David Burnham (PegaAnalytics) Martin Owen (GlaxoSmithKline) Gill Turner (GlaxoSmithKline) GSK Scientists and Managers. Programmer. Modeller.

Transcription

1 Decision-making in the face of uncertainty A Major JMP Scripting Language Application for Decision Tree Analysis Programmer David Burnham (PegaAnalytics) Modeller Martin Owen (GlaxoSmithKline) Gill Turner (GlaxoSmithKline) End User GSK Scientists and Managers

2 R&D: We seek innovative solutions to deliver robust, sustainable, and high quality products to patients INVEST Added value Portfolio of ideas

3 What is the problem we want to solve? What we do Drug discovery research Product development Pilot plant Manufacture Prescription medicines Vaccines Consumer Health Quality by Design

4 Typical Issues Do we switch from one starting material to another? Do we switch from procedural control (engineering solution) to attribute control (continuous monitoring)? When is the best time to do this? Now, pre-launch, post-launch Cost at Risk Timing relative to milestones Opportunity at Risk What financial factors influence this decision? What uncertainties impact on the outcome? How is this decision impacted by attitude to risk? How do we get team consensus with diverse individuals? 4

5 Decision trees Why Decision Trees? Structured way of exploring options, and possible outcomes Provide a balanced picture of the risks and costs/rewards associated with each possible course of action. A decision Tree consists of 3 types of nodes:- Decision Probability Outcome 5

6 Excel Add-in JMP Add-in BUY OR BUILD? CASE STUDY 6

7 The decision is to build or buy 20 licences of A (bespoke software) Each licence costs X -2 = -40 Build of application costs licences of B (multipurpose software) Each licence costs (20 x -0.5) = -20

8 probability Three possible outcomes We already know the capability of A (100%) Cost is % chance the build will be successful Cost is % chance the build of B will fail, so have to buy A Cost is (-20) + (-40) = - 60

9 Option: Compare weighted averages to make a decision This node is -40 The weighted average of this node is -40 (50% of -20 plus 50% of -60) Weighted average is a risk-neutral strategy

10 Attitudes to risk: The flaw of averages We know the buy cost is -40 We have a 50% chance of reducing costs to -20 RISK-SEEKER BUT we have a 50% chance of Increasing cost to -60 RISK-AVOIDER How can we alter the odds to make the build option more attractive?

11 Proof of concept (PoC) (100%) (90% of 50%) (10% of 50%) (50%) Weighted average favours build!

12 Dave Burnham

13 13 BUY OR BUILD? JMP DEMO

14 Manage models

15 Present models

16 Dave Burnham

17 BUY OR BUILD? CASE STUDY IMPACT ON TEAM DECISION-MAKING 17

18 Uncertainty of the assumptions underpinning the model We have assumed there are 20 users the feasibility is 50% cost of A licence is -2 cost of B licence is -0.5 build cost is -10

19 One way sensitivity analysis

20 2 and 3-way sensitivity

21 The impact of the Proof of Concept Build Proof of concept, then build

22 Decision tree analysis models All models are wrong, but some are useful George Box It s not the science of giving the right answer, it s the art of asking the right question Stefan Scholtes Irrational Hyper-rational

23 23 Guiding principle

24 What is the value added? Before After Mindfulness of costs & benefits Openness & transparency of assumptions Builds consensus.. or identifies areas for resolution It would appear, Hopkins, that your gut feel was only indigestion What is important? What don t we need to worry about? The art of what is possible

25 Dave Burnham

26 26 WHAT NEXT?

27 Sunk costs and future costs GSK The 20K build cost is a sunk cost JMP Community Availability??

28 A Decision Tree Platform for JMP Sunk costs: the non-quantifiable benefits of build Secrets of Success Rapid prototyping and development in JMP Contractor relationship: flexibility & trust 28

29 Acknowlegments A Major JMP Scripting Language Application for Decision Tree Analysis Programmer David Burnham (PegaAnalytics) Sponsors Vern de Biasi Vance Novack Modeller Martin Owen Gill Turner End User John Strachan Kevin Cardwell Dean Edney