Sensitivity Analysis
Definition
A technique for assessing how much the posterior probabilities in a Bayesian network change when the values in the conditional probability tables are varied. Mandatory for forensic applications where CPT values are estimated from limited data; courts in several jurisdictions require sensitivity results to be disclosed alongside the main output.
- Applies to
- Bayesian network posterior probabilities
- Purpose
- Test how CPT value changes affect outputs
- Why mandatory
- CPT values often estimated from limited data
- Court expectation
- Disclosure alongside the main output, several jurisdictions
- Field
- Forensic Bayesian evidence evaluation
Common questions
What does it mean if a posterior probability is highly sensitive to a CPT input?+
It means a modest, defensible change in an estimated conditional probability produces a large swing in the conclusion, which signals that the result rests on a shaky assumption and needs either better data or a more cautious, qualified presentation to the court.
How does sensitivity analysis differ from simply checking the network's structure?+
Structural review asks whether the variables and dependencies in the network correctly represent the case, while sensitivity analysis asks a separate quantitative question, how much the numeric assumptions inside a correctly structured network matter to the final answer.
What practical form does a sensitivity result typically take in a report?+
Analysts often vary one or more uncertain CPT entries across a plausible range and report the resulting range of posterior probabilities, sometimes as a tornado chart or simple table, so the fact-finder can see how stable or fragile the conclusion is.
Related terms
- Belief Propagation
- An algorithm for computing marginal and posterior probabilities in a Bayesian network by passing messages between neighbouring nodes. Exact on tree-structured networks;...
- Conditional Probability Table (CPT)
- A table that specifies the probability distribution of a node given every combination of states of its parent nodes. Every non-root node...
- D-Separation
- A graphical criterion that determines whether two sets of nodes in a Bayesian network are conditionally independent given a third set. If...
- Directed Acyclic Graph (DAG)
- A graph in which edges have a direction (from parent to child node) and no path can return to a node it...
- Mixture Likelihood Ratio
- The ratio of the probability of observing a mixed DNA profile if the person of interest is a contributor to the probability...