Presentation is loading. Please wait.

Presentation is loading. Please wait.

Plenary Data Analysis Session PSI Conference Bristol 2006 John Matthews School of Mathematics and Statistics University of Newcastle upon Tyne.

Similar presentations


Presentation on theme: "Plenary Data Analysis Session PSI Conference Bristol 2006 John Matthews School of Mathematics and Statistics University of Newcastle upon Tyne."— Presentation transcript:

1 Plenary Data Analysis Session PSI Conference Bristol 2006 John Matthews School of Mathematics and Statistics University of Newcastle upon Tyne

2 Five Period Crossover volunteer study  Active treatments, A to F – six doses of a new compound  Two control treatments, c 1, c 2, namely a positive control S - a standard treatment already on the market a negative control P - a placebo with zero dose of the active compound  Alternating study, two cohorts, each of 10 volunteers

3  A to F are increasing doses d j in g  Exact doses not taken into account in analysis but doses must escalate  Contrasts d j - c i of equal and primary interest: i = 1,2 and j = 1,…,6 ABCDEF 103060150250400

4  Within each period, data (FEV 1 ) collected at baseline and at six times post administration  Main interest is in response at 12 hours  One volunteer withdrew after three periods (missed P and F), otherwise data complete  Substantial washout – no carryover, ?period effect?

5 Quick and thoughtless analysis  Fixed subjects effects  Period effect  No period to period baseline

6 ANOVA table Df Sum Sq Mean Sq F value Pr(>F) factor(subject) 19 50.672 2.667 115.4127 <2.e-16 factor(period) 4 0.120 0.030 1.2969 0.28017 factor(Rx) 7 0.467 0.067 2.8880 0.01067 * Residuals 67 1.548 0.023  No strong evidence of period effect  Some treatment effect – largely because of difference between +ve and –ve controls

7 Mean effects Dose Mean Difference (l) from negative control (baseline as covariate) SEP 10-0.087-0.0770.0690.0620.210.22 30-0.052-0.0450.0700.0630.460.48 60-0.039-0.0090.0670.0610.570.88 150-0.046-0.0340.0680.0610.500.58 250-0.029 0.0700.0630.680.65 400-0.043-0.0580.0740.0670.560.39 S0.1250.1580.0490.0440.0130.0008

8 Design V’rs Cohort 1 V’rs Cohort 2 1,2 PASCE 11,12 PBSDF 3,4 SACEP 13,14 SBDFP 5,6 ACPES 15, 16 BDPFS 7,8 APCSE 17, 18 BPDSF 9.10 ASCPE 19,20 BSDPF

9 Can a better design be found? 1. Need to establish criteria for how good a design is 2. Not all aspects are numerical 3. Practical constraints 4. Statistical criteria

10 Practical constraints  Doses must escalate  Don’t start too high  Dose increments should not be too large  Two cohorts - cohort 2 investigated while cohort 1 rests. Study to finish in 3/12

11 Statistical criteria  Model can be written as  Hence where

12 Variance of contrasts  Want to consider variance of estimate of   This might be thought to be C -1 but C is singular  Therefore use g-inverse C -  C - is not unique but  If A is a c  t matrix of contrasts of interest then dispersion of contrasts, AC - A, is well defined (need rank(C)=7=t-1 if all contrasts to be estimable)

13 Contrast matrix A SPABCDEF 1000000 100 0000  1000000 01 00000 010 0000  0100000

14 Statistical Criterion  If all else is equal, we prefer a design with lower mean variance for a contrast of interest, i.e. minimise trace( AC - A)  Might want to minimise max{( AC - A) ii } but this is not pursued here

15 Practical improvements to design  Split doses into {A,C,E} and {B,D,F} to permit alternating design  Also ensures that dose increments are not large  Given doses must escalate there is little room for use of different designs  Flexibility about when controls given  Once these are chosen, sequences are defined

16 Design controls Cohort 1Cohort 2 PSPS SPSP PSPS PSPS SPSP

17 Control disposition  Same pattern in two cohorts  P before S in 12 volunteers  trace( AC - A)=2.429  There are 5 C 2 = 10 different unordered pairs of places in a sequence  Allowing for order there are twenty possible sequences for each cohort

18 Possible control sequences PS PS PS PS PS PS PS PS PS PS  Type ‘PS’ sequences  Further 10 sequences with S preceding P, type ‘SP’ sequences  Fill in gaps with either {A,C,E} or {B,D,F}  Allocate {A,C,E} to 10 sequences and {B,D,F} to remainder – allows alternation and close to balance on volunteers

19 Allocation method 1 PS PS PS PS PS PS PS PS PS PS  40 possible sequences  10 with {A,C,E} in sequences shown  10 with {A,C,E} and Type ‘SP’ sequences  20 as above but with {B,D,F} not {A,C,E}  Choose random 20 from these 40. Perhaps search for a ‘good’ set

