ST4241: Design and Analysis of Clinical Trials

Chen Zehua
Department of Statistics & Applied Probability

12:00-2:00 pm, Friday, October 12, 2012

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

Lecture 15: Latin and Greco-Latin square design for particular prognostic variable control (II)
Analysis of scenario 2 of repeated Latin squares ANOVA approach
When squares are taken as the levels of another vector, there is an additional SS for the main eect of this factor. In addition, since this factor is crossed with both the column factor and the treatment factor with multiple response measurements at each crossed level, there are two SSs for two interactions. The levels of row factor are nested within the levels of the square factor. Consequently, the SS for the row factor must be computed within each level of the square factor.
Chen Zehua ST4241: Design and Analysis of Clinical Trials

The ANOVA table

Source Square Row(S) Column S by C Treatment S by T Residual Total df q1 q (g 1) g 1 (q 1)(g 1) g 1 (q 1)(g 1) q (g 1)(g 2) qg 2 1 SS SSS R(S)SS CSS SCSS TSS STSS by subtraction TTSS MS SMS R(S)MS CMS SCMS TMS STMS RMS F ratio SMS/R(S)MS

TMS/RMS STMS/RMS

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

Formulae for the SS

SSS = g 2 R (S )SS = g CSS = qg SCSS = g TSS = qg STSS = g TTSS =
(s ) () 2 (X X ) ,

X )2 , (X i
( ) 2 ( ) X (X ) , j (s ) (s ) ( ) ( )

(s )

(s )

X X +X )2 , (X j j
( ) 2 ( ) X (X ) , k (s ) ( ) 2 (s ) X ( ) + X (X X ) , k k

X )2 . (X ijk

(s )

()

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

Remark
In the F ratio for the square eect, the denominator is R(S)MS instead of RMS. This is because that the eect of squares and the eect of rows are partially confounded. The SS for square also partially accounts for the variation caused by rows. The expectation of SMS is the sum of two terms, the rst term measures the square eect and the second term measures the same quantity as that measured by the expectation of R(S)MS. In general, to form the F ratios in an ANOVA table in complicated designs, the expectations of the MSs are derived, and the F ratios are formed such that the expected numerator dier from the expected denominator by an additional term which is the eect to be tested. For rules to determine the expected MSs, see Montgomery (1991). Design and analysis of experiments, Chapter 8. John Wiley & Sons.
Chen Zehua ST4241: Design and Analysis of Clinical Trials

The rows can be either xed levels of a factor or a random sample of that factor. If the rows are random levels, the F ratio for treatment eects should be F = TMS /STMS .

Example 14.1 (cont.)

The dierent squares can represent dierent wards. Considering ward as another factor, the ANOVA table is given as follows: Source Square Row(S) Column S by C Treatment S by T Residual Total df 3 12 3 9 3 9 24 63 SS 0.861631 2.334013 2.425906 0.754557 0.725069 1.150394 3.575524 11.827094 MS 0.2872 0.1945 0.8086 0.0838 0.2417 0.1278 0.1490 F ratio 1.48

1.62 0.86

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

Linear model approach to the analysis of scenario 2

In scenario 2, additional dummy variables, i.e., those for squares are needed. The linear model is of the form:
qg g g

= 0 +
i =2 q

i ri +
j =2 q g

j cj +
k =2

k tk
q g

+
l =2

l s l +
l =2 j =2

lj sl cj +
l =2 k =2

lk sl tk + .

Note: the subject factor is nested under squares. If there is no missing value, the linear model approach yields the same result as the ANOVA approach.

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

Signicance tests
Signicance tests are done by using the ANOVA table.

Multiple comparison
If square by treatment interaction is signicant, multiple comparison should be made within each square. Interpretation of parameters: within square l , the eect dierence between treatment k (> 1) and treatment 1 is indicated by k + lk ; the dierence between treatment k and k is indicated by (k + lk ) (k + lk ). To do the comparison, a new model without the nested subjects should be tted. The variance matrix from the model should be adjusted by dividing the RMS of the current model and multiplying the RMS of the full model.

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

If the square by treatment interaction is not signicant, the overall multiple comparison can be made. A new model without the nested subjects and the square by treatment interaction should be tted. The variance matrix corresponding to the k s from the new model must be adjusted similarly.

Computation using R Example 14.1 (cont.)

Create data objects: in addition to infant, week and tmt, another factor square must be created. square = factor(c(rep(1,16), rep(2,16), rep(3,16),rep(4,16))) Call function lm: lm.t=lm(xsquare+week+tmt+square*tmt + square*week+square/infant)
Chen Zehua ST4241: Design and Analysis of Clinical Trials

Greco-Latin squares

If, in addition to the three factors (row, column and treatment), there is another factor to be controlled, a Greco-Latin square can be used. A Greco-Latin square is obtained by imposing the levels of the additional factor on a Latin square in a way such that any level of a factor appears once and only once at each level of the other factors. The following gives the standard 3 3 Greco-Latin square: A B C B C A C A B

Chen Zehua

ST4241: Design and Analysis of Clinical Trials

The following are standard 4 4 Greco-Latin squares: g = 4: A B C B A D C D A D C B D C B A A B C D B A D C C D A B D C B A

For standard Greco-Latin squares for g = 5 ,7,8,9, 11,12, see Cochran and Cox (1957). Experimental Designs. 2nd ed. New York, Wiley.

Randomization of Greco-Latin squares

A random Greco-Latin squares can be obtained from a standard square in three steps: 1) permute the rows of the standard square; 2) permute the columns of the standard square; 3) permute the Greco letters.
Chen Zehua ST4241: Design and Analysis of Clinical Trials

Analysis
The analysis of a single Greco-Latin square is similar to that of a single Latin square. The only dierence is that in the ANOVA table, there is an additional SS (i.e., the Greco SS). Like in Latin square case, the df of the residuals of a single Greco-Latin square is too small. Replicated squares are usually needed. Parallel to the situation in replicated Latin squares, there are corresponding two scenarios for replicated Greco-Latin squares. The analysis of replicated Greco-Latin squares are similar to that of Latin squares in both scenarios. The only dierence is that additional main eect and interactions related to Greco factor are included in the analysis.

Chen Zehua

ST4241: Design and Analysis of Clinical Trials