Chapter 18: Split-Plot, Repeated Measures, and Crossover Designs

Introduction

When the experiment involves a factorial treatment structure, the implementation of one or two factors may be more time-consuming, more expensive, or require more material than the other factors. In situations such as these, a split-plot design is often implemented. For example, in an educational research study involving two factors, teaching methodologies and individual tutorial techniques, the teaching methodologies would be applied to the entire classroon of students. The tutorial techniques would then be applied to the individual students within the classroom. In an agricultural experiment involving the factors, levels of irrigation and varieties of cotton, the irrigation systems must apply the water to large sections of land which would then be subdivided into smaller plots. The dierent varieties of cotton would then be plantted on the smaller plots. In a crossover designed experiment, each subject receives all treatments. The individual subjects in the study are serving as blocks and hence decreasing the experimental error. This provides an increased precision of the treatment comparisons when compared to the design in which each subject receives a single treatment. In the repeated measures designed experiment, we obtain t dierent measurements corresponding to t dierent time points following administration of the assigned treatment. The multiple observations over time on the same subject often yield a more ecient use of experimental resources than using a dierent subject for each obsevation time. Thus, fewer subjects are required, with a subsequent reduction in cost. Also, the estimation of time trends will be measured with a greater degree of precision. Medical researchers, ecological studies, and numerous other areas of research involve the evaluation of time trends and hence may nd the repeated measure design useful.

Split-Plot Designed Experiments

The yields of three dierent varieties of soybeans are to be compared under two dierent levels of fertilizer application. If we are interested in getting n = 2 observations at each combination of fertilizer and variety of soybeans, we would need 12 equal-sized plots. Taking fertilizer as factor A and varieties as a treatment factor T, one possible design would be an 2 3 factorial treatment structure with n = 2 observations per factor-level combination. However, since the application of fertilizer to a plot occurs when the soil is being prepared for planting, it would be dicult to rst apply fertilizer A1 to six of the plots dictated by the factorial arrangement of factors A and T and then fertilizer A2 to the other six plots before planting the required varieties of soybeans in each plot. An easier design to execute would have each fertilizer applied to two larger wholeplots and then the varieties of soybeans planted in three subplots within each whole plot. This design is called a split-plot design, and with this design there is a two-stage randomization. First, levels of factor A (fertilizers) are randomly assigned to the wholeplots; second, the levels of factor T (soybeans) are randomly assigned to the subplots within a wholeplot. Using this design, it would be much easier to prepare the soil and to apply the fertilizer to the larger wholeplots. Consider the model for the split-plot design with a levels of factor A, t levels of factor T, and n repetitions of the i levels of factor A. If yijk denotes the kth response for the ith level of factor A, jth level of factor T, then yijk = + i + ik + j + ij + where i : Fixed eect for ith level of A. j : Fixed eect for jth level of T. ij : Fixed eect for ith level of A, jth level of T. ik : Random eect for the kth wholeplot receiving the ith level of A. The ik are independent
2 normal with mean 0 and variance . ijk ,

ijk :

Random error.
ijk

The ik and

are mutually independent. 2

Table 1: An ANOVA table for a completely randomized split-plot design. Source A Wholeplot Error T AT subplot error Total SS SSA SS(A) SST SSAT SSE TSS df a1 a(n 1) t1 (a 1)(t 1) a(n 1)(t 1) atn 1 EMS
2 2 + t + tn 2 2 + t

2 + an 2 + n 2

The ANOVA for this model and design is shown in Figure 1. The sum of squares can be computed using our standard formulas. Wholeplot analysis H0 : = 0 (or, equivalently, H0 : all i = 0), F =
M SA M S(A) . M SAT M SE .

Subplot Analysis H0 : = 0 (or, equivalently, H0 : All ij = 0), F = H0 : = 0 (or, equivalently, H0 : All j = 0), F =
M ST M SE .