20 Method 1  For original design trace( AC - A)=2.429  Method 1 ensures no particular degree of balance  Optimal row-column designs are uniform on periods and subjects, i.e. each treatment appears equally often on each volunteer and in each period  Cannot achieve this but can we get ‘close’?  If we achieve a certain balance on volunteers, ‘closeness’ can be measured by treatment by period incidence matrix

21 Example of a Treatment Period Incidence Matrix PSABCDEF 44aa0000 44bbcc00 44ddeeff 4400ccbb 440000aa

22 Treatment  Period Incidence Matrix for original design PSABCDEF 44660000 44442200 44006600 44002244 44000066

23 Allocation method 2 PS PS PS PS PS PS PS PS PS PS  Allocate {A,C,E} to 5 randomly chosen ‘PS’ sequences

24 Allocation method 2 PSACE PASCE PS PACES APSCE PS PS PS ACPES PS  Allocate {A,C,E} to 5 randomly chosen ‘PS’ sequences

25 Allocation method 2 PSACE PASCE PS PACES APSCE PS PS PS ACPES PS  Allocate {A,C,E} to 5 randomly chosen ‘PS’ sequences  Allocate {B,D,F} other 5 ‘PS’ sequences

26 Allocation method 2 PSACE PASCE PBDSF PACES APSCE BPDSF BPDFS BDPSF ACPES BDFPS  Allocate {A,C,E} to 5 randomly chosen ‘PS’ sequences  Allocate {B,D,F} other 5 ‘PS’ sequences

27 Allocation method 2 PSACE PASCE PBDSF PACES APSCE BPDSF BPDFS BDPSF ACPES BDFPS  Allocate {A,C,E} to 5 randomly chosen ‘PS’ sequences  Allocate {B,D,F} other 5 ‘PS’ sequences  This gives full replication of ‘PS’ sequences  Allocate {A,C,E} to the 5 ‘SP’ sequences analogous to the ‘PS’ sequences just allocated to {B,D,F}  Allocate {B,D,F} to remaining ‘SP’ sequences  Gives balance over periods of two sets of doses

28 Treatment  Period Incidence Matrix for all method 2 designs PSABCDEF 44660000 44333300 44114411 44003333 44000066

29 Method 2 results  trace =2.3168  Variances of contrasts versus P given right (same as versus S) Original design Method 2Ratio A0.20680.19771.05 B0.20680.19771.05 C0.19380.18371.05 D0.19380.18371.05 E0.20680.19771.05 F0.20680.19771.05

30 Further method  Method 2 imposes balance but does not allow duplication of sequences  May be merit in allowing this

31 Allocation method 3 APCES   Choose a ‘PS’ sequence at random and allocate {A,C,E}

32 Allocation method 3 APCES  BSDFP   Choose a ‘PS’ sequence at random and allocate {A,C,E}  Allocate {B,D,F} to corresponding ‘SP’ sequence

33 Allocation method 3 APCES  BSDFP  SACPE   Choose a ‘PS’ sequence at random and allocate {A,C,E}  Allocate {B,D,F} to corresponding ‘SP’ sequence  Allocate {A,C,E} to ‘reverse’ ‘SP’ sequence

34 Allocation method 3 APCES  BSDFP  SACPE  PBDSF   Choose a ‘PS’ sequence at random and allocate {A,C,E}  Allocate {B,D,F} to corresponding ‘SP’ sequence  Allocate {A,C,E} ‘reverse’ ‘SP’ sequence  and {B,D,F} to the analogous ‘PS’ sequence

35 Allocation method 3 APCES  BSDFP  SACPE  PBDSF   Choose a ‘PS’ sequence at random and allocate {A,C,E}  Allocate {B,D,F} to corresponding ‘SP’ sequence  Allocate {A,C,E} ‘reverse’ ‘SP’ sequence  and {B,D,F} to the analogous ‘PS’ sequence  This allocates 4 volunteers – repeat a further 4 times, sampling with replacement at first step

36 Allocation method 3: chosen design ACEPS 33 PACSE 11 ACPES 11 plus other sequences as in method 3  trace =2.262

37 Treatment  Period Incidence Matrix for method 3 design PSABCDEF 55550000 44224400 22332233 44004422 55000055

38 Method 3 results  trace =2.262  Variances of contrasts versus P given right (same as versus S) Original design Method 3Ratio A0.20680.18851.10 B0.20680.18851.10 C0.19380.18831.03 D0.19380.18831.03 E0.20680.18851.10 F0.20680.18851.10

39 Conclusions  Little room for manoeuvre in design of dose-escalating studies  Positioning of controls is about limit  Nevertheless worth doing – proposed change equivalent to 10% reduction in variance at no cost  Work to be done to extend existing work on comparison with multiple controls to allow for other constraints


Download ppt "Plenary Data Analysis Session PSI Conference Bristol 2006 John Matthews School of Mathematics and Statistics University of Newcastle upon Tyne."

Similar presentations


Ads by Google