 79–88. https://doi.org/10.1016/0378-3758(94)90190-2
  • Iqbal, I., & Tahir, M. H. (2009). Circular strongly balanced repeated measurements designs. Cozzunications in Statistics—Theory and Zethods, 38(20), 3686–3696. https://doi.org/10.1080/03610920802642566
  • Iqbal, I., Tahir, M. H., & Ghazali, S. S. A. (2010). Circular first-and second-order balanced repeated measurements designs. Cozzunications in Statistics-Theory and Zethods, 39(2), 228–240. https://doi.org/10.1080/03610920902941728
  • Magda, C. G. (1980). Circular balanced repeated measurements designs. Communications in Statistics – Theory and Methods, 9(18), 1901–1918. https://doi.org/10.1080/03610928008828004
  • Pearce, S. C. (1964). Experimenting with blocks of natural size. Biometrika, 20(4), 699–706. https://doi.org/10.2307/2528123
  • Rasheed, U., Rasheed, H. M. K., Rasheed, M., & Ahmed, R. (2018). Minimal circular strongly balanced repeated measurements designs in periods of three different sizes. Communications in Statistics – Theory and Methods, 47(16), 4088–4094. https://doi.org/10.1080/03610926.2017.1361995
  • Schulz, K. F. (1995). Subverting randomization in controlled trials. JAMA: The Journal of the American Medical Association, 274(18), 1456–1458. https://doi.org/10.1001/jama.1995.03530180050029
  • Schulz, K. F., & Grimes, D. A. (2002). Generation of allocation sequences in randomised trials: Chance, not choice. The Lancet, 359(9305), 515–519. https://doi.org/10.1016/S0140-6736(02)07683-3

Appendix

Table A1. List of sets of shifts for CNBs in two different sizes of periods and treatments smaller than 30.

Table A2. List of sets of shifts for CNBs in three different sizes of periods and treatments smaller than 30.