A variation on this design introduces a blocking factor (such as farms). Thus for our example, there may be b = 2 farms with a = 2 wholeplots per farm and t = 3 subplots per wholeplot. The model for this more general two-factor split-plot design laid o in b blocks is as follows: yijk = + i + j + ij + k + ik +
ijk ,

where yijk denotes the measurement receiving the ith level of factor A and the kth level of factor T in the jth block. The parameters i , k , and ik are the usual main eects and interaction parameters for a two-factor experiment, whereas j is the eect due to block j and ij is the interaction between the ith level of factor A and the jth block. The analysis corresponding to this model is shown in Table 2. Wholeplot analysis. H0 : = 0 (or, equivalently, H0 : all i = 0), F = Subplot analysis. H0 : = 0 (or, equivalently, H0 : all ik = 0), F = H0 : = 0 (or, equivalently, H0 : all k = 0), F = 3
M SAT M SE . M SA M SAB .

M ST M SE .

Table 2: An ANOVA table for a randomized split-plot design (A, T xed; block random. Source Blocks A AB(Wholeplot Error) T AT subplot error Total SS SSB SSA SSAB SST SSAT SSE TSS df b1 a1 (a 1)(b 1) t1 (a 1)(t 1) a(b 1)(t 1) abt 1 EMS
2 2 + at 2 2 + t + bt 2 2 + t

2 + ab 2 + b 2

Example 2.1 Soybeans are an important crop throughout the world. A study was designed to determine if additional phosphorus applied to the soil would increase the yield of soybean. There are three major varieties of soybeans of interest (V1 , V2 , V3 ) and four levels of phosphorus (0, 20, 40, 65, pounds per acre). The researchers have nine plots of land available for the study which are grouped into blocks of three plots each based on the soil characteristics of the plots. Because of the complexities of planting the soybeans on plots of the given size, it was decided to plant a single variety of soybeans on each plot and then divide each plot into four subplots. The researchers randomly assigned a variety to one plot within each block of three plots and then randomly assigned the levels of phosphorus to the four subplots within each plot. The yields (bushels/acre) froom the 36 plots are given in Table 18.5 of the textbook. For this study, we have a randomized complete block design with a split-plot structure. Variety, with 3 levels, is the wholeplot treatment and amount of phosphorus is the split-plot treatment. The ANOVA analysis is as follows. The results indicate that there is a signicant variety by phosphorus interaction from which we can conclude that the relationship between average yield and amount of phosphorus added to the soil is not the same for the three varieties. The distinction between this two-factor split-plot design and the standard two-factor experiments discussed in Chapter 14 lies in the randomization. In a split-plot design, there are two stages to the randomization process; rst levels of factor A are randomized to the wholeplots within each block, and then levels of factor B are randomized to the subplot units within each wholeplot of every block. In contrast, for a two-factor experiment laid o in a randomized block design, the 4

Table 3: An ANOVA table for a randomized split-plot design (A, T xed; block random). Source Blocks V BV(Wholeplot Error) P PV subplot error Total df 2 2 4 3 6 18 35 SS 763.25 671.81 6.56 408.37 117.41 4.08 1971.48 MS 381.63 335.90 1.64 136.12 19.57 0.23 F * 232.60 * 601.04 86.40 p-value * < .0001 * < 0.0001 < 0.0001

randomization is a one-step procedure; treatments (factor-level combinations of the two factors) are randomized to the experimental units in each block.

Single-Factor Experiments with Repeated Measures

In Section 18.1, we discussed some reasons why one might want to get more than one observation per patient. Consider a design, three compounds are administered in sequence to each of the n patients. A compound is administered to a patient during a given treatment period. After a suciently long washout period, another compound is given to the same patient. This procedure is repeated until the patient has been treated with all three compounds. The order in which the compounds are administered would be randomized. The data is shown below. multicolumn4cPatient Compound 1 2 3 1 y11 y21 y31 2 y12 y22 y32 n y1n y2n y3n

The model for this experiment can be written as yij = + i + j +

ij ,

where is the overall mean response, i is the eect of the ith compound, j is the eect of jth patient, and
ij

is the experimental error for the jth patient receiving the ith compound.

