David Spiegelhalter
Download
Report
Transcript David Spiegelhalter
Opportunities for Bayesian
analysis in evaluation of
health-care interventions
David Spiegelhalter
MRC Biostatistics Unit Cambridge
[email protected]
Summary
•
•
•
•
•
•
•
What is the Bayesian approach?
Example: CHART
Why is it relevant to evaluation in health-care?
Example: HIPS
What areas might benefit most?
Example: ASTIN
What are key challenges?
What is the Bayesian approach?
A possible definition.
‘the explicit quantitative use of external
evidence in the design, monitoring, analysis,
interpretation and reporting of a health-care
evaluation’
But what does this mean?
Basic Bayesian ideas
• Uncertainty about unknown quantities
expressed as a probability distribution
• This ‘prior’ distribution is a judgement based
on all available evidence
• Bayes theorem provides a formal way of
revising this distribution as more evidence
accumulates
“Posterior prior x likelihood”
CHART trial in non small-cell lung cancer
The Data Monitoring Committee met annually and was presented
with full data.
Date
No
patients
No
deaths
Observed
hazard
ratio
95% CI
2-sided
P-value
1992
256
78
0.55
(0.35 to
0.86)
0.007
1993
380
192
0.63
(0.47 to
0.83)
0.001
1994
460
275
0.70
(0.55 to
0.90)
0.003
1995
563
379
0.75
(0.61 to
0.93)
0.004
1996
563
444
0.76
(0.63 to
0.90)
0.003
CHART Lung trial results
Why is it relevant to evaluation in
health-care?
• Can incorporate all relevant evidence in an
incremental way
• Can model potential biases in studies
• Answers question: how should new evidence
change our opinions?
• Directly make statements such as:
“Probability that X is cost-effective is 92%”
• Inference feeds naturally into decisionmaking and planning further studies
• Requires explicit, accountable judgments,
recognising context and multiple stakeholders
Comparison of Charnley and Stanmore
hip prosthesis (NICE, 2000)
Charnley
Stanmore
No. Patients Revision rate No. Patients Revision rate
Hazard Ratio
Source
HR
Registry
RCT
Case series
28,525
200
208
5.9%
3.5%
16.0%
865
213
982
3.2% 0.55
4.0% 1.34
7.0% 0.44
0
5000
15000
K = acceptable cost per Q ALY
(0.37-0.77)
(0.45-3.46)
(0.28-0.66)
1.0
(c) Equal weights
0.0
0.2
0.4
0.6
0.8
1.0
0.8
0.0
0.2
0.4
0.6
0.8
0.6
0.4
0.2
0% health discount
1.5%
6%
0.0
Prob(cost-effective)
1.0
(a) Medium weight to registry(b) Low weight to registry
(95% Interval)
0
5000
15000
K = acceptable cost per Q ALY
0
5000
15000
K = acceptable cost per Q ALY
What areas might benefit most?
• Planning and monitoring development
programmes
• Selection of compounds for further
investigation
• Data monitoring within studies
• Adaptive designs in proof-of-concept studies
• Evidence synthesis
• Cost-effectiveness analysis
• Value-of-information (payback) models
ASTIN study
• Adaptive dose-response study of UK-279,276 in acute
ischaemic stroke (Krams, Lees, Hacke, Grieve, Orgogozo, Ford
etc (2003)
• 15 doses available: placebo, 10 - 120 mg
• Primary outcome: increase in Scandinavian Stroke Scale (SSS)
at 90 days (adjusted for baseline)
• Next dose suggested is that which minimises the expected
variance of the response at the ED95 (minimal dose near
maximal efficacy)
• Randomisation: 15\% to placebo, 85\% `near' suggested dose
• Fits smoothly flexible curve: no imposed shape
• IDMC examined data every week
• Stop for efficacy when 90% probability that effect at ED95 > 2
• Stop for futility when 90% probability that effect at ED95 < 1
• Design approved by FDA (based on simulation studies)
• Stopped by IDMC for futility after 966 patients randomised
Changing dosing pattern
Final dose-effect curve
Doses finally given
Monitoring changing probabilities
What are key challenges?
• Marshalling appropriate evidence
• Robust, rigorous modeling with appropriate
sensitivity analysis
• Presentation in persuasive way to decisionmakers in companies and regulatory
authorities
• Integration of cost-effectiveness ideas into
product-development programmes
BUT
Cannot make silk purse …., so need good
studies and good data
References
Berry DA, Mueller P, Grieve AP, Smith M, Parke T, Blazek R, Mitchard N and
Krams M (2001) Adaptive Bayesian designs for dose-ranging drug trials.
Case Studies in Bayesian Statistics, Volume V. Eds Gatsonis C, Carlin B
and Carriquiry A. Springer-Verlag, New York. p 99-181
O'Hagan A Luce BR (2003) A Primer on Bayesian Statistics in Health
Economics. Centre for Bayesian Statistics in Health Economics, Sheffield
Parmar MKB, Griffiths GO, Spiegelhalter DJ, Souhami RL, Altman DG and van
der Scheuren E (2001) Monitoring large randomised clinical trials - a new
approach using Bayesian methods, Lancet, 358, 375—381
Spiegelhalter DJ, Abrams K, and Myles JP. Bayesian Approaches to Clinical
Trials and Health Care Evaluation. Wiley, Chichester, 2004.
Spiegelhalter DJ and Best NG (2003) Bayesian methods for evidence
synthesis and complex cost-effectiveness models: an example in hip
prostheses. Statistics in Medicine, 22, 000-000