Description
Treat a probability estimate as a designed mental experiment: choose how to break it into component judgements, elicit those, and combine them by an explicit rule.
A judged probability can be built in different ways, and the choice of decomposition affects its quality. Shafer and Tversky describe designs in two formal languages. Total-evidence designs split the target into scenarios with their probabilities and conditional chances, combined by the law of total probability. Conditioning designs form a prior from part of the evidence and update with likelihood ratios for the rest. Jeffrey designs fix probabilities for a partition directly and carry earlier conditionals forward. Belief-function designs treat each independent argument as a support function and combine them with Dempster's rule, making fewer commitments than the evidence warrants. Good design matches how the judge's knowledge is organised, by causes, by past cases or by separate arguments, and comparing designs exposes which components are guesswork.
Attributes
- Discipline
- Game Theory & Decision Science
- Loss type
- Unmigrated
- Stage
- Strategy, Definition
- Data
- Analytical / expert
- Effort
- Hours
Steps
- State the target event and list the bodies of evidence available.
- Choose a design: total evidence (scenarios and conditionals), conditioning (prior plus likelihood ratios), Jeffrey (direct partition probabilities) or belief functions (one support function per independent argument).
- Elicit only the component judgements the design calls for, for example P(scenario) and P(outcome given scenario).
- Combine with the matching rule: total probability, Bayes' rule in odds form, Jeffrey's rule or Dempster's rule.
- Where feasible, run a second design and compare results; note which component judgements felt speculative.
- Record the decomposition with the estimate so it can be revisited when evidence changes.
Sources
- Decision Making: Descriptive, Normative, and Prescriptive Interactions (Bell et al., 1988)
Also known as
- Probability judgement design
- Decomposition of probability assessments
- Likelihood-ratio partitioning design
- Belief-function (Dempster-Shafer) design
- Total-evidence design
Categories
- Strategy & Collaboration
- Game Theory & Decision Science