Types of decision
Download
Report
Transcript Types of decision
Economic evaluation of health
programmes
Department of Epidemiology, Biostatistics
and Occupational Health
Class no. 16: Economic Evaluation using
Decision Analytic Modelling II
Nov 3, 2008
Plan of class
Decision-analytic modeling: General
considerations
Markov models
Patient-level simulations
Measurement vs. Support to
decision-making
Classes 1 to 14 had to do with measurement:
Costs
(Outcomes)
Utilities associated with outcomes
Essential for individual studies
Need to integrate results of individual studies,
and go beyond, to inform decision-making
To inform decision-making, a single
study using one set of primary data
is not enough
Integrate all relevant evidence
• Multiple studies
• Consider all relevant alternatives
• Extrapolate from intermediate to final
endpoints
• Extrapolate further into the future
• Make results applicable to decision-making
context
Multiple studies of effects of
an intervention
Results of any one study influenced by:
Sampling variability
Study design details (e.g., inclusion and
exclusion criteria, drug dosage)
Contextual factors (e.g., health care system
characteristics)
Averaging across multiple RCTs or other
comparative studies can help us attain
true value
Consider all relevant
alternatives
Good decision requires considering more alternatives
Individual studies usually consider few alternatives
Ex: Tx of rheumatoid arthritis (RA): NSAIDs vs diseasemodifying antirheumatic drugs (DMARDs) vs newer biologics.
Many possible Tx options, including regarding timing of
introduction of DMARDs.
Not all trials consider all options.
• Ex: one trial considers homeopathy vs NSAIDs vs DMARDs.
Extrapolate from intermediate
to final endpoints
Many trials consider intermediate clinical
endpoints:
% reduction in cholesterol level
CD4 count and viral load test for HIV
Change in Health Assessment Questionnaire (HAQ)
score for functional disability (RA)
Medication adherence
Distant from outcomes meaningful for decisionmaking
Need to extrapolate, using other studies
Extrapolate further into the
future
Most trials short-term
Long-term consequences often relevant
E.g., supported employment, Tx of RA
Modeling can provide plausible range for
LT consequences
Extrapolate survival data using various
assumptions
Extrapolate using modeling
Make results applicable to
decision-making context
Economic analysis : costs and consequences
under normal clinical practice
O’Brien et al. 95: Adjust for rates of asymptomatic
ulcers (Box 5.1)
Make results applicable to other setting
Subgroups with different baseline effects – see Figure
9.2
• Do this on basis of plausible clinical explanation, not data
mining
Common elements of all
decision-analytic models
Probabilities
Bayesian vs frequentist notions of probability
Frequentist – probability is a measure of the true
likelihood of an event.
• Probability of rolling a 1 with standard die: 1/6
Bayesian – probability is a subjective estimate of the
likelihood of an event.
In decision-analytic models, we do not know
probabilities in the frequentist sense. So we use
expert judgement.
• Is it a weakness? Not necessarily. May be the best that we
can do.
Expected values
Multiply outcome by probability;
See Box 9.3
Stages in development of
model
Define decision problem
Define model boundaries
Structure the model
Types of decision-analytic
models
3 basic options:
– Decision trees
– Markov models
– Patient-simulation models
Why use a Markov model instead of a
decision tree?
• Decision tree can get too complicated if the
sequence of events is too long.
– Especially likely to occur when modeling treatment of
chronic illness
Example:
Welsing, Severens et al. (2006). Initial
validation of a Markov model for the
economic evaluation of new treatments for
rheumatoid arthritis. Pharmacoeconomics
24(10): 1011-1020
Purpose: Initial validation of Markov model
to carry out cost-utility analyses of new
treatments for treatment of rheumatoid
arthritis
Limitations of Markov models
Memory-less state transition probabilities
May be excessively unrealistic
3rd alternative: patient-level
simulation
Each individual encounters events with
probabilities that can be made pathdependent
Virtually infinite flexibility
But how to “populate” all model
parameters?