You are on page 1of 12

# 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

Chen Zehua ST4241: Design and Analysis of Clinical Trials . there are two SS’s for two interactions. there is an additional SS for the main eﬀect of this factor. Consequently. the SS for the row factor must be computed within each level of the square factor. The levels of row factor are nested within the levels of the square 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.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.

The ANOVA table Source Square Row(S) Column S by C Treatment S by T Residual Total df q−1 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 .

(X ijk (s ) (·) Chen Zehua ST4241: Design and Analysis of Clinical Trials . (X i ·· ( ·) 2 ¯ ( ·) − X ¯··· (X ) .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 . ··k ··k ¯ −X ¯··· )2 . ··k (s ) ( ·) 2 ¯ (s ) − X ¯··· ¯ ( ·) + X ¯··· (X −X ) . (X ·j · ·j · ( ·) 2 ¯ ( ·) − X ¯··· (X ) . ·j · (s ) (s ) ( ·) ( ·) (s ) (s ) ¯ −X ¯··· − X ¯ +X ¯··· )2 .

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

62 0.575524 11.334013 2.86 Chen Zehua ST4241: Design and Analysis of Clinical Trials .2872 0.861631 2.8086 0. Example 14.150394 3. the F ratio for treatment eﬀects should be F = TMS /STMS .827094 MS 0. Considering ward as another factor.725069 1.48 1.1278 0.The rows can be either ﬁxed levels of a factor or a random sample of that factor.1945 0.754557 0. 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 0.2417 0.425906 0.) The diﬀerent squares can represent diﬀerent wards.0838 0.1 (cont. If the rows are random levels.1490 F ratio 1.

If there is no missing value. i.Linear model approach to the analysis of scenario 2 In scenario 2.. the linear model approach yields the same result as the ANOVA approach. Note: the subject factor is nested under squares.e. The linear model is of the form: qg g g X = µ0 + i =2 q αi ri + j =2 q g βj cj + k =2 γk tk q g + l =2 κl s l + l =2 j =2 ξlj sl cj + l =2 k =2 ηlk sl tk + . Chen Zehua ST4241: Design and Analysis of Clinical Trials . additional dummy variables. those for squares are needed.

the eﬀect diﬀerence between treatment k (> 1) and treatment 1 is indicated by γk + ηlk .Signiﬁcance tests Signiﬁcance tests are done by using the ANOVA table. Multiple comparison If square by treatment interaction is signiﬁcant. a new model without the nested subjects should be ﬁtted. multiple comparison should be made within each square. 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. the diﬀerence between treatment k and k is indicated by (γk + ηlk ) − (γk + ηlk ). To do the comparison. Interpretation of parameters: within square l . Chen Zehua ST4241: Design and Analysis of Clinical Trials .

rep(2. week and tmt.16). rep(3. another factor square must be created.16))) Call function lm: lm. square = factor(c(rep(1.rep(4. Computation using R —Example 14.ﬁt=lm(x∼square+week+tmt+square*tmt + square*week+square/infant) Chen Zehua ST4241: Design and Analysis of Clinical Trials . A new model without the nested subjects and the square by treatment interaction should be ﬁtted. the overall multiple comparison can be made.16).16).) Create data objects: in addition to infant.If the square by treatment interaction is not signiﬁcant. The variance matrix corresponding to the γk ’s from the new model must be adjusted similarly.1 (cont.

there is another factor to be controlled. 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.Greco-Latin squares If. column and treatment). a Greco-Latin square can be used. in addition to the three factors (row. The following gives the standard 3 × 3 Greco-Latin square: Aα B β C γ B γ C α Aβ C β Aγ B α Chen Zehua ST4241: Design and Analysis of Clinical Trials .

Wiley.8.The following are standard 4 × 4 Greco-Latin squares: g = 4: Aα B β C γ B δ Aγ D β C β D α Aδ Dγ C δ Bα Dδ Cα Bγ Aβ Aα B β C γ D δ B γ Aδ D α C β C δ D γ Aβ B α D β C α B δ Aγ For standard Greco-Latin squares for g = 5 . Experimental Designs. 2nd ed.7. 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.12. see Cochran and Cox (1957). 3) permute the Greco letters. 11. Chen Zehua ST4241: Design and Analysis of Clinical Trials .9. New York. 2) permute the columns of the standard square.

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