For this model, we make the following assumptions: 1. i s are constants with a = 0.
2 2. The j are independent and normally distributed N (0, ).

3. The 4. The 5. The

ij

ij s ij s ij s

are independent of the j s. are normally distributed N (0, 2 ). have the following correlation relationship:
ij ij

and

ij

are correlated for i = i ; and

and

are independent for j = j .

That is, two observations from the same patient are correlated but observations from dierent patients are independent. From these assumptions it can be shown that the variance of yij is
2 + 2 . A further assumption is that the covariance for any two observations from patient j,

yij and yi j , is constant. These assumptions give rise to a variance-covariance matrix for the observations, which exhibits compound symmetry. 6

The ANOVA table for the experiment is shown below. Source Patients A Error Totals SS SSP SSA SSE TSS df n1 a1 (a 1)(n 1) an 1 EMS (A xed, patients random)
2 2 + a

2 + n 2

When the assumptions hold, and hence compound symmetry holds, the statistical test on factor A (F = M SA/M SE) is appropriate. The conditions under which the F test for factor A is valid are often not met because observations on the same patient taken closely in time are more highly correlated than are observations taken farther apart in time. So be careful about this. In general, when the variance-covariance matrix does not follow a pattern of compound symmetry, the F test for factor A has a positive bias, which allows rejection of H0 : all i = 0 more often than is indicated by the critical F -value. Example 3.1 An exercise physiologist designed a study to evaluate the impact of the steepness of running courses on the peak heart rate (PHR) of well-conditioned runners. There are four ve-mile courses that have been rated as at, slightly steep, moderately steep, and very steep with respect to the general steepness of the terrain. The 20 runners will run each of the four courses in a randomly assigned order. There will be sucient time between the runs so that there should not be any carryover eect and the weather conditions during the runs were essentially the same. Therefore, the researcher felt condent that the model yij = + i + j + would be appropriate for the experiment. The ANOVA table for the experiment is shown below: Source Runner Course Error Totals SS 4048.44 3619.25 1083.51 8751.19 df 19 3 57 79 EMS (A xed, patients random) 213.08 1206.41 19.01 F 11.21 63.47 Prob < 0.0001 < 0.0001
ij

From the output we have that the p-value associated with the F test of H0 : 1 = 2 = 3 = 4 versus H1 : not H0

has p-value< 0.0001. Thus, we conclude that there is signicant evidence of a dierence in the mean heart rates over the four levels of steepness. The estimated variance components are given by Error = M SE = 19.01 2 Runner = 2 M SRunner M SE 213.08 19.01 = = 48.52 4 4

Therefore, 72% of the variation in the heart rates was due to the dierences in runners and 28% was due to all other sources.

Two-factor Experiments with Repeated Measures on One of the Factors

We can extend our discussion of repeated measures experiments to two-factor settings. For example, in comparing the blood-pressure-lowering eects of cardiovascular compounds, we could randomize the patients so that n dierent patients receive each of the three compounds. Repeated measurements occur due to taking multiple measurements across time for each patient. For example, we might be interested in obtaining blood pressure readings immediately prior to receiving a single dose of the assigned and then every 15 minutes for the rst hour and hourly thereafter for the next 6 hours. This experiment can be described generally as follows. There are m treatments with n experimental units randomly assigned to the treatments. Each experimental unit (EU) is assigned to a single treatment with t measurements taken on each of the EUs. The form of the data is shown below. Note that this is a two-factor experiment (treatment and time) with repeated measurements taken over the time factor. Time Period Treatment 1 EU 1 . . . n . . . m 1 . . . n ym11 ymn1 ym12 ymn2 ym1t ymnt 1 y111 y1n1 2 y112 y1n2 t y11t y1nt

