Bayesian Diagnosis¶
Diagnostic reasoning — instantiates Bayesian Belief Updating
Combines a base rate or pretest probability with test evidence to revise the plausibility of a condition, cause, or hidden state.
Bayesian Diagnosis asks which hidden condition is actually present and answers it by starting from how common that condition is before any test, then revising in light of one specific observed result. Its defining discipline is that the result is never read on its own: a positive finding is worth only as much as the gap between how often it appears when the condition is present and how often it appears when it is absent — weighed against how rare the condition was to begin with. The output is a posttest probability — the plausibility of the condition given both its base rate and the evidence — not a yes/no verdict. What makes it THIS mechanism and not a generic update is the base-rate anchor: it refuses to let a dramatic result erase a low prevalence.
Example¶
An on-call engineer is paged for a spike in checkout latency. From the incident history, the candidate hidden causes carry base rates: a slow downstream payments dependency (common, ~50% of past spikes), a bad cache node (~30%), a memory leak in the checkout service (~15%), and a rare DNS/config fault (~5%). The new evidence: the payments dependency's own dashboards are green, and the latency histogram is sharply bimodal.
Each cause's base rate is the pretest probability. The engineer folds in the observation's diagnostic value — how expected that green-payments-plus-bimodal signature is under each cause. Payments-green sharply lowers the payments hypothesis; a bimodal latency profile is very expected under a failing cache node. The cache-node posterior jumps from 30% to roughly 70%. That posterior — a ranked plausibility, not a certainty — is what directs the first mitigation (drain the suspect node) while keeping the memory-leak hypothesis alive as the fallback if draining doesn't help.
How it works¶
- Name the hidden states and their base rates. Draw the pretest probabilities from a reference class — past incidents, prevalence, historical frequency — before looking at the new result.
- Ingest the observed result as evidence, noting its reliability and what could produce a false signal.
- Update in odds form. Multiply the pretest odds by the evidence's diagnostic value (its likelihood ratio, taken as a given input) to get posttest odds; convert back to a probability.
- Report the ranking, not a label. Carry the posttest probability of each candidate into the response so the runner-up survives.
The base-rate step is load-bearing: skip it and a striking-but-common signal will masquerade as a diagnosis.
Tuning parameters¶
- Reference class — how narrowly you define the population the base rate comes from; narrower is more relevant but noisier.
- Candidate set breadth — how many hidden states you carry; too few excludes the true cause, too many spreads the evidence thin.
- Evidence granularity — one decisive finding vs. a panel of weaker ones combined.
- Posttest reporting form — a full ranked probability vs. the top candidate with a confidence band.
- Revision trigger — how surprising a new finding must be before you re-open the diagnosis.
When it helps, and when it misleads¶
Its strength is that it is the standard corrective for treating a vivid result as decisive: forcing the prevalence back into view prevents the classic error where a low-base-rate condition gets "diagnosed" off one striking signal. Its failure mode is that it is only as good as the base rate — a wrong or stale reference class, or a mis-specified false-positive rate, quietly poisons every posterior. The classic misuse is base-rate neglect — reading the test's accuracy as if it were the probability of the condition and ignoring how rare the condition is[1]. The guarding discipline is to name the reference class out loud, sanity-check the posttest against the pretest (a huge jump from a weak test is a red flag), and keep the runner-up hypothesis alive rather than collapsing to the leader.
How it implements the components¶
base_rate_reference— its opening move: it anchors on the condition's prevalence or historical frequency as the pretest probability.evidence_observation— it ingests the specific observed result and weighs its reliability before letting it move belief.posterior_estimate— its output is the posttest probability of each candidate hidden state.
It does not build the likelihood contrast or run the odds-form arithmetic — it consumes likelihood_model and update_rule from Likelihood-Ratio Reasoning. Its nearest twin is Posterior Risk Estimation, which produces a calibration_check-governed, posterior_communication_note-wrapped continuous risk score for a future event; Bayesian Diagnosis instead infers which categorical hidden cause is present now. It leaves the action cut-point (decision_threshold) to Adaptive Decision Threshold.
Related¶
- Instantiates: Bayesian Belief Updating — the diagnostic specialization: base-rate-anchored inference of a hidden state.
- Consumes: Likelihood-Ratio Reasoning supplies the diagnostic value (likelihood ratio) of each finding.
- Sibling mechanisms: Likelihood-Ratio Reasoning · Posterior Risk Estimation · Adaptive Decision Threshold · Prior Sensitivity Analysis · Sequential Forecast Update · Bayesian Model Update
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Combines a base rate or pretest probability with test evidence to revise the plausibility of a condition, cause, or hidden state, making its operative form a computation or analytic transformation that produces an inference, comparison, or optimized result.
Independent corroboration: The frozen evidence defines Bayesian Diagnosis as 'Combines a base rate or pretest probability with test evidence to revise the plausibility of a condition, cause, or hidden state', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Medicine & Healthcare
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Clinical diagnosis combines disease prevalence with test sensitivity and specificity to compute posttest probability.
Related originating lineages:
- Psychology — Base-rate-neglect research explains why unaided diagnostic judgment misreads vivid positive evidence.
- Statistics & Experimental Design — Bayes' theorem provides the prior-likelihood-posterior structure.
Review resolution: Medicine is the agreed primary lineage through pretest-to-posttest diagnostic reasoning. Bayesian statistics is independently formative, while psychological judgment research informs misuse and debiasing rather than the core diagnostic mechanism.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The question this mechanism answers — how likely is this cause? — is not the same as how much does it matter? Diagnosis ranks the plausibility of hidden states; the magnitude of the harm each carries, and whether to act, belong to Posterior Risk Estimation and Adaptive Decision Threshold. Keeping them separate is what lets a team correct a bad base rate without re-litigating the whole response.
References¶
[1] Kahneman, D., & Tversky, A. "On the Psychology of Prediction". Psychological Review 80(4), 237–251 (1973). Shows that intuitive predictions can be largely insensitive to prior outcome probabilities—the pattern known as base-rate neglect. registry ↩