a national unit for bayesian health decision science
TRANSCRIPT
Decision science is the study of how people make decisions and how they can make better decisions in the presence of uncertainty, complexity and competing
values.
If the p-value is < 0.05, then we reject the null hypothesis, and say that there is a ‘treatment effect.’
Otherwise, we have insufficient evidence to be able to claim a ‘statistically significant’ finding.
If the drug is safe and there is a statistically significant treatment benefit, then licence the drug.
Otherwise, don’t grant it a licence.
‘Traditional’ new drug decision.
Which would you recommend …
Treatment (A) or (B)?
An RCT published in NEJM shows that (A) works better … (p=0.013)
More Important (?) Concerns …
Which would you recommend …
Treatment (A) or (B)?
An RCT published in NEJM shows that (A) works better … (p=0.013)
OR = 6.0 (95% CI 1.5 – 24.7)
More Important (?) Concerns …
Which would you recommend …
Treatment (A) or (B)? … given that …
… treatment with (A) costs €600 compared to €30 and requires 2 months off work. The condition in question is ingrown toenail. The outcome is recurrence free at 1 year.
More Important (?) Concerns …
Which would you recommend …
Treatment (A) or (B)? … given that …
… treatment with (A) costs €600 compared to €30 and requires 2 months off work. The condition in question is a congenital heart defect. The outcome is survival after 5 years.
More Important (?) Concerns …
Bayes is different …
• You don’t (and will never) know anything … ‘hold onto your uncertainty’.
• Everything you have experienced counts when doing inference (prior belief).
• Values (as utilities) are naturally a part of the analytical framework.
• Sequential (or complex) decision problems can be dealt with coherently.
21
Results of experiment
What I believed before I did the experiment
Probable values for the outcome of interest
… Bayes theorem only tells us of how to update our beliefs (in probabilistic terms) … it does not tell us how to decide whether something is statistically significant or otherwise … we can use the same ‘rule’ as the 0.05 – but we can do otherwise – utility is a natural fit …
Maximise subjective expected utility.
That is, consider the courses of action that are available. If there are some uncertain parameters, use the probability distributions.
Choose the path that leads to the best outcome ‘on average.’
Bayesian decision-making
Bayesian decision-making
Tx A
Tx B
Utility of (Failure of A combined with costs etc)
Utility of (Success of A combined with costs in financial, time and side effects)
Utility of (Success of B combined with costs etc)
Utility of (Failure of B combined with costs etc)
Bayesian decision-making
Tx A
Tx B
U(A+,-costs)
U(A-,-costs)
U(B-,-costs)
U(B+,-costs)
P(B+|Data)
P(A+|Data)
Bayesian decision-making
Tx A
Tx B
U(A+,-costs)
U(A-,-costs)
U(B-,-costs)
U(B+,-costs)
Expected Utility(A|Data)
Expected Utility(B|Data)
Expected Utility = sum(P*U)
choose the biggest
Bayesian decision-making
Tx A
Tx B
U(A+,-costs)
U(A-,-costs)
U(B-,-costs)
U(B+,-costs)
P(B+|Data)
P(A+|Data)
OR = 6.0 (95% CI 1.5 – 24.7)
Treatment with (A) costs €600 compared to €30 and requires 2 months off work.
The condition in question is a congenital heart defect. The outcome is survival after 5 years.
The condition in question is ingrown toenail. The outcome is recurrence free at 1 year.
More Important (?) … revisited
Motivation
• Health technology assessment (HTA) is a structured methodology, helping identify which technologies or pharmaceutical agents should have priority in being introduced into the health system.
• This contrasts with regulatory concerns, which are designed to ensure ‘fitness for purpose’.
Decision Making - the C/E plane
• The cost effectiveness plane is a core aspect of how outcomes are communicated and interpreted. It trades off gains in health outcomes (on the x-axis) and costs (on the y-axis).
• Note that it is only a part of the process and only ‘informs’ decision.
System
Model
Input Output
NewInput
Inference
Costs, QALYs
Progression rates, Efficacy etc.
Probability Distribution
System
Model
Input Output
NewInput
Inference
Costs, QALYs
Progression rates, Efficacy etc.
Probability Distribution