The analysis of a repeated measurement design can, under certain conditions, be approximated by the methods used in a split-plot experiment. Each treatment is randomly assigned to an EU. This is the wholeplot in the split-plot design. Each EU is then measured at t time points. This is considered the split-plot unit. The major dierence is that in a split-plot design, the levels of factor B are randomly assigned to the split-plot EUs. In the repeated measurement design, the second randomization does 9

not occur, and thus there may be strong correlation between the measurements across time made on the same EU. The split-plot analysis is an appropriate analysis for a repeated measurement experiment only when the covariance matrix of the measurements satisfy a particular type of structure: Compound Symmetry: 2 Cov(yijk , yi j k ) = 2 0

when i = i , j = j when i = i , j = j when i = i

where yijk is the measurement from the kth EU receiving treatment i at time j. Thus we have Corr(yijk , yij k ) = . This implies that there is a constant correlation between observations no matter how far apart they are taken in time. This may not be realistic in many applications. One would think that observations in adjacent time periods would be more highly correlated than observations taken two or three time periods apart. The model can be written as yijk = + i + dik + j + ( )ij +
ijk

where i = 1, . . . , m, j = 1, . . . , t, k = 1, . . . , n, i is the ith treatment eect, j is the jth time eect, ( )ij is the treatment-time interaction eect, dik is the subject-treatment interaction eect
2 (random, independent, N (0, d ), ijk

independent N (0, 2 ), and dik and

ijk

are independently

distributed. Let = t/2(1 ). The ANOVA table for the split-plot analysis of a repeated measures experiment is given in Table 4, where the treatment and time eects are xed. Based on Table 4, the following tests can be performed: H0 : = 0 F = H0 : = 0 F = M Stime M SE M Strttime M SE

10

Table 4: An ANOVA table for a two-factor experiment, repeated measures on one factor. Source TRT EU(TRT) Time TRT*Time Error Total df m1 (n 1)m t1 (m 1) (t 1) m(t 1)(n 1) mnt 1 Expected Mean Squares
2 2 (1 + 2) + td + nt 2 2 (1 + 2) + td

2 + nm 2 + n 2

H0 : = 0 F = M ST rt M SEU (trt)

Example 4.1 In a study, three levels of a vitamin E supplement, zero (control), low, and high, were given to guinea pigs. Five pigs were randomly assigned to each of the three levels of the vitamin E supplement. The weights of the pigs were recorded at 1, 2, 3, 4, 5, and 6 weeks after the beginning of the study. This is a repeated measurement experiment because each pig, the EU, is given only one treatment but each pig is measured six times. The ANOVA table for the example is as follows. Source TRT PIG(TRT) Week TRT*Week Error df 2 12 5 10 60 SS 18548.07 105,434.20 142,554.50 9762.73 32,552.60 28510.90 976.27 542.54 52.55 1.80 < 0.0001 0.0801 MS 9274.03 F 1.06 p-value 0.3782

From this table, we nd that there is not signicant evidence of an interaction between the treatment and time factors. Since the interaction was not signicant, the main eects of treatment and time can be analyzed separately. The p-value=0.3782 for treatment dierences and p-value< 0.0001 for time dierences. The mean weights of the pigs vary across the 6 weeks but there is not signicant evidence of a dierence in the mean weights for the three levels of vitamin E feed supplements. Therefore, the

11

two levels of vitamin E supplement do not appear to provide an increase in the mean weights of the pigs in comparison to the control, which was a zero level of vitamin E supplement.

12

Crossover Designs

In a crossover design, each experimental unit (EU) is observed under each of the t treatments during t observation times. It is important to emphasize the dierence between a crossover design and the general repeated measurement design. In a repeated measurement design, the EU receives receives a treatment and then the EU has multiple observations or measurements made on it over time or space. The EU does not receive a new treatment between successive measurements. The crossover designs are often useful when a latin square is to be used in a repeated measurement study to balance the order positions of treatments, yet more subjects are required than called for by a single latin square. With this type of design, the subjects are randomly assigned to the dierent treatment order patterns given by a latin square. Consider an experiment in which treatments A, B, and C are to be administered to each subject, and the three treatment order pattern are given by the latin square Order Position pattern 1 2 3 1 A B C 2 B C A 3 C A B

