Diagnostic Validity — Bayes' Theorem

Post-test probability, predictive values and likelihood ratios · VPP = Se·P / (Se·P + (1−Sp)(1−P))

Test parameters

Reference population

Hypothetical cohort size used for the tree and the 2×2 table

people

Prevalence

Pre-test probability of having the disease

0.001100

Sensitivity

P(test+ | disease) — true positive rate

0100

Specificity

P(test− | healthy) — true negative rate

0100

Clinical scenarios

Bayesian metrics

PPV — Positive Predictive Value

20.00

NPV — Negative Predictive Value

99.67

Accuracy

80.75

LR+

4.75

LR−

0.06

Youden's J Index

0.750

Diagnostic Odds Ratio

76.00

Diagnostic confidence of a positive test

Low confidence

A positive test does not confirm the disease

Natural-frequency interpretation

Of the 2 375 positive cases in this cohort, approximately 475 truly have the condition.

TP: detected disease · FN: missed disease · TN: correctly ruled-out healthy · FP: healthy with wrong positive test

Fundamentals & Explanation

What is Bayes' Theorem in diagnostics?

Bayes' Theorem updates a probability in light of new evidence. In medical testing it answers the clinically relevant question: "given a positive test result, what is the actual probability that the patient has the disease?". That post-test probability is the Positive Predictive Value (PPV) and depends on three inputs: the disease prevalence, the test sensitivity (P(test+|disease)) and specificity (P(test−|healthy)).

PPV = \dfrac{Se \cdot P}{Se \cdot P + (1-Sp) \cdot (1-P)}

Symmetrically, the Negative Predictive Value (NPV) measures how confidently a negative test rules out disease:

NPV = \dfrac{Sp \cdot (1-P)}{Sp \cdot (1-P) + (1-Se) \cdot P}

The prevalence paradox

A test can have excellent intrinsic properties yet fail as a diagnostic tool when the disease is rare. Applied to a population with 0.1 % prevalence (Se = Sp = 99 %), the PPV collapses to ~9 %: only 9 out of every 100 positive results are true cases. The reason: although the false-positive fraction is small, it is applied to a huge group (the healthy population), generating many more spurious positives than true ones.

The "PPV vs Prevalence" tab visualises this collapse: the curve drops sharply as prevalence decreases, with sensitivity and specificity held fixed. This is why screening programs are followed by confirmatory tests: the second test "sees" a much higher effective prevalence (the patients who already tested positive on the first one).

Reference: Gigerenzer & Edwards (2003), BMJ — "Simple tools for understanding risks"

Natural frequencies

Reasoning in natural frequencies (how many people out of 10,000) reduces interpretation errors compared to conditional probabilities. The "Frequency Tree" tab shows the full cascade: the population is split first into diseased and healthy by prevalence, and then each branch into true/false positives and true/false negatives by the test properties. The 2×2 table presents the same numbers in canonical epidemiological form (rows = test result, columns = true status).

Reference: Gigerenzer & Hoffrage (1995), Psychological Review — "How to improve Bayesian reasoning without instruction"

Caveats

•Se and Sp are not universal: they depend on the studied population (clinical spectrum, severity, comorbidities). A test validated in severe patients may degrade in population screening.
•Cut-off threshold: Se and Sp trade against each other when the positivity threshold moves. The ROC curve traces that frontier.
•Test independence: when chaining tests, errors are typically assumed conditionally independent given the true status — an assumption frequently violated in practice.

Likelihood ratios

LR^{+} = \dfrac{Se}{1-Sp}

LR^{-} = \dfrac{1-Se}{Sp}

Likelihood ratios compare the probability of a result between diseased and healthy individuals. They are independent of prevalence, which makes them portable across populations. Practical clinical heuristics:

LR+ > 10— strong evidence to rule in.
LR+ 5–10— moderate evidence.
LR+ 2–5— weak evidence.
LR− < 0.1— strong evidence to rule out.

Reference: CEBM Oxford — Likelihood ratios

Youden's J and DOR

J = Se + Sp - 1

J ∈ [−1, 1]. J = 0 corresponds to an uninformative test (equivalent to chance); J = 1 indicates a perfect test. Negative values mean the test performs worse than chance — its interpretation should be inverted.

DOR = \dfrac{LR^{+}}{LR^{-}}

The Diagnostic Odds Ratio summarises test performance in a single number that is independent of prevalence. It is widely used in diagnostic meta-analyses.

Reading the traffic-light

PPV < 50%— a positive does not confirm; false positives dominate.
50% ≤ PPV < 80%— moderate evidence; confirmation advised.
PPV ≥ 80%— high probability of true disease.

Diagnostic Validity — Bayes' Theorem

Test parameters

Bayesian metrics

Fundamentals & Explanation

What is Bayes' Theorem in diagnostics?

The prevalence paradox

Natural frequencies

Caveats

Likelihood ratios

Youden's J and DOR

Reading the traffic-light

You may also like:

What is Bayes' Theorem in diagnostics?

The prevalence paradox

Natural frequencies

Caveats

Likelihood ratios

Youden's J and DOR

Reading the traffic-light