Penalized designs of multi-response experiments

38
1 Penalized designs of multi-response experiments Valerii Fedorov Acknowledgements: Yuehui Wu GlaxoSmithKline, Research Statistics Unit

Transcript of Penalized designs of multi-response experiments

Page 1: Penalized designs of multi-response experiments

1

Penalized designs of multi-response experiments

Valerii FedorovAcknowledgements: Yuehui Wu

GlaxoSmithKline, Research Statistics Unit

Page 2: Penalized designs of multi-response experiments

2

OutlineRI approach

Reminder

Model

Information, utility functions, criteria

Design

Analysis of scenarios

Page 3: Penalized designs of multi-response experiments

3

RI approach: recycling of ideas

Page 4: Penalized designs of multi-response experiments

4

¾ came from A&F, 1988

Page 5: Penalized designs of multi-response experiments

5

Twenty years ago

Page 6: Penalized designs of multi-response experiments

6

Thirty years ago

Page 7: Penalized designs of multi-response experiments

7

Observations and penaltyReminder

Page 8: Penalized designs of multi-response experiments

8

Likelihood and information matrixReminder

Page 9: Penalized designs of multi-response experiments

9

Cost constrained designReminder

Page 10: Penalized designs of multi-response experiments

10

Design transformReminder

Page 11: Penalized designs of multi-response experiments

11

Equivalence theoremReminder

Page 12: Penalized designs of multi-response experiments

12

Specific cases from 1988, A&F

Page 13: Penalized designs of multi-response experiments

13

Bayesian* and composite designs

Stage Conservative Optimistic

Design Use prior Use prior

Estimation No prior Use prior

*

Page 14: Penalized designs of multi-response experiments

14

First order algorithmReminder

Page 15: Penalized designs of multi-response experiments

15

Adaptive designReminder

Page 16: Penalized designs of multi-response experiments

16

Binary observationsModel

Page 17: Penalized designs of multi-response experiments

17

Two drugs: typical responsesModel

Page 18: Penalized designs of multi-response experiments

18

Probit modelModel

Page 19: Penalized designs of multi-response experiments

19

More about probit model can be found in:

Model

Page 20: Penalized designs of multi-response experiments

20

Diagram from Pearson’s paper

Page 21: Penalized designs of multi-response experiments

21

Linear predictorModel

In what follows θ = {-1.26, 2.0, 0.9, 14.8; -1.13, 0.94, 0.36, 4} , ρ = 0.5

Page 22: Penalized designs of multi-response experiments

22

Information Matrix for a Single Observationor What Can be Learned from One Patient

Information

Page 23: Penalized designs of multi-response experiments

23

2D Penalty functionPenalty = { Probability of efficacy} – 1 ∗ {Probability of no toxicity} – 1.

Model

Page 24: Penalized designs of multi-response experiments

24

Utility and optimality criterion

Utility (or what we want): Probability of efficacy without toxicity, i.e. p10(x)Distance from desirability point: δ (x*) All unknown parameters

Optimality criteria:Var {estimated p10(x)}Var {estimated max[p10(x)]} + Var {estimated location of max[p10(x)]} Var {estimated x*}Var {estimated θ }

Utility and optimality criterion

Page 25: Penalized designs of multi-response experiments

25

Utility - I

arg max p10(x,θ) ={ 0.18, 0.68 }

Utility and optimality criterion

Page 26: Penalized designs of multi-response experiments

26

Utility - IIUtility and optimality criterion

Page 27: Penalized designs of multi-response experiments

27

Design scenarios

91 subjects were assigned to three doses per drug in initial design (7 support points)Another 90 subjects to be assignedTotal sample size: 181

Analysis of scenarios

Page 28: Penalized designs of multi-response experiments

28

Penalized locally D-optimalAnalysis of scenarios

Page 29: Penalized designs of multi-response experiments

29

Penalized composite locally D-optimal

Initial design

Added design points

Analysis of scenarios

Page 30: Penalized designs of multi-response experiments

30

Design comparison for D-criterioninitial design included

Design type Information per patient

Penalty per patient

Information per penalty u.

Locally D-optimal 1 2.07 0.48

0.79

0.62

0.52

0.50

0.39

0.50

Penalized locally optimal 0.79 1

Penalized composite locally optimal 0.77 1.23

Penalized composite (median, 500) 0.70 1.36

Penalized adaptive (median, 500) 0.75 1.50

“Do your best” adaptive (median, 500) 0.46 1.19

Restricted “Uniform ” , 12 points 0.67 1.34

Analysis of scenarios

Page 31: Penalized designs of multi-response experiments

31

Simulation results (500 runs): the estimated best combination

Analysis of scenarios

Adaptive D-optimalComposite D-optimal

Do your best

Best dose

Page 32: Penalized designs of multi-response experiments

32

Simulation results (500 runs): the estimated best response

Composite D-adaptive Do your best

Analysis of scenarios

Page 33: Penalized designs of multi-response experiments

33

CONCLUSION

Recycling helps

Page 34: Penalized designs of multi-response experiments

34

References Atkinson, A.C. and Fedorov, V.V. (1988). The optimum design of experiments in the presence of uncontrolled variability and prior information. In Optimal Design and Analysis of Experiments. Eds. Dodge Y., Fedorov V.V. and Wynn H.P. New York: North-Holland. pp 327- 344.Box, G.E.P. and Hunter, W.G. (1963). Sequential design of experiments for

nonlinear models. In "Proceedings of IBM Scientic Computing Symposium (Statistics)". 113-137.Cook, D. and Fedorov, V. (1995). Constrained optimization of experimental design. Statistics 26: 129-178.Dragalin, V., Fedorov, V., and Wu, Y. (2006). Optimal Designs for BivariateProbit Model. GSK Technical Report 2006-01. http://www.biometrics.com/8D97129158B901878025714D00389BEB.htmlFedorov, V.V. and Hackl, P. (1997). Model-Oriented Design of Experiments. Lecture Notes in Statistics, pp.125. Springer.Ford, I., Titterington, D.M. and Wu, C.F.J. (1985). Inference and sequential design. Biometrika 72: 545-551.Hu, I. (1998). On sequential designs in nonlinear problems. Biometrika 85: 446-503.Hu, F. and Rosenberger, W. (2006), The theory of response-adaptive randomization in clinical trials, Wiley.Ying, Z. and Wu, C.F.J. (1997). An asymptotical theory of sequential designs based on maximum likelihood recursions. Statistica Sinica 7: 75-91.

Page 35: Penalized designs of multi-response experiments

35

Appendix

Page 36: Penalized designs of multi-response experiments

36

Estimation of the best dose with and without dichotomization I

Dichotomized Continuous

Model

Page 37: Penalized designs of multi-response experiments

37

Estimation of the best dose with and without dichotomization II

Model

Page 38: Penalized designs of multi-response experiments

38

2D binaryResponders without toxicity -

Responders with toxicity

Non-responders without toxicity

Non-responders with toxicity

Model