Suppose that 3n subjects are available for the study. Then n subjects will be assigned at random to each of the three order patterns in a latin square crossover design. Note that this design is a mixture of repeated measures (within subjects) and latin square (order patterns from a latin square). For this experiment, the model can be written as yijkm = + i + j + k + m(i) +
ijkm ,

where i = 1, . . . , t, j = 1, . . . , t, k = 1, . . . , t, and m = 1, . . . , m. The term i denotes the eect of the ith treatment order pattern, j denotes the eect of the jth order position, k denotes the eect of the kth treatment, and m(i) denotes the eect of subject m which is nested within the ith
2 treatment order pattern. Here we assume m(i) are independent N (0, ), ijkm

are independent

N (0, 2 ) and independent of the m(i) . The ANOVA table for the experiment is as follows.

13

Source of Variation Patterns(P) Order Position(O) Treatments(TR) Subjects Error Total

SS SSP SSO SSTR SSS SSE SST

df t1 t1 t1 t(n 1) (t 1)(nt 2) nt2 1

EMS
2 2 + r + nr

2 + nr 2 + nr
2 2 + r

The formulas for the sums of squares follow the usual pattern: SST =
i j m

(yijkm y... )2 (i... y.... )2 y

i

SSP = nt SSO = nt
j

(.j.. y.... )2 y (..k. y.... )2 y

k

SST R = nt SSS = t
i m

(i..m y.... )2 y

SSE = SST SSP SSO SST R SSS. Example 5.1 The following table contains data for a study of three dierent displays on the sale of apples, using a latin square crossover design. Six stores were used, with two assigned at random to each of the three treatment order patterns shown. Each display was kept for two weeks, and the observed variable was sales per 100 customers. Two-week Period(j) Pattern(i) 1 Store m=1 m=2 2 m=1 m=2 3 m=1 m=2 The ANOVA table for the data is as follows: 14 1 9(B) 4(B) 12(A) 13(A) 7(C) 5(C) 2 12(C) 12(C) 14(B) 14(B) 18(A) 20(A) 3 15(A) 9(A) 3(C) 3(C) 6(B) 4(B)

Source of Variation Patterns Order positions Displays Stores (within patterns) Error The test for the treatment eect is F =

SS 0.33 233.33 189.00 21.00 20.33

df 2 2 2 3 8

MS 0.17 116.67 94.50 7.00 2.54

94.50 M ST R = = 37.2 M SE 2.54

which is greater than F0.05,2,8 = 4.46. Therefore, we conclude that there are dierential sales eects for the three displays. Tests for pattern eects, order position eects, and store eects were also carried out. They indicated that order position eects were present, but no pattern or store eects. Order position eects here are associated with the three time periods in which the displays were studied, and may reect seasonal eects as well as the results of special events, such as unusually hot weather in one period. If the order position eects are not approximately constant for all subjects (stores, etc.), a crossover design is not fully eective. It may then be preferable to place the subjects into homogeneous groups with respect to the order position eects and use independent latin squares for each group. Carryover Eects If carryover eects from one treatment to another are anticipated, that is, if not only the order position but also the preceding treatment has an eect, these carryover eects may be balanced out by choosing a latin square in which every treatment follows every other treatment an equal number of times. For t = 4, an example of such a latin square is Period Subject 1 2 3 4 1 A B C D 2 B C D A 3 D A B C 4 C D A B

15

Note that treatment A follows each of the other treatments once, and similarly for the other treatments. This design is appropriate when the carryover eects do not persist for more than one period. When t is odd, the sequence balance can be obtained by using a pair of latin squares with the property that the treatment sequences in one square are reversed in the other square. For the earlier apple display illustration in which three displays were studied in six stores, the two latin squares might be as shown in the next table. The stores should rst be placed into two homogeneous groups and these should then be assigned to the two lattin squares. Two-week Period(j) Square Store 1 1 2 3 1 A B C 2 B C A 3 C A B

4 2 5 6

C A B

B C A

A B C

16